Company News

We are excited to share that dida, together with our partner Consileon, has successfully secured a framework agreement on AI with the Land Oberösterreich. The agreement runs for 4 years (with an option for 2 more). This long-term perspective is a fantastic opportunity to move beyond short-term pilots and implement sustainable AI solutions for the public administration that follow a coherent strategy and vision. The partnership is a classic "best of both worlds" setup: while Consileon brings their extensive expertise in public sector consulting and strategy, dida will take charge of the technical realization and AI software development. By combining these two areas of expertise, we will be able to implement solutions that are innovative and actually work in the context of public administration. We're looking forward to the next few years of collaboration and to bringing digital transformation to life in Upper Austria.

Last week, we kicked off our "BLISS x dida" workshop series. What started as an idea to bridge the gap between AI research and industry practice turned into a packed room of over 90 curious minds at TU Berlin. A big thank you to Reduan Achtibat from Fraunhofer HHI for leading our first session: "Opening the Black Box with LRP." Reduan took us on a deep dive into Explainable AI (XAI) for LLMs. In a world where LLMs are often seen as black boxes, sessions like these are vital for building safe, faithful, and steerable AI systems. Key takeaways for the participants of our live-coding session: Tracing reasoning from tokens to predictions using LRP Identifying how specific attention heads retrieve information during in-context learning Intervening in latent representations to steer generation Understanding why an LLM makes a decision is just as important as the decision itself. This was a fantastic first session of our workshop series. Stay tuned for the next session, which will already follow in the next weeks!

Earlier this week, we were honored for our work on xJuRAG, our joint project with the Fraunhofer Heinrich Hertz Institute (HHI). Our colleagues Emilius Richter, Julius Lauenstein and Manu Keiper presented xJuRAG to the audience of the Digital Justice Summit, explaining how we are tackling the black box problem of AI generations for the legal and justice sector. For these domains, it is not enough for an AI system to answer complex legal questions - it must show how it actually derived an answer. So with xJuRAG, we are working on: Transparency: Citing sources and revealing its decision-making foundation. Trust: Using Attention-aware Layer-wise Relevance Propagation (AttnLRP) to show how the model reaches its conclusions. Efficiency: Accelerating legal research without sacrificing accountability. The consensus of the Digital Justice Summit was quite clear: The digitization of the justice system is urgent. In order to cope with the increased pace of modernized workflows at legal firms, the justice sector needs to also pick up the pace, especially due to factors like a skilled labor shortage caused by demographic shifts - often referred to as reaching "Equality of Arms" (Waffengleichheit). However, the crucial condition remains: AI in justice must be sovereign, compliant, and above all explainable. That is why we at dida are incredibly proud to have won the Digital Justice Award!

Last week, our colleague Dr. Maximilian Trescher presented our work on GENIAL4KMU at the Mittelstandsforum 2025 in Berlin. The project focuses on optimizing RAG approaches and LLM generations for German SMEs dealing with specific industry terminologies. Together with the Hub of Computing and Data Science (HCDS) Hamburg, we presented a poster and answered many questions from interested visitors. The Mittelstandsforum, held in November, focused on the megatrends of AI and sustainability in the construction industry. The aim was to provide decision-makers with insights into more sustainable construction practices and the use of AI. We would like to thank the organizers for the opportunity to present our project, and we look forward to more events like this in the future.

On Tuesday Markus Düttmann and David Berscheid had the opportunity to join IPAI's Experience Day. The event brought together the IPAI community to showcase their latest work in AI and share their learnings with other IPAI members. While we've had plenty of interesting conversations, some of our highlights where: Aleph Alpha's talk on "Specialists build the world, generalists talk about it", where Yasser Jadidi presented research efforts of Aleph Alpha like tracing in LLMs and how they relate to the final customer value - concepts we might consider in some of our own LLM-focused research projects; Getting great demos from the MANN+HUMMEL team about their AI in manufacturing and data sharing platform; Learning from AUDI AG how they are working on their MLOps platform to enable machinery optimizations across manufacturing plants; Talking to the ebm-papst team on topics like demand forecasting and tooling decisions; Overall the Experience Day was packed with interesting demonstrations. This event was a great example of how the typical narrative of European enterprises lacking ambition in AI adoption often falls short. What we saw here was great momentum, ambitious AI agendas, and realistic approaches to achieving them. Thank you to the IPAI team for the invitation and for the great organization!

We are excited to announce the start of our xJuRAG project, which is short for "Explainable AI for legal RAG applications", a new joint research initiative. As the project coordinator, dida is partnering with the renowned Fraunhofer Heinrich-Hertz-Institut (HHI) to tackle one of the biggest challenges in artificial intelligence: the "black box" problem. While LLMs show great potential, their tendency to "hallucinate" or invent facts makes them risky for high-stakes, sensitive fields like the legal sector. This lack of trust is a major barrier to the widespread adoption of AI in professional legal environments. The goal: A fully transparent legal assistant The primary goal of xJuRAG is to build a legal assistance system that is not only intelligent but also completely transparent. We are developing an application based on Retrieval Augmented Generation (RAG) , specifically designed for the legal field. The system will be able to: Answer complex legal questions Reliably cite its sources Clearly explain how it arrived at its conclusions The project will initially focus on decisions in traffic and tenancy law. The innovation: Opening the black box xJuRAG's innovation lies in its deep integration of advanced Explainable AI (XAI). We will adapt a powerful method developed at Fraunhofer HHI, Attention-aware Layer-wise Relevance Propagation (AttnLRP). This approach goes far beyond standard RAG systems. It allows us to create a detailed "audit trail" that quantifies exactly which words in a source document were most relevant to the final answer. The justice system is built on clarity and verifiability - principles that standard AI cannot guarantee. xJuRAG aims to meet these principles directly. By providing a reliable and transparent tool, we will help legal professionals automate time-consuming research and build confidence in AI-supported decisions. This will free up experts to focus on higher-level strategy and argumentation. By proving the viability of explainable AI in a sensitive field, xJuRAG will also help promote public trust and accelerate responsible AI adoption in other critical sectors like medicine and public administration. The project is structured in three phases and is set to run for 36 months Contact If you represent a legal firm or legal-tech company and are interested in learning more or exploring potential applications, please reach out to us.

Last week, Dr. Lorenz Richter accepted STAC's invitation to speak at the STAC Summit in London about solving partial differential equations (PDEs) in high-dimensional settings. The conference, which was aimed at technical decision-makers in finance, proved to be a great opportunity to share some of our more technical insights about physics-informed neural networks (PINNs) and methods based on backward stochastic differential equations (BSDEs). In computational finance, overcoming the curse of dimensionality is a constant challenge. Therefore, Lorenz demonstrated the connections of PINNs and BSDEs to stochastic optimal control and reinforcement learning and how these approaches open new frontiers for solving complex problems in finance. The summit was a two-day event, filled with fantastic speakers and deeply sophisticated topics around AI in finance. It was a great pleasure to meet so many experts in their field. We were glad to contribute and are already looking forward to the next event. A big thank you to the STAC team for the great organization!

Today we’re very happy to share that our paper on stochastic optimal control, which we developed in collaboration with colleagues from Karlsruher Institut für Technologie (KIT), NVIDIA, Microsoft Research and Cornell University has been accepted to NeurIPS 2025. The paper with the title “Trust Region Constrained Measure Transport in Path Space for Stochastic Optimal Control and Inference” explores how a trust-region-based optimization algorithm performs for applications of stochastic optimal control against existing methods. We could show that our novel method can improve tasks like fine-tuning diffusion models, diffusion-based sampling or transition path sampling. A big thank you to all contributors (Denis Blessing, Julius Berner, Lorenz Richter, Carles Domingo-Enrich, Yuanqi Du, Arash Vahdat, Gerhard Neumann) and congratulations to the team on receiving a spotlight mentioning! The paper can be found here: https://arxiv.org/pdf/2508.12511

We are excited to announce that dida has joined IPAI - the “Innovation Park Artificial Intelligence”. The innovation platform from Heilbronn with partners and members such as Schwarz Digits, Aleph Alpha, Telekom, Porsche, Audi, or Würth Group, connects stakeholders from research, industry, and politics, to collaborate on responsible AI developments. With this partnership, dida will contribute especially its expertise in Natural Language Processing, Computer Vision, and MLOps to support IPAI members in pursuing their AI journey, and in developing custom AI solutions that are optimized for security and explainability. As Dr. Markus Düttmann (Managing Director of dida) explains: “In our daily work with German and European corporations, we see a lot of promising AI developments that can be built upon. Networks like IPAI play a central role in bringing these potentials to reality by creating a ground for collaboration between members, and in sharing expertise and resources. We joined IPAI because we are motivated to contribute our deep technical knowledge to such collaborations to further de-risk ambitious AI strategies for European organizations. Let's bring together world-class AI scientists with world-class domain expertise.” On this note, dida is looking forward to future exchanges, initiatives and events within the IPAI ecosystem. If you would like to talk about responsible AI implementations with us or IPAI, please don’t hesitate to contact us.

On September 11th and based on dida's solar planning project with Enpal, Dr. Lorenz Richter gave a talk at BIG BANG KI FESTIVAL on solving complex geometrical problems within AI projects and how to transform 2D information into 3D. The Big Bang KI Festival 2025 was a two-day event focused on AI and digital trends, branded as "Europe's largest AI event," with 8,500 sold-out tickets. We're glad to have contributed to this large conference with speakers from public institutions, various sectors and technology providers. Thank you for the invitation and especially to the audience of our masterclass for the engaging session and participation, which led to a great discussion.

Last week, our Managing Director and CTO, Dr. Lorenz Richter, had the opportunity to meet with one of Germany’s leading reinforcement learning groups, the Autonomous Learning Robots Lab at Karlsruher Institut für Technologie (KIT), and present his latest ICML paper "Reinforcement Learning with Random Time Horizons”. The ALR Lab, led by Professor Dr. Gerhard Neumann, aims to develop novel machine learning methods for robotics, which are suitable for complex robot domains such as grasping and manipulation or dynamic motor tasks. In his presentation, Lorenz shared his approach of transferring concepts from stochastic optimal control to reinforcement learning - two seemingly separated research domains, with interesting parallels - followed by a technical discussion with the group’s researchers. We would like to say thank you for the invitation and the possibility to discuss our latest research with experts from KIT!

A delegation of the dida team (Dr. Marty Oelschläger, Jakob Scharlau, Dr. Konrad Schultka, Dr. Sebastian Thomas) had the pleasure of a first-time visit to nVent Schroff's facilities here in Germany. We not only discussed the progress of our joint project but also had the incredible opportunity to learn about Schroff's rich company history, a leader in electronics packaging and infrastructure solutions for over 65 years. What's particularly inspiring is how nVent Schroff continues to innovate. Their expertise helps make critical systems resilient, providing vital infrastructure for areas like energy storage and cutting-edge data centers. Looking forward to what we'll achieve together!

Let’s recap this year’s International Conference on Machine Learning (ICML), which took place in Vancouver, Canada. We are happy to once again have been part of the ICML and, in this case, we are proud to have presented our first paper related to reinforcement learning (RL). In joint collaboration with Enric Ribera Borrell and Christof Schütte, we discussed our paper “Reinforcement Learning with Random Time Horizons”. In our work, we could extend RL to settings with random, policy-dependent time horizons by rigorously deriving new policy gradient formulas - both stochastic and deterministic - and show that these improve optimization convergence. Apart from our presentation, we mixed and mingled with other international researchers. The most dominant topics at the conference this year could be clustered to the following: (1) LLMs & Agents, (2) Diffusion Models, (3) Interpretability & Alignment, (4) Optimization & Scaling Laws, and (5) Fairness It once again was a truly great atmosphere - another reason why we like to share our research on an international stage. Link to our paper: https://arxiv.org/pdf/2506.00962

Last week, Julius Lauenstein and Lorenz Richter followed the invitation by publicplan to attend and contribute to their annual “Anwendertag” in their headquarters in Düsseldorf. The event is publicplan’s dedicated event to showcase digital administration in practice, with a focus on AI and open source solutions for the public administration. With representatives from authorities, public administration, and ministries, it was a great day to celebrate publicplan’s 15th anniversary. Lorenz contributed a talk about "Information extraction with AI and LLMs: from complex forms to handwriting recognition", where he showed edge cases of information extraction scenarios that are not trivial to solve. We would like to say thank you to Christian Knebel, Fabian Nagel and Moritz-Maximilian Katzer for the invitation and congratulations once again to the organization of this great event. We’re looking forward to many more joint developments for the public sector in the future!

Last week we had the opportunity to visit IPAI for the first time. Our colleague David Berscheid gave a presentation on "Production-ready AI projects: Why many projects fail without individual optimizations" for the members of the IPAI community. IPAI, which stands for Innovation Park Artificial Intelligence, is a highly ambitious project in Heilbronn, which aims to build one of Europe’s leading ecosystems for applied AI technology, bringing together corporations, startups, researchers, and public sector actors to develop and commercialize AI solutions with a focus on ethical and responsible use. It was really great to see the IPAI spaces, getting to know the IPAI team and learning more about the vision for this whole region. Thank you to the event and partner management teams for the invitation and the possibility to share some insights into our work of how we optimize our projects through mathematical optimizations. We’re looking forward to more IPAI events in the future!

One week ago we had the pleasure to attend the Zukunftsforum Kunststoffkreislauf hosted by DECHEMA Gesellschaft für Chemische Technik und Biotechnologie. Our colleagues Dr. William Clemens and Dr. Eva Höning represented dida and presented our work on KIOptiPack - a project focused on identifying the appropriate recycling method for plastic products using AI. At the event, they showcased a poster on our latest technological developments: "Probabilistic Modeling of Unknown Plastic Properties with Denoising Diffusion." This approach aims to improve the accuracy of material classification when product information is incomplete - supporting smarter and more sustainable recycling decisions. Thank you DECHEMA for organizing such an inspiring event and to all the participants for the valuable conversations!

dida's AI-supported document processing product called smartextract and the company on-geo GmbH are officially announcing their exclusive partnership in the field of land register data extraction for the financial and real estate industries. On-geo is the market leader in software-supported market and mortgage lending valuations as well as a solution provider for experts, appraisers and service providers in the real estate industry. By integrating the AI solution smartextract into existing software products such as on-geo's LORA, the company enables fast and digital extraction of land register data for banks, savings banks, financial brokers, appraisers, real estate agents and insurers. Among other things, this replaces the manual transfer of land register information to the valuation software, meaning that AI significantly reduces the amount of work required by users and increases data quality. In addition, the use of smartextract significantly optimizes the digitization of land register content already offered via the geoport web store, so that data can be provided much faster than before. This also applies to the on-geo service for processing entire portfolios. The option of integrating smartextract into third-party systems will also be included in the scope of services in future. This will make the entire real estate valuation process more efficient and economical. Matthias Knabe, Managing Director at on-geo, says: With the integration of smartextract into our software solutions, we are taking a further step in the digital transformation of the real estate and finance industry. Through our collaboration with smartextract, we can now offer our customers an even more efficient solution for land register data extraction and actively shape the future of the real estate industry. Axel Besinger, VP Product & Growth at smartextract, comments the collaboration as follows: We are united with on-geo by the conviction that AI technology not only accelerates workflows, but can also set new efficiency standards. The partnership is a logical and consistent step towards sustainably supporting the digitalization of the real estate and finance industry with intelligent document processes. About smartextract smartextract is a brand of dida - one of Germany's leading AI and Machine Learning agencies. smartextract offers specialized AI solutions for automated document processing and enables the extraction, classification and comparison of data from business documents and emails - flexibly as an API or as white label solution for platform providers and software manufacturers. More information at: smartextract.ai About on-geo 80,000 customers, 1.6 million valuations, 1.9 million data and document retrievals and 250,000 property viewings per year. These are the statistics from on-geo. The Erfurt-based company develops software solutions for greater transparency and efficiency in real estate valuation. These include the digital construction financing tool baufi.me, LORA - the market-leading process solution for determining mortgage lending values and accumate, an AI-based automated valuation model. The portfolio is complemented by the research data store geoport. More information at: on-geo.de Press contact on-geo GmbH Andy Dietrich andy@strategiekollegen.de Mobil: 0170 2283299

Last Thursday our colleague Julius Lauenstein attended the Creative Bureaucracy Festival in Berlin. The Creative Bureaucracy Festival is an annual event that brings together public sector innovators, creative minds, and changemakers from Germany and the DACH region. Its core purpose is to rethink the principles, processes, and practices of public institutions and explore how better bureaucracy can emerge. The event provided interesting presentations such as the session by Iliya Nickelt from the Federal Ministry of Economic Cooperation and Development (BMZ) Data Lab on balancing agile, cloud-based prototyping with the long-term demands of secure public services, or the impressive AI-driven tools presented by the Federal Ministry of Labour and Social Affairs and the German pension insurance to support large-scale inspection processes that were previously highly labour intensive. It was great to see the shared sense of purpose and drive across all these initiatives, focussing on progress rather than scepticism. The public sector is not lagging behind in innovation - it’s actively shaping the future with responsibility, expertise, and momentum. We're glad to be part of ongoing digitalization efforts within German and European governments, developing AI-based process automations for more efficiency. If you would like to speak to us about AI-related public sector initiatives, please contact us.

Last Friday, we had the honor of welcoming around 300 AI and machine learning professionals at fizzforum in Berlin. With our annual conference, we aim to provide a platform where leading scientists and engineers can present their theoretical and applied work around AI and Machine Learning. This year's program included a mix of all the topics that are important to us: Bringing together public sector stakeholders at our pre-conference breakfast, who are working hard to make German and European governments AI-ready. Providing insights from leading AI organizations into the currently hottest topics, such as by Dr. Julian Risch (deepset) with his workshop on AI agents or by Andrey Vasnetsov (Qdrant) with miniCOIL. Looking at AI developments holistically, such as in the Responsible AI workshop (Florian Dohmann, Dina Dürkop, Dr. Liel Glaser), with Dr. Sebastian Lapuschkin's talk on Explainable AI, or by using confidential computing techniques (Thomas Strottner, Edgeless Systems). Offering a stage to many great researchers during our poster session, where guests could learn about AI projects from Technische Universität München, Universität Hamburg, BalkonSonne.App, Zuse Institute Berlin, Fraunhofer Heinrich Hertz Institute HHI, Humboldt-Universität zu Berlin or Universität zu Lübeck. And last but not least: Bringing AI into production - with NVIDIA’s Triton Inference Servers (Dr. Marty Oelschläger, Fabian Dechent), in embedded platforms (Edwin Fung, Dr. Tassilo Glander) or in AI-driven products such as Mirelo AI (Dr. Florian Wenzel) and smartextract (Axel Besinger, Dr. Augusto Stoffel). We would like to say thank you to everyone who helped contribute to this special day! Thank you to all contributors, long-term partners, guests and the whole dida team for joining the dida conference 2025!

The International Conference on Learning Representations (ICLR) 2025 in Singapore once again brought together leading researchers to discuss new findings in machine learning. Dr. Lorenz Richter, Managing Director and CTO of dida, participated in the event, presenting dida's joint research about diffusion models, and discussing latest machine learning progress with fellow researchers. dida presented two papers created in collaboration with the California Institute of Technology (Caltech), Cambridge University, the Karlsruhe Institute of Technology (KIT), and NVIDIA. The first paper, "Underdamped Diffusion Bridges with Applications to Sampling" introduced a new approach for diffusion-based generative modeling, leveraging theoretical physics to improve simulation methods for specific sampling challenges. The second paper, "Sequential Controlled Langevin Diffusions" (SCLD), explored a method that integrates Sequential Monte Carlo (SMC) with diffusion-based sampling techniques. This approach demonstrated a reduction in training resource requirements for sampling models. These contributions underscore dida's commitment to advancing AI research and developing solutions to complex problems. It is dida's mission to actively contribute to latest AI research and bridge the gap between world-class research and industrial application. For more information about these studies or dida’s ongoing projects, please visit our company website or contact us directly.

At the end of April, our Machine Learning Scientists Fabian Dechent, Dr. Marty Oelschläger and Dr. Liel Glaser traveled to Frankfurt (Main) where the LLM4Anamnese team met to discuss the progress on the joint research efforts to utilize Large Language Models to improve the analysis of unstructured medical texts. LLM4Anamnese is part of the "Interactive Technologies for Health and Quality of Life" research program, funded by the Federal Ministry of Education and Research (BMBF) with a budget of €1.02 million. Over 36 months, the project explores the application of LLMs for classifying anamneses and linking them with standardized medical terminologies, while ensuring data privacy. The project meeting was held at the Universitätsmedizin Frankfurt and included stakeholders from involved organizations, such as AWS and VDI VDE IT. AWS has committed to supporting the project with cloud credits, ensuring a reliable and secure cloud infrastructure. In our discussion with the IMI (Innovative Medicines Initiative), we learned a lot about the use of LLM4Anamnese in practice, including how LLM4Anamnese can enable automated medical documentation processes such as discharge letters and thus offer considerable process optimization. We would like to thank all participants for this engaging interim meeting and are looking forward to the progress that we'll be making in the next months.

Today we are happy to share that we published our latest paper in the Nature Scientific Data journal. Dr. Jan MacDonald and Dr. Lorenz Richter from dida, in joint collaboration with external partners Joana Reuss and Dr. Marco Körner (both Technical University of Munich, Chair of Remote Sensing Technology), and Dr. Simon Becker (ETH Zürich, Department of Mathematics), developed EuroCropsML - a time series benchmark dataset for few-shot crop type classification in Europe. It is the first time-resolved remote sensing dataset designed to benchmark transnational few-shot crop type classification algorithms that supports advancements in algorithmic development and research comparability. It comprises more than 700.000 multi-class labeled data points across more than 100 crop classes. If you are active in the area of remote sensing analysis we recommend you take a look at the publication and the respective code. Feel free to reach out to us with remarks, questions or remote sensing-related project initiatives.

Last week, dida attended the opening of a new KI Park satellite in Lingen (Lower Saxony), organized by KI Park and CORNEXION. With 250 guests, the event was a great success for the launch of AI-related initiatives in this strong industrial region, which is home to companies of the typical German Mittelstand and some global niche market leaders. Our colleague David Berscheid represented dida and had the opportunity to pitch dida to the audience, followed by interesting discussions about specific AI use cases that attendees were eager to get started with. The event marked the launch of a broader vision to build an AI campus in Lingen, including office space, green energy-powered data centers and a coherent strategy with stakeholders from academia and politics. Among our highlights were insights from regional keynote speakers who gave insights into their AI developments: Prof. Dr. Goy Hinrich Korn (CDO, KRONE GROUP) presented how KRONE is using AI-based revenue forecasting models with a time horizon of up to three years. Benjamin Wolters (Head of AI, ROSEN Group) shared his organization’s take on the development of safe and trustworthy AI systems. Dr. Henning Krüp (Head of AI, LIST Gruppe) explained how to disregard the hype surrounding the topic. We would like to thank the KI-Park for the kind invitation and excellent organization of the event. dida is proud to be a member of the KI Park ecosystem and looks forward to contributing to the responsible and effective integration of AI technologies across industries in Europe.

We are excited to announce that dida has two research papers accepted at the International Conference on Learning Representations (ICLR) 2025, one of the leading global conferences in the field of machine learning, taking place in Singapore this April. Both papers are developed in collaboration with the California Institute of Technology (Caltech), NVIDIA and the KIT. The first paper, "Underdamped Diffusion Bridges with Applications to Sampling“, introduces a novel framework for diffusion-based generative modeling in so-called underdamped settings. The idea is motivated by concepts from theoretical physics and culminates in more advanced simulation methods for generative processes. Applied across various sampling problems, those methods achieve state-of-the-art performance, outperforming alternative methods while requiring fewer discretization steps and no hyperparameter tuning. The second paper, "Sequential Controlled Langevin Diffusions“ (SCLD), introduces a new method that integrates Sequential Monte Carlo (SMC) with diffusion-based sampling techniques. By combining the strengths of both methods, SCLD shows improved performance on benchmark tasks, often using only 10% of the training resources required by earlier diffusion-based methods. With these publications, dida continues to contribute to latest AI research. We look forward to further integrating these insights into our solutions. If you're interested to learn more about these specific topics or dida's research activities in general, please reach out.

On January 16th, dida took part in Circula’s AI Operators Talk, where our CTO & Managing Director, Dr. Lorenz Richter, joined a panel discussion on AI’s role in modern businesses and SaaS solutions. The event provided great insights into how AI can optimize workflows and create real value, particularly in areas like expense management. Circula specializes in optimizing financial workflows, providing automation for expense management, travel costs, and benefits. With features like a DATEV interface that integrates receipt images and pre-recorded data, they help businesses simplify financial processes and improve efficiency. It was an interesting panel discussion, where we could share our experiences related to developing individual ML projects. Thank you to the Circula team for the invitation — we look forward to more discussions like this in the future!

We’re excited to announce that our third dida conference will take place on May 23, 2025, at frizzforum in Berlin. What started as an experiment in 2023 has grown into a highly anticipated event. This year, we’re moving to a bigger venue to accommodate the high demand from previous years. The conference will once again feature a mix of technical and applied presentations on AI and Machine Learning. In past years, we’ve had speakers and panelists from companies like Google, Bundesdruckerei-Gruppe, idealo internet GmbH, Enpal, and APCOA Group GmbH. For 2025, we’re working on putting together an equally impressive lineup. As spots are limited, we recommend registering early to secure your place. Registration is free! You can register for the conference now on our official conference page. To get a glimpse of last year’s event, watch the highlight video here.

At the end of last year, our machine learning scientist Marty Oelschläger accepted an invitation from idealo internet GmbH and Jan Anderssen to speak about "Automatic Product Generation: On our way to price comparison for the entire catalog". In the joint presentation at the idealo Collaboration Day, Jan and Marty presented their entity matching solution for automatically grouping offers from the huge idealo catalog into products and variants using a combination of fast retrieval and high-precision matching. We really enjoyed the event, where different idealo teams get together and inform each other about relevant developments within their departments. Thank you once again, Jan Anderssen, for the invitation and the opportunity to present our project - it's always a rewarding challenge to work with idealo's large amounts of data. (Photographer: Stefan Hähnel)

On November 20th, Maximilian Trescher presented our work on "Anomaly Detection in Track Scenes" at PyData Berlin 2024. In collaboration with Deutsche Bahn as part of the Digitale Schiene Deutschland initiative, we developed a machine learning solution to detect hazardous and anomalous objects on train tracks using onboard RGB cameras. The system uniquely identifies not just predefined classes but ranks objects by anomaly level. Maximilian highlighted the challenges we faced, the development of our multi-component pipeline—featuring depth estimation, segmentation, and anomaly detection—and the use of the OSDAR23 dataset alongside self-supervised learning. We are grateful to PyData Berlin for the opportunity to share this project with the community. Thank you to everyone who attended the talk and contributed to the engaging discussions that followed.

dida is pleased to announce “In-terra-gate”, our latest project to transform satellite data analysis through natural language interaction. This initiative uses advancements in large language models (LLMs), natural language processing (NLP), and visual foundation models to create an intuitive tool that translates complex satellite data into actionable insights. The project is an ESA Kickstart Activity and enables users to query geospatial information using natural language, providing accurate answers about geological and geographic features in satellite imagery. In-terra-gate uses text-image embeddings trained on high-resolution Sentinel-2 satellite data to provide insights into urban areas, residential zones, and regions exposed to natural hazards. This tool was developed with insurance companies in mind, as they need accurate, real-time data to assess individual property risks and expedite claims, e.g. for homeowners and farmers. With In-terra-gate, dida is setting a new standard in satellite data accessibility, providing non-technical users with the opportunity to intuitively harness the potential of complex satellite data analysis in various scenarios and industries.

dida was happy to participate in the recent Generative AI Exchange at Celonis headquarters in Munich, where over 130 AI developers, specialists, and executives gathered to explore the latest advances in AI. Our machine learning scientist, Fabian Dechent, led a hands-on workshop, sharing insights from our deployment of a real-time license plate matching service for our client, APCOA. Fabian's session focused on optimizing the performance of a TrOCR-based OCR system tailored to handle low-quality or partially truncated license plate images. He covered techniques for improving accuracy through metric learning and supervised learning, as well as deployment strategies using Nvidia Triton as a microservice. We would like to thank Celonis, NVIDIA, and KI Park for fostering an inspiring atmosphere, and the attendees for their enthusiasm and engagement. Thanks to the Generative AI Exchange team for the opportunity to contribute to this interesting event!

dida’s ML Scientist Jakob Scharlau had a presentation at the data2day conference, where he shared insights on "Semantic Search and Machine Learning: Finding the Right Document." His talk introduced the audience to the use of machine learning and pre-trained language models to build semantic search systems that understand text contextually. He highlighted recent advances in text embeddings and large language models (LLMs), and provided a technical overview of modern approaches to building tailored search solutions. Using examples from real-world projects, he demonstrated how these techniques are implemented to improve search accuracy and relevance. His goal was to provide a comprehensive yet accessible introduction that would allow attendees with any level of technical expertise to gain an understanding of embedding techniques, their evolution, and live applications of domain-specific semantic search. We sincerely thank data2day for this opportunity to contribute to the advancement of data and AI expertise in the industry.

We're happy to announce that Machine Learning Scientist Dr. Maximilian Trescher has contributed a chapter to the recently published whitepaper of the X-KIT Cluster Computer Vision. Developed as part of the X-KIT project - funded by the German Federal Ministry of Food and Agriculture (BMEL) and supported by the Federal Agency for Agriculture and Food (BLE) - the whitepaper highlights eight innovative projects using computer vision to address challenges in agriculture and livestock management. The X-KIT project promotes collaboration between AI initiatives in these sectors, maximizing synergies and minimizing redundant efforts. Our contribution comes from the KAMI project, where we've developed a prototype that uses image-based technology to automatically and accurately measure the respiratory rate of dairy cows. This system helps farmers monitor the health of their animals efficiently and non-invasively. We're glad to have worked with other contributors to explore how computer vision can create practical, impactful solutions in these areas.

Last week, our machine learning researcher, Dr. Jan Macdonald, presented at the Interdisziplinäre Fachtagung (Interdisciplinary Conference) titled Fast Forward Future at Eberhard Karls University of Tübingen. Jan introduced our research project, PretrainAppEO, which explores transfer learning and meta learning techniques for earth observation. His talk focused on crop type classification using satellite imagery, a key application in agriculture and environmental monitoring. The presentation was designed as an "impulse talk," aimed at introducing different transfer learning methods to the audience. Jan highlighted the relevance of these methods in adapting models to new tasks with limited data, as well as the value of meta learning in enabling models to quickly learn from new environments, which is particularly important for analyzing satellite data under varying conditions. The event brought together experts from diverse fields such as machine learning, environmental sciences, and data analysis. We are grateful for the opportunity to share our work and contribute to discussions on how AI can address challenges in Earth observation, especially in agriculture and land use. We thank Eberhard Karls University of Tübingen for hosting this event and look forward to continued collaboration.

The LLM4Anamnese project aims to use Large Language Models (LLMs) to automatically classify medical histories, which are often recorded in free text form, and link them to standardized medical terminology. This automation not only facilitates the further processing of the texts in digital systems, but also improves data quality without losing the individual aspects of the patient histories. A particular focus of the project is on researching how resource-efficient and robust LLMs can work in practice. The application in everyday clinical practice will be tested under real conditions to ensure that the technology meets the high requirements of the healthcare system. The adaptation of LLMs to medical data is a key step in optimizing the classification of texts and making the results usable for other medical applications. In the long term, the project should help to relieve the burden on medical staff and improve the quality of patient care.

dida had the opportunity of attending the recent AI & Data Summit and Quantum Summit, organized by Bitkom. The event brought together major companies from across the AI and technology sectors, each presenting their innovative AI-related solutions while addressing current challenges and future perspectives in the field. A significant focus of the summit was placed on local German developments, highlighting the contributions to its global advancements. Topics such as Large Language Models (LLMs), Generative AI, Responsible AI, and cutting-edge tools were extensively discussed, with key insights into their implications for both industry and society. We look forward to applying the knowledge gained and remaining at the forefront of AI advancements in the coming years.

We are pleased to announce that all recorded talks from the dida conference 2024 are now available for viewing on our website and our Youtube channel. This year’s conference featured an impressive lineup of industry experts and thought leaders who delivered deep insights on the latest advancements in data science and AI. Whether you attended the event or were unable to join us, you can now access the full range of presentations and discussions at your convenience. To explore the sessions and enhance your knowledge, visit our channel. This is your chance to revisit key moments from the conference and catch up on the sessions you may have missed.

Our CTO, Lorenz Richter, recently attended the International Conference on Machine Learning (ICML) in Vienna. At the conference, Lorenz presented our paper, "Bridging Discrete and Continuous State Spaces: Exploring the Ehrenfest Process in Time-Continuous Diffusion Models." The paper made significant strides in advancing the understanding of generative AI by examining time-continuous discrete-space diffusion models. It established a connection between these models and the widely-used score-based generative modeling, which is prominently applied in image generation. You can read the paper here.

We are pleased to announce the official launch of our GENIAL4KMU project. This project aims to develop an advanced Retrieval Augmented Generation (RAG) system that effectively harnesses unstructured data assets in organizations. GENIAL4KMU leverages the latest advances in Artificial Intelligence, in particular large language models (LLMs), to provide companies with accurate and contextualized data processing. Organized in collaboration with our partner, the University of Hamburg Hub of Computing & Data Science, the project is divided into two phases: First phase: Structured data Here, parameter-efficient fine-tuning of LLMs is carried out first, followed by the development of a RAG system for processing structured data. Second phase: Unstructured data In this phase, the system will be extended to unstructured company data, including further fine-tuning and integration into the RAG system. The developed solutions will be continuously evaluated and transferred into demonstrators to ensure that they meet the highest standards and meet the needs of SMEs. In summary, GENIAL4KMU represents a significant step in the application of AI technologies with the aim of sustainably increasing the efficiency and competitiveness of small and medium-sized enterprises.

We are pleased to look back at another great dida our summer workshop, which took place on July 4-5, 2024. Over the course of two days, our team gathered for a series of engaging and collaborative sessions designed to enhance teamwork, inspire creativity, and foster stronger connections among colleagues. The workshop featured a variety of activities, including a survival camp , strategic workshops, and opportunities for getting to know some of our new colleagues better. Participants were able to share ideas, discuss industry trends, and explore new strategies for future projects. We are grateful to everyone who attended and helped make this workshop a memorable and productive experience. The insights and connections made during these two days will undoubtedly benefit our work going forward. Stay tuned for more updates and exciting events in the future!

On July 10th, the Branchen-Blick Digital- und Kreativwirtschaft took place at the Berliner Volksbank. Industry leaders, innovators and professionals from the digital and creative sectors gathered to discuss the latest trends, challenges and opportunities in the industry. Attendees enjoyed an evening dedicated to networking, sharing ideas and finding inspiration. The event provided a valuable opportunity to connect with other industry experts during the Get-together session, sparking new ideas and exploring innovative ways to grow their businesses. Known for its insightful sessions and networking opportunities, the event once again proved to be an important platform for fostering collaboration and innovation. Attendees had the opportunity to engage with thought leaders and participate in discussions that will shape the future of the digital and creative industries.

We are excited to share that our research has been published in the Journal of Machine Learning Research (JMLR), one of the most respected journals in the field. You can find the paper here. This new paper builds on the work from a previous study of ours, which received an outstanding paper award at the ICML’21 for its contributions to generative modeling and optimal transport. In this follow-up, we take a closer look at the underlying mathematical concepts and introduce practical methods for working with high-dimensional and complex data. The paper also explores new ways to solve related optimization problems and demonstrates how these approaches can be applied in real-world settings. We are happy to see our research evolve and to contribute further to the field. Many thanks to JMLR for publishing our work and to everyone who has supported us along the way.

Last week we had the pleasure of attending Tangere24, hosted by Siemens Global Marketing Services (GMS) in Berlin. Known for their expertise in marketing, sales, and communications, Siemens GMS is striving to be a leader in the use of AI in their business. Our Managing Director, Philipp Jackmuth, and Machine Learning Researcher, Jan Macdonald, represented us at the event. They had the opportunity to gain valuable insights into Siemens' comprehensive AI strategies and explore how AI is transforming their global processes. The event featured collaborative workshops focused on developing practical solutions to real-world challenges in marketing, sales and communications. It provided a dynamic platform for meaningful interaction with AI experts discussing the latest advances and best practices. We had the opportunity to engage directly with the Siemens GMS team, making valuable connections and exploring the potential for joint initiatives. We thank the Siemens team for this insightful event and look forward to future collaborations.

On Monday, July 8th, 2024, Lorenz Richter, CTO of dida, gave a talk at the Fourth Symposium on Machine Learning and Dynamical Systems, hosted by the Fields Institute in Canada, a leading center for mathematical research. His presentation explored generative modeling through the lens of measure transport, describing it as a dynamical system that evolves from an initial distribution to a target measure. Lorenz highlighted how this process, guided by differential equations like ordinary differential equations (ODEs) or stochastic differential equations (SDEs), connects to partial differential equations (PDEs) and demonstrated practical computational approaches. He discussed Backward stochastic differential equations (BSDEs), which tie into optimal control, and PINNs, which simplify training by avoiding trajectory simulation. We thank the Fields Institute for hosting this inspiring event and to the audience for their thoughtful questions and interest.

We are glad to announce that our success story with Enpal, Germany’s first green unicorn and leading solar panel company, has been featured by Google. Enpal, known for their innovative solar panel solutions, faced considerable challenges due to the extensive manual work required in their sales operations. Handling the task of creating customized offers in a manual way typically took 120 minutes per person and per sales offer, leading to inefficiencies and resource constraints. In response to these challenges, we developed a machine learning service tailored specifically to Enpal’s needs. Our solution has innovated their workflow, automating a significant portion of the manual tasks. As a result, what once took 120 minutes now only takes 15 minutes per person — an 87.5% reduction in time. Our ML product not only saves time but also increased solar panel estimations to an approximate accuracy rate of 93%. When we initially launched the software in 2019, 13 employees at Enpal were using it. Today, four years later, this number has evolved to 150 employees. Each of these employees is now able to save 87.5% of their time, allowing them to focus on more specialized and impactful tasks. We extend our sincere thanks to Google for sharing our story. Their support helps us showcase the transformative power of our machine learning solutions. You can read the full story on the Google Cloud Blog here.

We are happy to share insights from the Digital Farming Conference 2024, hosted by Bitkom. Represented by Julius Lauenstein, dida actively contributed to discussions at this pivotal event focused on the digital transformation of agriculture and food industries. Key topics included advancements in precision agriculture, smart farming, and leveraging AI and IoT for managing livestock, enhancing animal welfare, and promoting sustainable practices. The conference highlighted advancements in precision agriculture, focusing on AI-driven applications for resource optimization and sensor-based technologies. At dida, we regularly work on several projects in the field of remote sensing, such as crop classification, which underscores our commitment to enhancing agricultural efficiency and sustainability. Additionally, the event emphasized topics aligned with our project KAMI, which is dedicated to enhancing animal health and well-being, specifically focusing on the "Transformation der Tierhaltung" (Transformation of Animal Husbandry). This highlighted dida’s efforts in promoting more animal-friendly and sustainable farming practices. As active members of the innovation community that organized the panel, alongside partners like DLR Group and Space2Agriculture, we are dedicated to innovative solutions that shape the future of agriculture.

Our new research project, "FewTuRe," aims to help small and medium-sized businesses (SMEs) to automate their document tasks more efficiently. By using Artificial Intelligence (AI) and Natural Language Processing (NLP), we can streamline tasks like extracting information from invoices and receipts, which are often done manually and prone to errors. In this project, we'll create an AI model that can reliably pull out the important details from these documents. We'll also use a smart learning method called Few-Shot Learning, which allows us to customize the model for different business needs using just a few examples. Plus, we'll make sure our system works in multiple languages by using a diverse dataset. dida will lead the project, while the Machine Learning department at HU Berlin will provide research support.

The AMoKI project (Autonomous Mining Monitoring in Loose and Solid Rock based on Remote Sensing Data for Sustainable Land Management) has officially commenced, aiming to provide groundbreaking solutions for land management in Germany. Given the high demands for sustainable land management and ensuring the supply of stones and earth resources (S&E resources), the development of a (semi-)autonomous system is crucial. Current methods for capturing extraction areas and statistics in many federal states are inadequate for modern land management. Lack of transparency and traceability, as well as high potential for digitization, are making this project really interesting. The AMoKI project aims to address these gaps. Through industrial research of a machine learning (ML)-based system, authorities' and satellite imagery data are utilized to conduct extraction monitoring of S&E extraction operations. The AI system enables the capture, pattern recognition, and tracking of land use and changes in land use by S&E operations, including consideration of 3D volumes. Thus, the AMoKI project marks an important milestone towards sustainable land management in Germany, contributing to the long-term security of domestic resource supply.

Our ML-Scientist Jakob will give the talk "Semantic Search and ML: Finding the Right Document" on the 18th and 19th of September in Heidelberg, Germany. The talk will provide an introduction to the use of machine learning and pre-trained language models for the development of search algorithms with semantic understanding of the text. He will provide an overview of recent technical advances, such as text embeddings and the use of LLMs. It will also discuss different approaches to implementing customized search algorithms. Parts of the presented methods will be illustrated by concrete customer projects. We are looking forward to this event.

The dida conference, held on May 31, 2024, successfully brought together around 250 AI enthusiasts and professionals keen on expanding their knowledge around machine learning, MLOps and govTech. With contributions from dida but also from Bundesdruckerei, Eversheds Sutherland, appliedAI, or Germany's leading research institutes, guests could enjoy a great mix of theoretical and applied AI topics. Scientists, engineers and decision-makers had the opportunity to attend keynote speeches, take part in interactive workshops or discuss latest AI research in the poster session. To mention only a few highlights: Holger Pannhorst, Head of Data & Analytics at Bundesdruckerei, gave a presentation on ‘New Opportunities: applications of machine learning technologies in the public sector’. He showed how ML applications have already found their way into public processes and how they improve efficiency and transparency. Christian Dürschmied, attorney at Eversheds Sutherland, provided insights to the legal implications of generative AI and large language models, which legal frameworks are already in place, and where EU legislation will likely develop more guidelines in the future. Julius Richter from University of Hamburg spoke about how generative models can enhance audio and speech data. Through his research he proposed a method to learn conditional distributions of clean speech given noisy inputs and explained how a combined approach of predictive and generative AI could significantly improve audio and speech processing technologies. As the dida conference was not just about learning, it was also great to see so many lively exchanges, new connections being made and ending a long day with a little bit of live music and a cocktail. We are really happy about all the positive feedback we already received and would like to thank all participants in making our second dida conference once again a true success. Recordings of the talks and further impressions of the day will follow shortly.

Two weeks ago, we had the pleasure of hosting our partners from the KAMI project at dida. Representatives from the Leibniz Institute for Agricultural Engineering and Bioeconomy (ATB), the University of Hildesheim, the Teaching and Research Institute for Animal Breeding and Animal Husbandry (LVAT), and Wille Engineering joined us to discuss the progress and next steps of our joint research project on monitoring dairy cow respiratory rates. This initiative aims to enhance animal welfare through advanced AI and sensor technology. As the project nears its conclusion, our discussions centered on ensuring a successful wrap-up and maximizing the project's impact on dairy farming. The KAMI project is dedicated to improving animal health and well-being, providing an early warning system for stress and illness detection. We take pride in our collaboration and in working together to advance precision farming and animal welfare.

Our paper „An optimal control perspective on diffusion-based generative modeling” has been accepted at the journal Transactions on Machine Learning Research (TMLR), one of the leading journals on machine learning research. The paper establishes a connection between stochastic optimal control and diffusion based generative models. This perspective allows to transfer methods from optimal control theory to generative modeling and e.g. allows to derive evidence lower bound as a direct consequence of the well-known verification theorem from control theory. Further, a novel diffusion-based method for sampling from unnormalized densities is developed - a problem frequently occurring in statistics and computational sciences.

We are happy two announce that two papers with dida contribution will appear in the proceedings of this year’s International Conference on Learning Representations (ICLR), which will take place in Vienna in early May. Next to NeurIPS and ICML, the ICLR is one of the three main conference on machine learning. The paper Improved sampling via learned diffusions investigates how sampling from specified probability densities can be approached via diffusion processes in which the drift is learned via neural networks. A perspective from measures on path space is proposed which allows to generalize existing concepts such as Schrödinger bridges or diffusion based generative modeling. This new perspective in consequence allows to design more efficient loss functions that can lead to vast improvements in numerical experiments. The paper Fast and unified path gradient estimators for normalizing flows suggest more efficient algorithms for the variance-reduced training of normalizing flows for the sampling from complex probability distribution that for instance appear in physics or molecular dynamics. It is further derived how variance reduction methods can be applied to likelihood training which is relevant when target data samples are available. At dida, we are very happy that we can continue our ambition to conduct cutting edge research and contribute to the machine learning community. We can therefore continue our path to combine practical machine learning with innovative research. It is the company’s philosophy that both worlds strongly benefit from one another.

We are pleased to announce that the program for the upcoming dida conference has been published here. This year's event will feature a variety of workshops and talks, focussing on LLMs and MLOps, including notable sessions like Fabian Dechent's "Decision Process Automation with Large Language Models" and Anton Shemyakov's "Flexible training pipelines with Pytorch Lightning and Hydra". You will furthermore find interesting program points dealing with ML applications, research on computer vision-based meta and transfer learning or legal implications of generative AI. Be sure to check the full schedule and join us for a great day of knowledge-sharing at B-Part, Berlin.

We recently visited the Institute for Plastics Processing in Industry and Craft at RWTH Aachen University. Together with our partners, we discussed our joint research project “KIOptipack”, which aims at developing AI-supported tools to optimize product design and production methods for plastic packaging, concerning sustainability. During our visit, we took a tour of the laboratories. This gave us first-hand experience and a better understanding of the complex equipment and data sets that are essential to our work. We talked about the different data augmentation techniques that are specifically designed for various measurement modalities. Unlike the intuitive application of augmentation methodologies in visual datasets, such as images, where techniques like rotations and zooms are common, we conducted a thorough analysis to develop customized strategies for optimal augmentation in our specific research contexts. We want to express our gratitude to the team at IKV for their valuable contributions to our discourse. We're happy to incorporate these insights into our ongoing research efforts and can't wait to share the results!

We are happy to share our key insights from the 2024 Digitaler Staat Conference. Over two enriching days, we had the opportunity to explore fascinating applications, engage in AI topic discussions, and gain valuable new perspectives. Two standout topics caught our attention: A project by Nortal, in which the risk of long-term unemployment among applicants was predicted in the Estonian state during the Corona period to support counselors in the unemployment fund. The project is a technical success and a visionary application. We look forward to discussing the implications of the AI Act with the project members in the near future. The Servicestandard, which forms the framework for standards in software development and commissioning by the public sector and was developed by the DigitalService des Bundes in consultation with the Nationaler Normenkontrollrat (NKR), will be decisive for software development within the framework of the Onlinezugangsgesetz (with the current amending law). The Servicestandard will increasingly guide us through our public assignments. We are looking forward to the Digital State 2025 and will be present at further conferences this year!

Last year's dida conference was a great success, featuring machine learning talks, engaging panels, and insightful workshops from organizations like Google, TU Munich, and codecentric. We were thrilled by the positive feedback, and we're excited to share with you the highlights of last year's event. Click below to watch: Whatch last year's conference highlights here. At this year’s dida conference, you will get the opportunity to once again receive input to the latest advancements in machine learning. Here's what you can expect: ◾ Talks covering key ML areas like LLMs and MLOps ◾ Hands-on workshops to actively discuss technical ML aspects in more detail ◾ Time for networking with fellow enthusiasts and industry experts ◾ Delicious cuisine to fuel your brainpower ◾ Plus, a bit of live music and refreshing cocktails to wrap up the day in style. And the best part? We've still got some open slots for registration, but they're filling up fast. You can secure your spot for free right now: Register for the conference here: https://dida.do/conference 𝐃𝐚𝐭𝐞: May 31st 𝐋𝐨𝐜𝐚𝐭𝐢𝐨𝐧: B-Part Berlin We're eagerly anticipating your presence at the dida conference 2024.

Our CTO Lorenz Richter visited the group of Anima Anandkumar (director of machine learning research at NVIDIA and Bren professor of computing) at Caltech in Los Angeles and gave a talk on diffusion-based generative modeling. It was a great opportunity to exchange on cutting-edge AI trends and to get some California sun. In his presentation, Lorenz delved into the intricacies of diffusion-based generative modeling. The talk established a connection between generative modeling based on Stochastic Differential Equations (SDEs) and three foundational pillars of mathematics: stochastic optimal control, Partial Differential Equations (PDEs), and path space measures. This presentation reflects dida's dedication to pushing the limits of mathematical research, cultivating collaboration among diverse fields of study, and propelling tangible progress.

Lorenz Richter (Co-Founder & CTO of dida) followed the invitation to speak at the Measure Transport, Diffusion Processes and Sampling Workshop at the Flatiron Institute in New York City. The workshop brought together experts on the mathematical aspects of diffusion based generative modeling and was a great opportunity to exchange with colleagues. In his presentation, Lorenz delved into the intricacies of diffusion-based generative modeling. The talk established a connection between generative modeling based on Stochastic Differential Equations (SDEs) and three foundational pillars of mathematics: stochastic optimal control, Partial Differential Equations (PDEs), and path space measures. This presentation reflects dida's dedication to pushing the limits of mathematical research, cultivating collaboration among diverse fields of study, and propelling tangible progress. For more information about the talk, you can visit: https://events.simonsfoundation.org/event/335b8972-e63d-4672-82f3-4dd95868cb40/summary

Philipp Jackmuth got invited to the latest episode of "In Numbers We Trust - Der Data Science Podcast" by INWT Statistics GmbH In conversation with INWT's Managing Director Amit Ghosh, the two talked about machine learning projects and how they differ from projects in the field of data science and classic software development. Philipp shared success factors that dida has been able to identify again and again in past projects, the importance of multidimensional metrics in the evaluation of machine learning models and strategies for dealing with black box models. We would like to thank our colleagues from INWT for the invitation and recommend that you listen to this and other very exciting episodes of the podcast. You can find the podcast on all common platforms as well as here: https://inwt.podbean.com/

On October 16, 2023, our managing director Philipp Jackmuth, gave a talk to the Advisory Board of the Faculty of Business and Social Sciences at the University of Hohenheim. The presentation centered around real-world applications of Artificial Intelligence (AI). The ensuing discussion expanded to the seamless integration of AI advancements into the curriculum, with all participants unanimously underscoring the critical importance of nurturing critical thinking as an essential component of university education.

On October 10th the FAZ Institute and German AI Association honored German outstanding AI solutions. The AI Innovation Award, recognizes pioneering achievements in the field of Artificial Intelligence. It seeks companies utilizing AI technologies to positively impact their own or other companies, their industry, and society at large. The solution to automatically analyze service charge statements developed for Mieterengel made it into the final round (top three) of the award. It enables tenants to swiftly and efficiently verify the legality of their annual cost bills by simply uploading a photo. Since its usage in 2021, the solution has substantially reduced review times for lawyers by 60%, serving over 10,000 customers. Tenants receive a free initial assessment of their utility bills, empowering them to make informed decisions by offering a cost-effective alternative to traditional legal consultations. Thus, Mieterengel is democratizing access to legal resources, ensuring more tenants can confidently advocate for their rights. We are pleased that the committee recognized the innovative and positive aspects of this solution.

On October 5, the AI Accelerator Institute hosted its Generative AI and Computer Vision Summit at the Waldorf Astoria in Berlin. Technical leaders from amongst others Airbus, AstraZeneca, GSK or Bosch presented aspects of their latest work and findings related to generative and computer vision models. As we regularly attend AI-related conferences and often present our findings ourselves, it was once again an interesting day. Talking to other developers about their projects allows us to learn about the state of the art in the industry and compare our work to theirs. It was good to see that our approach of actively participating in AI research and our internal R&D efforts put us in a good position to help shape future developments in generative AI and computer vision.

On September 28, David Berscheid (virtually) spoke at the OGS World Café in Koblenz on the topic of "AI - Hype vs. Real Value" and provided insights into text and image-based AI projects that are creating real value in production settings. In addition, he offered dida's perspective on current developments and explained what we believe is and will be relevant for companies faced with plans to develop and implement AI solutions. Hosted by OGS GmbH, a Koblenz-based IT company specializing in ERP software, the OGS World Café put together a program consisting of informative presentations and discussions around the topic of AI in the ERP context, among others. We are pleased to have been able to contribute to this year's World Café and are looking forward to the next one.

こんにちは (Konnichiwa) Lorenz Richter (Co-Founder & CTO of dida) followed the invitation to speak at the ICIAM 2023 in Tokyo. This gathering, known for its significance in uniting professionals in applied mathematics and related fields, provides a platform for insightful discussions and knowledge exchange. In his presentation, Lorenz delved into the intricacies of diffusion-based generative modeling. The talk established a connection between generative modeling based on Stochastic Differential Equations (SDEs) and three foundational pillars of mathematics: stochastic optimal control, Partial Differential Equations (PDEs), and path space measures. We are happy that we could once again contribute to latest research and and demonstrate our commitment to innovation in the field of machine learning. This presentation reflects dida's dedication to pushing the limits of mathematical inquiry, cultivating collaboration among diverse fields of study, and propelling tangible progress. For more information about the talk, you can visit: https://iciam2023.org/registered_data?id=00087#04799

On August 31, another GovTech Campus Community Meetup took place in Berlin. This regular series of events aims to connect the federal, state and local governments with the most innovative players in the tech scene, civil society, the open source community and applied research. The aim is to promote Germany as a pioneer in the development and application of digital technologies and solutions for the public sector w.r.t. government technology. As a member of the GovTech Campus dida had the privilege to participate in this event. Our team member Axel Besinger used the opportunity to exchange ideas with experts from industry and the public sector. The highlight was the presentation by Bernd Schlömer, State Secretary for Digitalization at the Ministry of Infrastructure and Digital Affairs of the State of Saxony-Anhalt, as well as the State Government's Commissioner for Information Technology (CIO), on the topic: "Platform capability - What does it mean for the German administration in the future?". It was exciting to hear how agile Bernd Schlömer is in building up his own administration platform in Saxony, with the goal of integrating citizen-relevant product components (whether developed in-house or purchased) into this platform in order to bring Saxony, and thus Germany, one step closer to digitization.

Successful Innovation Day with Vereinigte Volksbank Raiffeisenbank eG! Last Friday, thanks to Christoph Baier, we were able to welcome Vereinigte Volksbank Raiffeisenbank eG (VVRB) to smartextract.ai / dida as part of their Innovation Day on the topic of artificial intelligence. In an exciting discussion, we explored the diverse application scenarios of information extraction in the financial sector. The innovative spirit and great interest in new technologies under the leadership of Mr Ralf Magerkurth, CEO of Vereinigte Volksbank Raiffeisenbank eG, impressed us deeply and should serve as a role model for others. The entire smartextract.ai and dida team says THANK YOU for the opportunity to present our ideas and solutions and to meet this impressive team. See you at the next Innovation Day!

Our ML Scientist & Project Lead, Jakob Scharlau, attended a great event at the Google office in Berlin, hosted by KI Park e.V., Google, and pacemaker® by thyssenkrupp. Under the title "AI Innovation: LLMs, Federated Learning and Slime Mold Algorithms", it focused on advanced methodologies of Artificial Intelligence and their exciting implications for our future. It featured two keynote sessions discussing the complexities of operationalizing Large Language Models and the role of #AI in demand forecasting using federated learning and slime mold algorithms. Overall, it was an insightful and exciting evening dedicated to the future of AI, offering valuable insights and opportunities with industry leaders. We are happy to regularly participate on such events, learn from our colleagues and offer some of our insights in exchange.

We are happy to announce that the paper „Improved sampling via learned diffusions“ has been accepted at the workshop „New Frontiers in Learning, Control, and Dynamical Systems“ which takes place at the International Conference on Machine Learning (ICML) this week in Hawaii. The paper investigates how sampling from specified probability densities can be approached via diffusion processes in which the drift is learned via neural networks. A perspective from measures on path space is proposed which allows to generalize existing concepts such as Schrödinger bridges or diffusion based generative modeling. This new perspective in consequence allows to design more efficient loss functions that can lead to vast improvements in numerical experiments. At dida, we are very happy that we can continue our ambition to conduct cutting edge research and contribute to the machine learning community. We can therefore continue our path to combine practical machine learning with innovative research. It is the company’s philosophy that both worlds strongly benefit from one another.

Together with our partners Ostfalia University of Applied Sciences, A&B Peine and GE-T, dida started to work on the research project „Wertstoff-Informations-System mit Künstlicher Intelligenz”, which is funded by the Federal Ministry for the Environment, Nature Conservation, Nuclear Safety and Consumer Protection. Our common goal is to develop an AI-powered mobile guidance application that helps to find the right disposal method for products. Further information can be found on the project website.

Yesterday, we followed the invitation of Bitkom to speak at the hub.berlin conference about the topic of AI co-creation between startups and corporates. In his presentation, David Berscheid spoke about three factors that, in our experience, are important for successful co-creation projects: Setting realistic scopes and expectations, the repeated integration of domain expertise, and a strong focus on cultural and communication aspects between the startup and the corporate. The conference was a real success and focused mainly on the topics of tech and digitalisation. We saw some great speeches such as the opening keynote from Tim Höttges (CEO of Deutsche Telekom) or 'How Generative AI Redefines Business Decision-Making' from Martin Heinig (Head of New Ventures & Technologies at SAP). Perhaps most exciting for us as NLP specialists and daily users of pre-trained transformer models was listening and talking to Jonas Andrulis about Aleph Alpha and his plans to develop safe and explainable large language models for Europe. Also, thanks to David Kregler for our moderation!

Letzte Woche hatten wir die Gelegenheit, in die Welt der Automobil-Tests einzutauchen und KI-Potenziale mit Leiter*innen der Test-Abteilungen von Vitesco zu diskutieren. Ziel der vom KI Park organisierten Veranstaltung war es, Unternehmen mit KI-Spezialisten zusammenzubringen, sich gegenseitig vorzustellen und erste Expertenmeinungen zu Anwendungsfällen, potenziellen Herausforderungen und aktuellen KI-Entwicklungen zu geben. Für uns war es ein hochinteressanter Austausch, da Testumgebungen in der Automobilindustrie vielversprechende Datenmengen für KI-Automatisierungen bieten, gleichzeitig aber auch viel auf dem Spiel steht, wenn es um fehlerhafte Tests geht. Wir hoffen, dass wir Viteso wertvolle Impulse geben konnten und freuen uns auf deren weitere KI-Entwicklungen sowie auf zukünftige KI-Park-Workshops.

Led by Bitkom e.V., we teamed up with other Machine Learning / AI experts to highlight ways in which SMEs can successfully deploy algorithmic solutions within their businesses. The whitepaper discusses different forms of collaboration, potentials and challenges, as well as success factors that the co-authored startups have discovered on their journeys. You will also find 8 use cases, where we, as well as other Machine Learning-related companies present concrete examples of successful collaborations with SMEs in the past. We would like to thank Bitkom for coordinating and finalizing the whitepaper! You can find the whitepaper here: https://www.bitkom.org/Bitkom/Publikationen/Co-Creation-and-Collaboration

Last week, David Berscheid gave a speech at the co³ - corporate content conference - in Munich. In his keynote on "Successful application of AI: Best practices from 5 years of AI projects", he presented some of dida's insights on how to identify the right problems and processes within an organization to solve using AI, how to implement solutions most effectively, and the role of corporate culture. The conference was aimed at marketing and communication professionals, who are faced with a rapidly changing environment in their field of expertise. Apart from dida’s best practices, the conference also dealt with ethical and legal aspects arising with the advances of generative AI. Other valuable contributions came from organizations like IBM iX, Technical University of Munich or German Entrepreneurship. We’d like to thank the corporate-content.com team for the invitation and the great organization!

On May 11th, the KI Park e.V. Summer Event 2023 was held at Marienpark (Berlin, Germany). This event aims to foster connections within the artificial intelligence ecosystems in Germany and Europe. As a member of KI Park e.V. dida had the privilege of participating in this event. Our team members, Augusto Stoffel and Axel Besinger, took the opportunity to engage with industry experts and showcase dida's latest work in Natural Language Processing and Computer Vision. The event was a resounding success, featuring esteemed speakers, panelists from the AI research community and practitioners.

On June 14th, David Berscheid will give a talk at the Corporate Content Conference - CO³ - in Munich. The topic of the conference is phrased as “AI - The Future of Corporate Content” and is directed at marketing experts of large corporations. In his talk, David will speak about best practices on how to successfully apply AI, based on the experience of 25+ projects at dida. We’re looking forward to the conference, to lots of interesting talks and panels, as well as meeting many new contacts. Ticket can be bought here.

Last Friday we hosted our first machine learning conference at B-Part Berlin. It was a full day of talks, workshops, a panel discussion, as well as time to chat and network. We hope that our guests and speakers enjoyed the event as much as we did and we are happy that 150 - 200 people came to celebrate our 5th anniversary - much more than we initially expected. We have attached some impressions of the event for those who could not be there. We are also already working on the video footage, which will be online soon. A big thank you to everyone who has supported us on our way and will continue to do so!

How can administration become future-proof? Our managing director, Philipp Jackmuth, attended an event on March 28, 2023, organized by GovMarket and PwC in Berlin, which addressed this interesting topic. In addition to contributions from municipalities, former State Secretary of the Federal Ministry of Finance, Volker Halsch, and Ekin Deligöz, Parliamentary State Secretary of the Federal Ministry of Family Affairs, Senior Citizens, Women and Youth (BMFSFJ), also spoke. Ms. Deligöz emphasized that by improving conditions for child care and elderly care, a workforce potential of up to one million full-time equivalents could be raised. Other speakers focused on the possibilities of increasing digitization and process automation to make administration more future-proof. Everyone agreed that action must now be taken so that after the impending wave of retirements in the coming years, the administration remains functional.

The authors Frank Weilandt, Lorenz Richter, Tiago Sanona, and Jona Welsch from dida published a paper about the early classification of crops using satellite data together with our cooperation partner FERN.Lab, GFZ Potsdam. The classification of crops is a well-known problem in Remote Sensing and is the basis for applications in public and private sector like yield forecasts. The paper presents a deep learning based method. This method allows predictions at arbitrary time points during the year with a single model, as well as the seamless combination (fusion) of different data sources (e.g. optical and radar data). The findings in this paper are based on research from the cooperation project "CropClass" with FERN.Lab.

From February 8-10, we had the pleasure of attending the Google Cloud Sunnyvale startup summit, which Google invited us to attend in Silicon Valley. The workshop offered a lot of interesting talks from high-ranking Google employees, such as the VP Databases, the VP Business Development and Growth, the Lead Cloud Quantum Engineer and the Director of Product Management of Google Brain. There was a balanced mix of technical as well as business related topics - a quite inspiring surrounding from the areas of Big Data and Machine Learning. The historical environment did its part to make the workshop a complete success. We look forward to further exchanges with Google and are happy to be listed as a service partner in the Google Cloud.

Last week, we had a great time at the Deutsche Bahn mindbox pitch day! Proud to have been represented by Moritz Besser and Augusto Stoffel. The event was a perfect platform for brainstorming ideas about digitalization at Deutsche Bahn and getting to know interesting teams from Klarso GmbH and Scriptomat. Excited about future collaborations and opportunities. Thanks a lot for the initial invitation!

On January 24th the leam.ai conference took place. It's an initiative to promote large european AI models (leam) led by Bundesverband KI, which aims to establish a European AI ecosystem that takes a competitive position in AI development. As dida is a member of Bundesverband KI, our managing director Philipp Jackmuth seized the opportunity to meet with fellow industry experts and to discuss the iniative's status quo. It was a well organized conference with speakers and panelists from the AI research community, practicioners as well as representatives of the legislative body.

On January 19, dida was co-host of the “KI Park Meetup”. Our ML scientists Angela Maennel and Konrad Mundinger moderated the discussions around the topic of Natural Language Processing. It was a great event and we had many interesting conversations with other ML scientists and engineers. We are already looking forward to the next events of KI Park.

On January 19th 11 am our Machine Learning Scientist Angela Maennel will hold a tech talk explaining how NLP can be used to create user-friendly interfaces. While customers prefer being able to express their needs using free-form text, for processing structured data is often necessary. She will show how this conflict can be resolved using NLP. The talk is based on a project we did together with felmo resolving exactly such a conflict for their website. You can register now for free using the following link: Click here

Our employee Marty gave a talk at the open NLP meetup about "Ethics in Natural Language Processing"! If you wonder what implications large language models have on us as society check it out: https://youtu.be/7Idjl3OR0FY

We are happy to announce that dida has made scientific contributions to this year's Conference on Neural Information Processing Systems (NeurIPS) in New Orleans. NeurIPS is one of the largest and most important conferences on innovations in machine learning and related topics. After a contribution in terms of an accepted paper in 2020 (as well as ICML papers in 2021 and 2022), dida employees contributed a paper and workshop presentation (“An optimal control perspective on diffusion-based generative modeling”) on diffusion based generative models as well as a paper with oral presentation (“Optimal rates for regularized conditional mean embedding learning”) on nonparametric statistics to NeurIPS this year. We are proud to say that dida is able to maintain a position amongst internationally renowned researchers from leading universities as well tech companies. Out of 9634 submissions 2672 (27.7%) papers got accepted and only 184 (1.9%) received an oral contribution. The paper “An optimal control perspective on diffusion-based generative modeling” establishes a connection between stochastic optimal control and diffusion based generative models. This perspective allows to transfer methods from optimal control theory to generative modeling and e.g. allows to derive evidence lower bound as a direct consequence of the well-known verification theorem from control theory. Further, a novel diffusion-based method for sampling from unnormalized densities is developed - a problem frequently occurring in statistics and computational sciences. The paper “Optimal rates for regularized conditional mean embedding learning” solves two open problems related to a commonly used nonparametric regression model: we prove convergence rates for a setting when the model is misspecified (meaning that the ground truth can not be analytically represented by the model) as well as lower bounds for convergence rates of all learning algorithms solving the general underlying problem. In total, these results give conditions under which the convergence speed of the investigated model can not be beaten by any other model solving the problem.

Project Kickoff! We are happy to be part of the research project KIOptiPack, or "AI-Application Hub: Circular Economy for Plastic Wrapping via Artificial Intelligence". Head to our german page to learn more about the project.

dida takes part in the virtual AI-Days of ibi research Institute "Artificial Intelligence in trade and financial services". dida's Machine Learning Project Lead will share real world experiences from AI projects in trade in his talk "Natural Language Processing (NLP) for (un-)structured documents". You can listen to the presentations for free, register under https://ki-tage.de/ (presentations will be in german).

Our Machine Learning Scientist Angela presented our Natural Language Processing project for felmo at the Deep Learning World on the 5th of October. In the questions and answer sessions afterwards it was often asked, if and how one can apply our solution to other domains. While one of the main takeaways was that project specific adaptations greatly improve model performance, the general approach can be applied in many different contexts.

This week dida participated at the leaders circle which happened as part of the Google Cloud Next conference in Munich. We were happy to connect with many interesting companies which aim to advance internal process via efficient data management and digitalization. It turns out that more and more companies rely on machine learning in order to improve their business processes or to develop novel innovative tools. With their flexible and powerful cloud service, the Google Cloud indeed provides a platform that is highly suitable to approach the related challenges and dida is proud to be an official Google service partner.

Our Machine Learning Project Lead Jona Welsch presented our research project CropClass during the workshop "geodata as open data for artificial intelligence" at the Annual Conference of the German Informatics Society in Hamburg on September 27th, 2022. There was a lively exchange in a following panel discussion with presenters from industry and research, especially regarding the usage of freely available satellite data from the ESA Copernicus program.

We will give a talk at the Deep Learning World conference on the 5th October 2022. In Berlin our ML Scientist Angela Maennel will present our NLP project for felmo, a company offering home vet visits. The goal was to extract symptoms, diseases and requested services from customer requests. This data felmo then can use to improve internal processes. You can learn more about the project in our case-study.

In July, dida was participating at this year’s International Conference on Machine Learning (ICML 2022), which - after two years of online events - was happing on site in Baltimore, USA. It was a pleasure to have presented our work on robust SDE-based PDE solvers, which got accepted as a conference paper. You can find the recording of the presentation here. It was good to get connected with cutting edge machine learning research again in a personal manner and exchange with researches from around the world. We could in fact see that robust machine learning and mathematically rigorous analyses of blackbox algorithms become more prominent - a trend that we expect to continue and that fits well to dida’s philosophy of getting deep into deep learning.

dida presents at this year's Data Natives Conference (August 31st - September 2nd). Jona Welsch, Machine Learning Project Lead at dida, presented on the Topic of "Information Extraction - Key learnings from real NLP projects". Information Extraction and Natural Language Processing (NLP) have a large variety of business applications, ranging from the digitalisation of invoices or purchase orders to the automatic interpretation of free-form text. The talk contains key learnings and insights from projects for industry clients and public institutions, which were successfully brought into production. For example, we will talk about the pitfalls when processing PDF-Documents or where to effectively use Graph Neural Networks.

The best feeling is when the work you do is appreciated and used by others. Even better it is if it also helps in a philantropic way. Because of that we are very glad that researchers approached us and asked if they could use our ASMSpotter to see how one can protect the Amazon rainforest better. Here a description of the researcher's work: "Artisanal gold mining is a growing driver of deforestation in the Amazon and the principal driver in Guyana where our research is focused. Assisted forest restoration and natural regeneration are crucial to offset the loss of biomass following such land-uses, however, little is known about the impact of gold mining on natural regeneration compared to traditional forms of deforestation in the Amazon such as agriculture expansion. Some mines show no signs of natural recovery and need assistance, while others regenerate naturally. So we asked, under what conditions do mines recover naturally? More specifically we chose to focus on whether there is a relationship between mine geometry (whether it is long and thin or wide and circular) and the rate of regeneration. We saw that dida have been developing a powerful tool that uses deep learning to generate near real time predictions of mining areas and their shape and size from satellite imagery. We reached out to them for a collaboration and are now using their models to build up a dataset of mines of different shapes and sizes in Guyana which we will use for analysis." - Elmontaserbellah Ammar BSc, MSc candidate, Durrell Institute for Conservation and Ecology, University of Kent, Canterbury, UK

We are happy to announce that a recent paper which dida contributed to has been accepted at this year’s International Conference on Machine Learning (ICML). The ICML is one of the most renowned conferences on machine learning innovations, on which multiple universities, but also companies like Google, Facebook and DeepMind frequently publish their latest findings. This year, 1235 out of 5630 submissions have been accepted, which yields an acceptance rate of 21.9%. At dida, we are very happy that we can continue our ambition to conduct cutting edge research and contribute to the machine learning community. After a paper at NeurIPS 2020 and a paper at ICML 2021, which received one out of five outstanding paper awards, we can therefore continue our path to combine practical machine learning with innovative research. It is the company’s philosophy that both worlds strongly benefit from one another. The current paper can be found here. It proposes and rigorously analyzes more robust and more efficient algorithms for solving linear Kolmogorov type partial differential equations (PDEs). Such PDEs are relevant in multiple applications for instance in finance, physics, reinforcement learning or diffusion based generative modeling. If you have any questions, do not hesitate to contact us.

Our team of three dida people took part in the ESA Living Planet Symposium in May 2022 in Bonn. dida’s ML Scientist Dr. Frank Weilandt presented our project on "Early crop classification via multi-sensor fusion and temporal attention", a research project with the German Research Centre GFZ Potsdam. We also met our fellow Remote Sensers from around the world to have an exciting exchange about new developments in Earth Observation.

We will give a talk at the ESA Living Planet Symposium in May 2022 about classifying crops from satellite data. In Bonn, our ML Scientist Frank Weilandt will present our project on "Early crop classification via multi-sensor fusion and temporal attention". We are developing software which can recognize crops growing on agricultural parcels from satellite data, e.g. wheat or maize. One central goal is the classification during the year of observation. This is joint work together with the GFZ Potsdam (German Research Centre for Geosciences).

This Thursday (24th of March) our ML Consultant Moritz Besser will give a short talk on Machine Learning in real projects at the "Technologiesalon" of the GEOkomm Verband der GeoInformationswirtschaft Berlin-Brandenburg e.V. Participation is free and does not require any registration. The event is held in German. You can attend via the following link.

We’ve been accepted to give a talk at AMLD in Lausanne, Switzerland on the LaserSKI project. Together with a consortium of manufacturing companies we have been working on an approach to automate the process of detecting defects in diode laser wafers using a FasterRCNN object detection model. Our Machine Learning Scientist William Clemens, PhD will be speaking on the AI and manufacturing track and show how AI can be used to optimise quality assurance in manufacturing.

At the start of the year our ASMSpotter project was featured by the UNESCO. Now the reknowned Forbes magazine wrote an article about the Top 10 list covering our tool. With the help of the ASMSpotter it is possible to detect informal mining sites in the rainforest and is co-developed with our partners from Levin Sources. Currently, it is getting deployed in Guyana. You can learn more about the project here.

Let us refer to the a new research paper on a machine learning attempt for solving high-dimensional PDEs that has recently appeared on Arxiv and has been done in collaboration with dida. In the paper, we analyze existing algorithms that are based on residual minimizations („physics informed neural networks“) or backward stochastic differential equations („deep BSDE“) and suggest a new method that is based on the novel diffusion loss. Crucially, this method can be interpreted as an interpolation between the two existing methods that, at the same time, offers certain numerical advantages. Those advantages may allow for more efficient and more accurate PDE approximations and are therefore likely to find further applications whenever high-dimensional PDEs have to be solved.

Together with our friends and partners from Polyteia and K'UP we will organize a Meetup session on the topic of how public sector and private enterprises can use environmental data to make their communities and businesses more resilient against the ever increasing risks posed by the climate crisis. We will be happy for you to join the discussion on the 10th of February so we all can learn from each other through different perspectives! (The first event will be held in German)

We presented our ongoing project LaserSKI on the 10th of January at the All-Hands-Meeting for BMBF funded AI research projects. It was an interesting experience doing a Poster presentation online, but thanks to the good preparation of the host, Munich Center for Machine Learning, a positive one.

The UNESCO's International Research Centre on Artificial Intelligence awarded our solution ASMSpotter as one out of ten outstanding AI solutions to achieve the SDG 12: "Responsible Consumption and Production" and SDG 15: "Life on Land". Together with our Partners Levin Sources we help detect informal mining sites in the Amazon rainforest to make mining more sustainable. Find out more at our ASMSpotter project page.

Together with our partner idealo, dida will present a Machine Learning solution to keep a huge product database up to date. Dr. Jan Anderssen (Domain Lead Inventory Business at idealo) and Jona Welsch (Machine Learning Project Lead at dida) will show how state of the art Deep Learning methods (such as Transformers, BERT) were used in connection with smart text matching in order to extract relevant information from free-form text. They also explain how the existing database was leveraged in order to create an automatically labelled training dataset, which reduced the required amount of manual labelling drastically. The developed model enables continuous automatic updates of idealos database.

The availability of Remote Sensing data and especially satellite data has seen a strong increase in the last years. With increasing data volumes, manual evaluation of these data becomes less efficient. Machine Learning methods are predestined to bridge this gap between data availability and need for evaluation expertise, making it possible for a larger user group to extract information from satellite data and apply this information in an enterprise context. In the upcoming webinar, Moritz Besser (Machine Learning Consultant) and Jona Welsch (Machine Learning Project Lead) will give an overview of the different types of available satellite data, Machine Learning methods used for their evaluation, as well as practical use cases. Register now!

Together with our project partners Bond3D and Fraunhofer IPT, dida has kicked off a 2-year research project funded by the Federal Ministry of Education and Research. The goal of the project is to increase the efficiency is additive manufacturing ("3D printing") by applying modern machine learning algorithms to the various data collected from the Fused Decomposition Modelling machines of Bond3D. By interlinking sensory and image recordings of the production process, rich data sets are collected, which are suitable for the application of machine learning algorithms. For example, dida will investigate how these data can be mined to set process parameters and to predict product quality during the production process. We are very much looking forward to the cooperation with our project partners and to delve into this exciting research project.

On 20th of October dida participated at the Lacos Innovation Day, a fair focussed on bringing together leading tech innovators in the agricultural sector. Though new to this economic field, we were welcomed very well and saw that the potential of AI in agriculture is recognized by the community. We are looking forward to continue the talks with our new contacts to bring machine learning to the world of food production. See here the local television report (German).

A new research project in agriculture started: in the KAMI project AI technologies help improve animal well being and health while improving yields and reducing the use of animal drugs. Computer vision techniques will be used to measure the breath of diary cows recorded with infrared cameras. Under the lead of the ATB Potsdam the team of dida, Stiftung Universität Hildesheim, the Lehr- und Versuchsanstalt für Tierzucht und Tierhaltung e.V. and the ITEMA GmbH work together on the project Here you can find further information.

The progress made in machine learning (ML) in the last 10-15 years is so impressive that many companies in Germany have now also set up their own departments for this area. We have had the privilege of supporting some of these companies in recent years, for example in the transfer of proof-of-concepts (POCs) to production. In our upcoming webinar, Philipp Jackmuth (Managing Director of dida) and Dr. Petar Tomov (Machine Learning Project Manager) will share their experiences on the decisive factors that distinguish successful from failed ML projects. Sign up here.

Last week, we were honored to be invited to the Applied Machine Learning Days (AMLD) Africa, organized by the Swiss Federal Institute of Technology (EPFL) in Lausanne. In contrast to last years AMLD, where our machine learning scientist William Clemens gave a talk, this years edition focussed on machine learning solutions designed to tackle problems on the African continent as well as introducing and connecting innovators from there. Moritz Besser, machine learning consultant at dida, introduced the audience to the ASMSpotter, which detects informal artisanal mining (ASM) sites on satellite imagery using artificial intelligence. As ASM is a rising problem on the continent, the tool may help to formalize new sites and therefore make minerals sourcing ecologically and socially more sustainable.

Augusto Stoffel, machine learning scientist at dida, delivered a talk on graph neural networks and their application for information extraction in this year's edition of EuroPython. EuroPython is one of the major Python community conferences, with over 1200 participants. It includes one session dedicated to data science topics. The talk aimed at a mostly technical audience and introduced the basic ideas around graph machine learning. Enriching this conceptual viewpoint, use cases derived from our experience developing ML projects that require table understanding capabilities were also covered. Do you have a potential machine learning application and would like to discuss the possibilites with a specialist? Then book an ML talk with us!

We are happy to announce that this year’s ICML paper by our CTO Lorenz Richter got awarded as an outstanding paper. Only 5 out of 1184 accepted and 5513 submitted papers (so roughly 0.1%) got this honorable mentioning. The other winning papers can be found here. We are happy that dida can continue to contribute to cutting-edge machine learning research which is also highly relevant for practical applications. The paper suggests an efficient numerical solver for high-dimensional parabolic partial differential equations, which are relevant for example in engineering, physics and finance. It relies on the tensor train format and contrasts this approach to deep neural networks as an alternative method. You can find the paper here and the video presentation here.

dida was part of an expert round at this year's Machine Learning Week Europe from June 14 - June 16th 2021. Jona Welsch, Machine Learning Project Lead at dida, presented on the Topic of "Information Extraction - Key learnings from real NLP projects". In interesting panel with other industry experts he afterwards also discussed important aspects of productionising NLP models. Information Extraction and Natural Language Processing (NLP) have a large variety of business applications, ranging from the digitalisation of invoices or purchase orders to the automatic interpretation of free-form text. The talk contained key learnings and insights from projects for industry clients and public institutions, which were successfully brought into production. For example, we talked about the pitfalls when processing PDF-Documents or where to effectively use Graph Neural Networks.

Live webinar: "NLP trend: profitably integrating neural language models into business practice". Neural language models, such as Google's BERT, are large neural networks that have been trained using large collections of text and have thus learned semantic and structural relationships between terms in a language. Neural language models are currently the dominant development in the NLP field with many practical applications, from neural search and question answering to information extraction. There are a number of freely available pre-trained models. In order to use them effectivelyy in one's own business practice, some things have to be taken into account. Which ones exactly and how to do that, we will show in our webinar after an overview of how neural language models work and examples of their possible applications. How do neural language models work? Application examples from the areas of search and question answering How can a neural language model be optimized for my business/application area? How does the integration with my data and applications work?

At the virtual technology brunch of GEOkomm e. V. (the association for geoinformation economy in Berlin/Brandenburg) on 4 June 2021, Jona Welsch, Machine Learning Project Lead and Robert Heesen, Director Product were invited to share dida’s experiences under the topic “Information Extraction with AI - tips from expert for a successful implementation”. Information Extraction and Natural Language Processing (NLP) have a large variety of business applications, ranging from the digitalisation of invoices or purchase orders to the automatic interpretation of free-form text. The talk contained key learnings and insights from projects for industry clients and public institutions, which were successfully brought into production. For example, we talked about the pitfalls when processing PDF-Documents or where to effectively use Graph Neural Networks.

We are happy to announce that dida will contribute a paper and a long presentation to this year’s International Conference on Machine Learning (ICML), which is one of the most popular conferences on artificial intelligence worldwide. Out of 5513 submissions only 166 (i.e. roughly 3%) got offered this opportunity. Our paper suggests an efficient numerical solver for high-dimensional parabolic partial differential equations, which are relevant for example in engineering, physics and finance. It relies on the tensor train format and contrasts this approach to deep neural networks as an alternative method. You can find a preprint version here. After a paper at last year’s NeurIPS conference, for which you can find a short video presentation, we are pleased that we can continue to foster our ambition to combine practical machine learning with cutting edge research. It is the company’s philosophy that both worlds can strongly benefit from one another. If you have any questions regarding the papers or other topics, please do not hesitate to contact us.

Ewelina Fiebig and Konrad Schultka, Machine Learning Scientists at dida, will present one of our new projects - "PreTrainAppEO" - at the 3rd Symposium on Applied Satellite Earth Observation "New Perspectives in Earth Observation" organized by DLR. The goal of the project is to develop a new methodology using the approach of pre-trained AI models to achieve generalizability to various standard applications in Earth observation and remote sensing. The project is realized in cooperation with the Remote Sensing Methodology Chair of the TU Munich. The presentation will take place on June 16 at 13:10 in Session 7: AI - New Methods.

Since May 2021 dida works together with the Remote Sensing Methodology Chair (LMF) of the Technical University of Munich (TUM) on a project "PreTrainAppEO" funded by the German Federal Ministry for Economic Affairs and Energy (BMWI). The project is granted with an amount of 199k EUR and is designed for a period of two years. The name "PreTrainAppEO" stands for Pre-Training Applicability in Earth Observation. Earth observation and remote sensing have become an important part of, for example, transportation and utility sectors, as well as earth sciences and meteorology, due to technological advances in space and imaging technologies and sensors. At the same time, the use of machine learning (ML) methods in the analysis and interpretation of image and sensor data is becoming increasingly important. This is also very evident from some of our projects: ASMSpotter, CropClass, Estimation of solar roof capacity, 4D Urban Insights, DWD Clouds. Although ML approaches are characterized by very high performance and accuracy, their performance is limited to the specific use case and training dataset. As a result, the training effort and the acquisition of large datasets must be repeated for each new use case, leading to high costs in application development. Unfortunately - in contrast to already established pre-trained AI models such as ResNet or VGG, which are based on the generic image dataset ImageNet - there are currently no pre-trained AI models available for satellite data or Earth observation applications. All the more we are happy to be a part of "PreTrainAppEO". The goal of this project is to make the use of AI in the field of Earth observation and remote sensing more attractive and efficient by developing a methodology that uses the approach of pre-trained AI models to achieve generalizability to different standard applications in this field.

At the virtual network meeting of the ZIM network "agrASpace" entitled "Artificial Intelligence in Agriculture" on 4 March 2021, Robert Heesen, Director Product and Jona Welsch, Machine Learning Scientist at dida presented the project "CropClass". The project titled " CropClass- Infra-annual Crop Classification with Remote Sensing Data and Machine Learning" aims to reliably determine crop types on agricultural land as early as possible in the year based on publicly available satellite data such as Sentinel-1 and Sentinel-2 data and to make them accessible via an user interface. The project is being carried out jointly with fern.lab, the remote sensing department of GFZ Potsdam, and will run until June 2022. Potential users of the software are ministries of agriculture for the control of EU agricultural subsidies, sellers of seeds and fertilisers and sprays as well as traders of agricultural goods. For more information, please see the presentation.

dida is represented three times at the DLR Copernicus Forum, which takes place virtually this year from 23-24 March. Our Managing Director Philipp Jackmuth will participate in the panel discussion "Artificial Intelligence - Key to the Copernicus Data Treasure?" together with representatives of DLR (German Aerospace Center), DFKI (German Research Center for Artificial Intelligence), ECMWF (European Centre for Medium-Range Weather Forecasts) and the Federal Ministry of Transport and Digital Infrastructure on 23 March at 2.40 pm. In addition, Moritz Besser, Machine Learning Consultant at dida, will present the project "ASMSpotter - Identification of small-scale mining based on satellite imagery" in the technical session "Energy and Natural Resources" hosted by BGR on 24 March from 2.30 pm. In addition, Robert Heesen, Director Product, and Moritz Besser will be available for questions in a virtual booth. We are looking forward to this year's Copernicus Forum and to dida presenting its projects to an expert remote sensing audience.

dida is again listed in the "AI Startup Landscape 2021" of the appliedAI initiative as one of 278 promising startups with an AI focus in the category "Professional, scientific and technical activities". For this, a "longlist" of AI startups was compiled from public sources such as Crunchbase and LinkedIn. The selection of the startups was done in personal interviews by the appliedAI team to verify their focus on artificial intelligence. In addition, the startups were evaluated by a jury consisting of well-known AI companies and investors such as Google, NVIDIA, HV Ventures and Lakestar. The clustering of startups was based on Shivon Zilis' "landscape of machine intelligence". We are pleased to be represented again this year in the AI Startup Landscape by our very esteemed colleagues from appliedAI.

On 4 January 2021, dida started work on the research project "AuBeSa" within the framework of the BMBF (Federal Ministry of Education and Research) funding programme "KMU-innovativ" with a total project volume of EUR 800k. "AuBeSa" is the acronym for "Automatic detection, measurement and mineralogical classification of tailings heaps and ponds with satellite remote sensing". Waste rock piles are wastes from mining operations that occur during the extraction of raw materials and ores and are usually accumulated near the open pit. Depending on the processing, valuable raw materials are stored, especially in older tailings piles of abandoned opencast mines, for which there was no use at the time. Today, these tailings are becoming the focus of mining companies due to high exploration costs of new economic (primary) deposits and improved processing methods. If exploitable material is found in a tailings by exploration companies, so-called secondary deposits are established and new open pits are built to extract the raw materials. The difficulty in the exploration of secondary deposits, however, is to be able to estimate the economic viability without an on-site study. As high costs are associated with an on-site study, the automatic analysis of remote sensing data can contribute to the assessment of economic viability. In "AuBeSa", an AI algorithm is trained to automatically locate tailings based on a database of known tailings locations. In order to estimate the economic usability, the volume of the tailings is calculated and the mineralogical composition is determined on the basis of satellite data. In addition to dida's AI expertise, geological expertise is therefore necessary for the implementation of the project. We are happy to have the Institute of Mineral Resources Engineering (MRE) at RWTH Aachen University as a strong partner and look forward to continuing the successful cooperation from the ASMSpotter project.

Next Sunday, December 6th, the internationally leading AI conference NeurIPS, which will take place virtually this year, will begin. A total of 9,450 scientific papers were submitted this year, of which 1,900 papers were accepted, an increase of almost 200% compared to 2017. According to an analysis by Criteo, American technology companies and universities dominate the list of conference papers this year as well. In terms of countries, the USA leads by far the list of the strongest AI countries in research with more than 1,186 accepted contributions, followed by China (259) and Great Britain (205). Germany follows in 6th place (70), just behind Switzerland (90). In total, Europe including Great Britain and Switzerland provides 520 expert contributions. A breakdown by institution shows that, in addition to US universities such as Stanford (104), MIT (100) and UC Berkeley (89), it is the American technology companies and their research departments that are leading in AI research. First and foremost Google. Including its London-based subsidiary DeepMind, the US company contributes 237 , followed by Microsoft (95) and Facebook (58). However, the analysis by dida of the German contributions also reveals some glimmers of hope for Germany as an AI location. After all, Germany can also boast top-level AI research. Measured by the number of accepted contributions, the Max Planck Institute for Intelligent Systems (13) based in Tübingen is the leader, ahead of the Technical University of Munich (10), the University of Tübingen (8) and the Max Planck Institute for Computer Science (8) based in Saarbrücken. However, companies such as Robert Bosch GmbH (6) also carry out top-class AI research. Viewed on a map, the south or southwest of Germany is a leader in AI research, with a focus on the university city of Tübingen (22), Munich (13) and Saarbrücken (12). Germany is also well represented at the event itself. Besides Prof. Dr. Ulrike von Luxburg (University of Tübingen) on the board of directors, Prof. Dr. Bernhard Schölkopf (Max Planck Institute for Intelligent Systems, Tübingen) is also represented on the advisory board of the NeurIPS organisation. In addition to Zalando, the Federal Ministry of Education and Research (BMBF) is also acting as sponsor this year. dida also contributes a technical paper at the NeurIPS this year. An article in which dida founder and CTO Lorenz Richter participated was accepted. Lorenz Richter: "We at dida see ourselves as a company at the interface between research and application. For this reason, we at dida are looking for mathematicians and physicists who have many years of experience in research. In this way, we want to contribute to ensuring that AI researchers do not migrate abroad and at the same time are able to apply the results of their research. At the same time, we want to enable our employees to continue doing top-class AI research in order to further advance the development of AI in Germany and not to rely solely on the development of new AI methods by American technology companies and universities". You'll find Lorenz paper here.

At the Sweden Innovation Days on November 17th, the European Landscape of AI startups was presented, including more than 500 startups from Sweden, France and Germany. The initiative was supported by the German Entrepreneurship and the German Accelerator program as well as Vinnova, Sweden's Innovation Agency - appliedAI (Germany), Ignite Sweden, AI Sweden, RISE Research Institutes of Sweden (Sweden) as well as Hub France IA (France). The purpose of the European Landscape of AI startups is to give companies interested in implementing AI solutions guidance which AI companies might be contractors experienced in their application domain. All startups were tested and interviewed in order to investigate their focus on AI technologies and to categorize them into one out of 25 categories. As being part of the landscape, dida is also awarded the "Trusted AI Startup" badge. We are honoured to be part of the landscape and be in the club with other fellow AI startups and hope that more European AI initiatives join the effort to create a pan-European map of AI Startups and to improve transparency in the AI market. You can find dida in the European AI Landscape here: https://www.ai-startups-europe.eu/

Applications of Natural Language Processing such as semantic search (Google), automated text translation (e.g. DeepL) or text classification (e.g. email spam filter) have become an integral part of our everyday life. In many areas of NLP, decisive progress is based on the development and research of a class of artificial neural networks that are particularly well adapted to the sequential structure of natural languages: recurrent neural networks, in short: RNNs. The webinar will give an introduction to the functioning of RNNs and illustrate their use in an example project from the field of legal tech. It will conclude with an outlook on the future importance of RNNs amidst alternative approaches such as BERT and Convolutional Neural Networks. This webinar takes place on November 24th 2020. Sign up here

In this webinar we will present a method based on the BERT model for automated answering of questions. The potential applications are manifold: the ideas of this approach can be used for example in chatbots, information extraction from texts and Q&A sections of websites. As a concrete example, we discuss the extraction of information from biomedical research using the open CORD-19 data set for COVID-19 research. More information here.

dida has won the internationally renowned "Microsoft AI for Earth" prize worth USD 100,000 (link). In the competition, in which solutions for regulating informal small-scale mining were awarded prizes by the American organization Conservation X Labs, dida prevailed with its ASMSpotter software against more than 100 companies from 40 different countries. Informal small-scale mining provides the livelihood for more than 100 million people and is of great importance for the global supply of raw materials. Due to the lack of opportunities for large-scale monitoring and effective controls, however, informal mining often goes hand in hand with major damages to nature and the population. This is where ASMSpotter comes in. ASMSpotter uses AI algorithms to analyze satellite images from the European satellite Sentinel-2 of the European Space Agency ESA from the Copernicus program and from Planet, a commercial provider of satellite images, and automatically identifies informal mining sites and makes changes to the last satellite image visible. The software was developed in the context of a grant from the European Space Agency. In the course of the project, Levin Sources, a UK partner, was acquired to help governments implement programs to formalize small-scale mining so that the solution is not misused and is implemented effectively. Robert Heesen, Director Product at dida: "We feel very honored to be recognized by such an outstanding jury and by Microsoft. With ASMSpotter we want to show that AI can make a positive contribution to the big problems of mankind. ASMSpotter helps to minimize environmental destruction and improve the working conditions of the local population. Until now local authorities lack solutions for automatic, large-scale monitoring of mining activities. Governments using our solution can regulate the sector much better because you can detect changes and act early. Together with our partner Levin Sources, we ensure that the solution is not misused and can develop its full potential. According to Heesen, the next steps will be the introduction of the solution in Ghana, together with the Ministry of Mining in Ghana and the partner Levin Sources. In addition, dida would also like to make the solution available for other raw materials such as cobalt or building materials such as gravel. Heesen: "Technically speaking, the AI solution is well scalable to other countries and raw materials. Now we are looking for partners who will work with us to ensure that the solution is used responsibly and effectively". About dida: dida is an AI company specializing in the development of computer vision applications, i.e. the automatic analysis of image files. dida is headquartered in Berlin and has twenty employees, fifteen of whom work in the development of AI applications. Contact person: Robert Heesen robert.heesen@dida.do +49 151 65266677

Webinar: BERT for Semantic Search | September 24rd from 11-12 o'clock (CEST) During the course of the webinar, we will answer these questions: What is Semantic Search and which problems can I solve with it? What is BERT and how can I use it for Semantic Search? (including a real example) What data do I need for that? Here you can sign up for the webinar.

Together with GFZ Potsdam, the national research centre for geosciences, dida has been awarded a 2-year contract by the Federal Ministry of Economics. The aim of the project is the early classification of crop types on agricultural parcels based on satellite data from the European Satellite system Sentinel using Machine Learning algorithms. The earliest possible classification of the crop type and its availability in an online service will enable a number of additional applications, for example price determination in commodity futures exchanges or the prediction of harvest amounts by crop types. We are looking forward to the cooperation with GFZ Potsdam and an exciting project over the next two years!

In February 2020 dida was selected to present ASMSpotter at the Nvidia Inception Summit / GTC (see Presentation Gallery > Computer Vision / Intelligent Video Analytics / Video & Image Processing). ASMSpotter is an ESA-project developed by dida and the Institute of Mineral Resources Engineering (MRE) at RWTH Aachen. The goal of the project is to automatically locate artesanal and small-scale mining (ASM) sites in the Amazonian rainforest of Suriname. dida was chosen with seven other companies to present their work. Nvidia decided to award our presentation with a Titan RTX GPU by Nvidia. We are very happy to announce that as of today we received the GPU and, of course, in no time upgraded our on-premise machinery. Our gratitude goes out to Alexander Martynau and Nvidia!

-- Thank you for participating in the first webinars! -- In a series of 3 webinars dida's Ewelina Fiebig explains in detail the most important steps leading to the successful implementation of an NLP project. Webinar 1: Text recognition (OCR) | April 23rd from 11-12 o'clock (CEST) Link to the recording in German language. Webinar 2: NLP Labeling Tools | May 28th from 11-12 o'clock (CEST) Link to the recording in German language. We will address the following aspects: What does labeling mean and what can be labeled? Why is labeling necessary at all? Challenges in choosing a suitable tool. Our tips on what is important to consider before and during the selection of a tool. Short presentation of selected tools. Other tools worth mentioning. Webinar 3: NLP Case Study - Entity Recognition in Documents | June 18th from 11-12 o'clock (CEST)

dida's Dr. Petar Tomov and Philipp Jackmuth had been invited to speak at this year's Deep Learning World (DLW) on May 12th. DLW is a conference with a focus on commercial deployment of deep learning. In the talk Petar and Philipp shared learnings with the audience from a recent project done for the European Space Agency (ESA). The project's goal was to detect illegal mines in developing countries from satellite data. We want to particularly thank the audience for the big interest in the topic and many insightful questions. We are already looking forward to being part of the next edition of the conference.

dida successfully completed the ESA project "ASMSpotter". The project investigated the feasibility of automatically identifying artisanal gold mining sites (ASGM) on satellite images using machine learning. In the implementation, dida has worked with both PlanetScope and Sentinel-2 satellite images and, together with the Institute of Mineral Resources Engineering (MRE) from RWTH Aachen University, generated a training data set of more than 10,000 annotated images for model training. For the test area in the Northeast Surinam a classification accuracy of over 80% was achieved. Furthermore it could be proven that the ML model can be applied to areas of similar topography. Due to the positive assessment by ESA, the project will be continued with an ASGM partner in the future. The software will be implemented in a ASGM project in Ghana to identify illegal mining activities at an early stage. We would like to thank ESA and the MRE of RWTH Aachen University for the trustful cooperation!

-- Thank you for participating in the first webinars! -- In a series of 3 webinars dida's Ewelina Fiebig explains in detail the most important steps leading to the successful implementation of an NLP project. Webinar 1: Text recognition (OCR) | April 23rd from 11-12 o'clock (CEST) In this talk we will deal with the topic of text recognition and introduce you to: What does OCR mean? Example of use Why is OCR needed? What OCR tools are available? How are these tools used? Which tool fits to which problem? Here you can find the recording in German language. Webinar 2: NLP Labeling Tools | May 28th from 11-12 o'clock (CEST) We will address the following aspects: What does labeling mean and what can be labeled? Why is labeling necessary at all? Challenges in choosing a suitable tool. Our tips on what is important to consider before and during the selection of a tool. Short presentation of selected tools. Other tools worth mentioning. Here you can find the recording in German language. Webinar 3: NLP Case Study - Entity Recognition in Documents | June 18th from 11-12 o'clock (CEST)

With dida as project lead Leibniz Ferdinand-Braun-Institut, eagleyard Photonics, Coherent DILAS and Innolume have started the research project LaserSKI on April 1st 2020, which is funded by the German Federal Ministry of Education and Research. The goal of the undertaking is to develop a machine learning algorithm for detecting defects at different stages of laser diodes production. Also, a defect classification reference frame will be worked out and made available to the entire Photonics industry. The overall project volume is almost EUR 2m. Among more than 400 submitted proposals LaserSKI is one of 21 successful projects, that were selected in the call for proposals "Artificial Intelligence methods applied". dida is looking forward to an exciting project in a high-tech industry with established companies and research institutes as partners!

Mathematical models cannot cure COVID-19, but they can help us to understand how the virus spreads and to take the appropriate actions to contain the pandemic. dida's external advisor Michael Borinsky contributed to a joint project by Robert Koch Institute and Humboldt-Universität zu Berlin which, amongst other things, provides country-specific predictions of COVID-19 case count numbers for the next 6 days. Please take the predictions with a grain of salt: even scientifically sound predictive models can fail (consider the weather forecast).

dida's Philipp Jackmuth participated in the 2-day conference Digital State, which brought together innovators, modernizers and trendsetters for intensive discussions from March 2nd to 3rd in Berlin. A central topic was the ongoing implementation of the Online Access Act (OZG), which is currently giving a great deal of momentum to administrative digitisation.In this context, the question of what should happen after the implementation of the OZG in order to successfully develop the way into the digital future of administration was discussed. Patron and opening speaker of the congress was once again the Minister of State for Digitisation Dorothee Bär.

It was an exciting day at Copenhagen where Lorenz Richter, CTO at dida, and Wolf Winkler, Senior Consultant at dida, presented the results of a 5-weeks GovTech challenge to the Danish Safety Technology Authority (DSTA). The proposed solution automates significant parts of the workflow of product inspectors based on the AI technologies computer vision (CV), image recognition and natural language processing (NLP). The goal is to free inspectors from administrative work and thus allow them to increase the number of product inspections. The challenge owner's feedback to our proposals was overwhelmingly positive. We thank DSTA and our partners from PUBLIC for the organisation of the challenge and for the warm and welcoming atmosphere. We very much look forward to following up on this. Especially as it will very likely involve visiting beautiful places in Denmark and cooperating with the leading digital public sector and government in Europe.

dida has been listed as one of Germany's AI Startups by the applied AI inititative by the TU Munich. The AI German Startup Landscape poster published on a yearly basis and creates an overview of AI companies active in different industries. In 2020, 247 startups were listed and dida can be found at "Industries" and then "Other". We are proud to be among the startups creating productive AI solutions to our partners and share the appliedAI goal to make AI accesible to the German industry.

On February 27th, an expert discussion on artificial intelligence in earth observation took place at Deutsches Luft- und Raumfahrtzentrum (DLR) in Bonn. From the private sector, companies such as Eftas, GAF AG and Live-EO participated. dida was represented by its managing director, Philipp Jackmuth. Together with representatives from science and the public sector, the participants discussed which conditions can be created to facilitate the use of AI procedures in earth observation. The participants agreed that in addition to a user-friendly data infrastructure, the provision of annotated training data would also be advantageous. Furthermore, successful applications of AI procedures should be made available to other market participants through showcases and an increased exchange of knowledge.

In February 2020 dida was selected to present ASMSpotter at the (Nvidia Inception Summit)[https://events.nvidia.com/inception-summit] on March 6th. ASMSpotter is an ESA-project developed by dida and the (Institute of Mineral Resources Engineering (MRE))[http://mre.rwth-aachen.de/en/] at RWTH Aachen. The goal of the project is to automatically locate artesanal and small-scale mining (ASM) sites in the Amazonian rainforest of Suriname. dida was chosen with seven other companies to present their work.

Agriculture is probably one of the industries with the most potential to apply Computer Vision, as the vegetation progress, vermins and leaf diseases can be observed from satellites, UAV or drones. dida is engaged in the ZIM innovation network agraSpace to exchange ideas on the applications of Computer Vision in agriculture. At Februar 27th, a network meeting was held at Leibniz-Institut für Agrartechnik und Bioökonomie e.V. (ATB) where network partners met to discuss potential projects and search for potential partners for joint funding proposals. Among the partners are research institutions like Leibniz-Institut für Agrartechnik und Bioökonomie e.V. (ATB), Leibniz Centre for Agricultural Landscape Research (ZALF), German Research Centre for Geosciences (GFZ) and private companies like Vereinigte Hagelversicherung, 365FarmNet and dida. Topics of the event included the new funding guidelines of the ZIM program, a presentation of research proposals from Leibniz-Institut für innovative Mikroelektronik (IHP), a presentation from the agricultural practice from Mr. Meise from agrafrisch, a agricultural company from Eastern Brandenburg and a presentation on a project from the BTU Cottbus-Senftenberg on a BMWI funded project on structural change in the Lausitz region.

dida is proud to be chosen for the ESA kickstart activity "4D Urban Insights" in the call for applications that create a social impact. With 4D Urban Insights, dida wants to create an automatic tool to detect urban sprawl and built-up to give urban planners and insurance companies a tool to track city growth in rapidly growing cities. In 2050, 68% of world population will live in cities, up from 30% seventy years ago. 90% of urban population growth will come from Asia and Africa. In Africa, the urban growth is fueled by population growth and migration to urban areas. Due to this, the ten fastest growing cities in next fifteen years will be in Africa. That's why the feasibility study focuses on five African cities located in the Golf of Guinea: Lagos, Abidjan, Luanda, Accra and Douala. We are looking forward to another interesting project together with ESA.

With a group of four dida visited the Applied Machine Learning Days at the SwissTech Convention Center in Lausanne. We contributed to the event through a talk by William Clemens and presented our services on the exhbition floor. Will shared some insights about detecting patterns in geostationary satellite data with convolutional neural networks. We were happy to see so much interest in our use cases and are thankful for many interesting discussions with colleagues from the industry as well as from academia. We will definitely return in 2021!

dida took part in the accelerator program from Deutsche Bahn, called DB mindbox. dida has been selected to be among the three companies, presenting their solution to a challenge to extract information from train documentation to be transmitted to an internal working handbook to maintain trains. On January 28th, Deutsche Bahn invited all teams to their office in Frankfurt to discuss the details of the project and to have workshops with all teams and to give the teams the chance to ask questions to the challenge owners. At the end, all startups presented their approach to solving the problem and were judged by a jury of the DB unit.

Johan Dettmar, Machine Learning Scientist at dida, was invited to present at the yearly event "National Program for Space travel and Innovation" which is hosted by DLR Raumfahrtmanagement in Bonn, Germany. dida was awarded with two ESA Kickstarter projects earlier this year, namely "ASMSpotter", a Machine Learning (ML) based software tool for detecting Artisanal Mining sites in Northeastern Surinam as well as "4D Urban Insights", an ML software tool for monitoring urban development of 5 West African cities. The two projects were presented to various members of ESA and DLR, as well as their academic- and industry partners.

Together with AI Spektrum we have created a poster about Natural Language Processing (NLP). It describes techniques and methods for pattern recognition in natural language. The goal is a direct communication between humans and computers based on natural language. The poster informs about approaches, functions and fields of application. You can download the poster in high resolution here (registration required). Have fun reading!

Robert Heesen, Director Product at dida, will present the ESA kickstart activity "ASMSpotter" at November 28th at the European Space Agency's (ESA) premises in Harwell, UK. The workshop "Space-Enabled Applications for the Mining Sector", hosted by ESA, Satellite Applications Catapult and the United Nations Crime Institute (UNICRI), will address topics such as monitoring and improving mining operations, facilitating social engagement in mining, Environmental Monitoring and funding opportunities for mining activities. Moreover, presentation of case studies from the mining and law enforcement sector and showcases of space-enabled projects and feasibility studies will be presented.

Will Clemens, Machine Learning Scientist at dida, will present one of our recently completed projects for Deutscher Wetterdienst (DWD) at this year's symposium for earth observation in Cologne, hosted by the DLR Raumfahrtmanagement. Under the title, "Machine Learning based detection of thunderclouds from METEOSAT satellite imagery" Will will present a project to improve weather monitoring at airports. The presentation will take place 13th of November at 4 p.m. in session 6a "AI applications".

dida is involved in two ZIM networks, agraSpace and CopServ and will therefore attend the yearly meeting of ZIM networks in Berlin September 24th. This year's topic "AI for the German Mittelstand - together with ZIM networks towards the digital future" is highly relevant to dida as AI will play a major role for the German Mittelstand in the future and ZIM networks provide a good way to cooperate with potential end users and detect use cases from different industries. Recently, ZIM funding has rarely been provided to AI software projects so that the topic of this conference might suggest that for the new ZIM funding cycle starting next year, significant changes can be expected for AI software companies like dida.

As dida has implemented several Remote Sensing applications for clients such as ENPAL and DWD in the past few months, Robert from dida will visit Intergeo, the leading geospatial trade fair in Europe and yearly meeting point for the geospatial community. The Intergeo takes place at the Stuttgart trade fair center from September 17th to 19th. Highlight topics of this years' Intergeo are Smart Cities and Digitalisation of public services, exhibitors include large companies and institutions such as Airbus and the European Commission and lots of SMEs and startups.

The dida machine learning scientists Lorenz Richter and Konrad Schultka participated in the conference Mathematics of Deep Learning, on the mathematical foundations of deep neural networks and their growing applications. Among others it featured talks on equivariant generalizations of convolutional neural networks (Taco Cohen), on connections to the theory of ordinary and partial differential equations (Eldad Haber), on the non-convex loss landscape of deep neural networks (René Vidal), on alternative learning strategies based on Langevin dynamics (Ben Leimkuhler) and on application to inverse problems and parametric partial differential equations (Gitta Kutyniok).

dida is part of agrASpace, an innovation network by GEOKOMM and is taking part in its network event at September 5th in Berlin. The network event will be held at the Forum Digitale Technologien at the Fraunhofer Heinrich-Hertz-Institut. The goal of the event is to network among the 30 partners of the ZIM network and to discuss innovative projects in agriculture that can be realized with satellite and in-situ data. Through its project experience in building Computer Vision applications with satellite data for ENPAL and DWD, dida has plenty of experience in this field and is excited to work in the Remote Sensing sector for the coming years.

On the 27th and 28th of August 2019 dida data scientists Johan Dettmar and Matthias Werner attended the “Train the Trainers: Machine Learning Bootcamp“ at the Google for Startups Campus in Warsaw. The workshop was all about tips and tricks for Google Partners to organize and run Machine Learning Bootcamps for potential and current customers. The aim of the ML Bootcamps is to introduce customers to the world of Machine Learning and Data Science, while presenting useful tools and infrastructure available on the Google Cloud Platform.

dida has been promoted to the "Surge" partner status in the Google Cloud Partner Program. We want to thank Google for their ongoing support and commitment to the partnership. Google Cloud's overall infrastructure, its data and API solutions as well as its computational power give dida the opportunity to build cutting edge machine learning applications.

The platform #AI_Berlin reflects the AI scene in Berlin and serves as a networking hub for the national and international AI community. It illustrates current and future developments, networks, initiatives and experts. They approached our CTO Lorenz Richter and our Managing Director Philipp Jackmuth to comment on use cases for computer vision applications and on the general state of AI in Germany (-> find here the full interview). The aim of AI_Berlin is to connect technology companies, start-ups, research labs and talents on one platform, put Berlin on the map and further strengthen the city as a hub for AI.

Each year the University of Mainz - School of Medicine hosts the DICOM-Conference for radiologists. Next to image data management, the focus was put on the topics of artificial intelligence and data security. dida was invited to give a talk on current research topics in the field of machine learning and pattern recognition in medical images. We were happy about the chance to discuss with radiologists and medical engineers the needs and expectations for process automation and decision support systems in radiology.

dida is going to the first BDLI Network Meeting in Berlin at FABRIK 23 at the 14th of June. The topic of the network meeting is "We are one- together towards the future- How can New Space and Classic Space work together?". As dida works a lot with space data and that more and more applications are run with the help of Machine Learning Algorithms, we are looking forward to interesting discussions and presentations. If you want to meet Robert from the dida team, feel free to reach out!

We are happy to announce that dida has become a member of the state-funded innovation network Space2Agriculture which was initiated by the DLR. Other partners of the network include various institutes from Fraunhofer and DLR, Deutscher Bauernverband, John Deere or KWS Saat. We share the DLR's vision that agriculture will be one of the first sectors that will capitalize on space data with various applications in crop classification, agricultural EU subsidy control, crop growth modelling and precision farming. Moreover, drones will deliver increasing amounts of visual data to be interpreted by Machine Learning algorithms. As dida has ample experience in pattern recognition and image segmentation of satellite images, we believe that agriculture is a promising industry with various interesting problems to solve with Machine Learning and Computer Vision software.

In our afternoon session on may 6th we presented learnings from a recent legal-tech NLP project. In 2018 we were asked to develop a software, that is able judge the legal validity of certain paragraphs within rental contracts. After exposing lawyers to an early prototype, we realised that we needed to give detailed explanations about how the neural net derives its decision. As a consequence, we decided to split up the problem into different modules which helped much in creating a better transparency. A key module employs a recurrent neural net (RNN) in the form of a bidirectional LSTM. Now that the software is in production we wanted to share our learnings and also discuss the quality of the predictions. The Deep Learning World was held together in Munich with the Predictive Analytics World conference, attracting hundreds of data scientists from all over Europe. If you could not attend our session, however, are interested in a live demo of our machine learning application, please feel free to contact us.

As dida becomes more active in the public sector, we will join a workshop on public tendering for IT businesses. The event is organized by "Deutsches Vergabenetzwerk" and supported by the European Central Bank and the federal ministry for the environment and nature conservation. We are already looking forward to an interesting exchange with all the different stakeholders involved in the field of public procurement.

We are happy to be part of the state-funded cooperation network CopServ. Together with internationally renowned partners and research institutions such as Ruhr-Universität Bochum, Eftas or Exasol, CopServ aims at fostering innovation and growth through promoting the use of data from the European satellite program Copernicus. The main economic areas of application are in renewable energies, forestry, agriculture and mining. As dida has succesfully completed multiple projects in the field of pattern recognition from satellite sensor data in the last years, we are fully aware of the potential progess to be achieved by a stronger market uptake of Copernicus remote sensing data. Hence, we agree with CopServ's agenda and decided to join their efforts.

dida representatives joined the "Copernicus meets Galileo" conference 8th of May. Host was the Technische Hochschule Georg Agricola in Bochum, participants were Copernicus users from all over Germany. Presentations were given from institutions like the DLR, T-Systems and RWTH Aachen on the use of Copernicus and Galileo data and on applications built with Copernicus data. The event closed with a space show at the Planetarium Bochum to Pink Floyd music.

dida representatives wil join the 5th digital future science match organized by Hasso Plattner Institute, Zuse Institute Berlin and the newspaper Tagesspiegel. The conference is organized around the question 'What's next in Artificial Intelligence?' and features prominent speakers such as Anja Karliczek, Federal Minister of Education and Research, and Christoph Schütte, President of the Zuse Institute Berlin. Please contact us to schedule a meeting, if you want to find out more what dida can do for your organization or to discuss best practises on AI projects in general.

Thanks so much to almost 100 customers and friends who joined our machine learning meetup at Mindspace Berlin. We had decided to split the event into 2 sessions: 1) Teaching the computer to see: deep learning and image segmentation, 2) From 2d to 3d: post processing and projective geometry. Below you may find yourself starring in our latest image video which was produced from the material filmed at the meetup.

We very much enjoyed the opportunity to deliver a talk on the inner workings of convolutional neural nets (CNNs) at the Deep Learning World conference in Berlin. In 2015, CNNs first won against humans in the yearly ImageNet challenge on image classification. This marked a turning point in object recognition within visual data. Since then, further major advances has been made on the algorithmic and hardware side. It has never been easier to setup and run a CNN. dida made use of these technological advancements in various projects. The presentation provided the audience with insights into how object recognition projects are planned and what services are used to technically set up the project. Also major obstacles and solutions were discussed. Thanks to the audience for your active participation in the Q&A session!

We were invited to present a poster on 27th - 29th of November at the National Forum for Remote Sensing and Copernicus 2018 from a recent remote sensing project. A key part of the project involved an image segmentation, for which we had decided to use a convolutional neural net developed by the University of Freiburg (-> U-Net). The conference is hosted by the Federal Ministry of Transport and Digital Infrastructure in Berlin. With over 450 participants and many experienced speakers, the Copernicus Forum is one of the most distinguished events in the field of remote sensing in Europe. We want to thank DLR (German Aerospace Center) for the invitation and all the participants for the interesting exchange. We hope to see you again at the next Forum!

This year we decided to jointly participate in the PyData conference in Berlin with all team members. Thanks so much to the organizers, the speakers and sponsors such as Anaconda or Rasa! For us it was an interesting event with diverse talks ranging from parsing time expressions in documents over generalized additive models to modern approaches to bayesian learning. As always, one of the most enjoyable things at PyData is the relaxed and friendly atmosphere. If you are interested, most of the talks were videotaped - you can find them here.

At this year's Predictive Analytics World Industry 4.0 conference, we were asked to present a case study on how we used deep learning in a business application. We decided to showcase a recently developed software program that is able to predict from image data the number of solar panels which can be fit on a specific house. This application needs to be fed with address data as input and delivers within seconds accurate estimates of the roof area and the number of solar panels. We started out with a live- demonstration of the application and then moved on to explain the inner workings, especially the image segmentation part where we used deep learning. Thanks so much to our customer Enpal for taking the time to co-present the application.

After having worked on mostly machine vision applications lately, we have decided to share some of our results with the python community. Pydata Berlin will give us the opportunity to give a lightning talk on January 17th in front of an audience of up to 200 folks. We are already looking very much forward to this and hope to see many familiar faces! Also a big thanks to Celera One for providing food, drinks & the location.