Recorded talks

dida talks
Fabian Dechent
Building a real time license plate matching service

Fabian Dechent

May 23rd, 2025

This talk covers strategies applied in bringing a real-time license plate matching service into production. Based on the project requirements, we discuss the chosen ML model architecture and training procedure, implemented software architecture components, as well as tools and infrastructure for the deployment. Particular focus will be on the deployed microservices architecture and model serving using NVIDIA Triton.
dida talks
Axel Besinger

Augusto Stoffel (PhD)
From OCR to LLMs: The journey to reliable data extraction from complex retail documents

Axel Besinger and Augusto Stoffel (PhD)

May 23rd, 2025

AI-powered data extraction works - until it doesn’t. When handling structured tables in invoices, orders, or financial documents, we expect OCR, LLMs, and Vision AI to extract data reliably. However, complex documents - e.g. nested tables, irregular structures, and edge cases - pose real challenges for document data extraction AI models. With our solution smartextract, we tackled a real-world customer challenge: automating order entry from complex order documents and tables for a German shoe retailer: OCR and text-based LLMs struggled, Vision LLMs were inconsistent. Only extensive customization could solve the appearing problems - including segmentation, few-shot prompting, fine-tuning, and even the possibility of training a custom computer vision model. In this talk, we will show why standard AI models struggle with complex tables and demonstrate in which cases segmentation helps. Further, we will show benchmarks of commercial vs. open-source models and discuss the trade-offs between OCR, LLMs, and computer vision models.
dida talks
Ma Li (PhD)
Detecting subtle events in videos with YOLO-in-time

Ma Li (PhD)

May 23rd, 2025

In this talk, we present a general-purpose model architecture that is suitable for detecting and counting subtle events in videos. It combines a convolutional neural network (CNN), a recurrent neural network (RNN) and a YOLO-type head in the time domain. The model can be trained end-to-end with time-resolved labels only, without the need for customization or domain-specific knowledge. We discuss possible use cases and show experimental results both on synthetic datasets and on real world datasets.
dida talks
Lorenz Richter (PhD)
Bridging discrete and continuous state spaces: Exploring the Ehrenfest process in diffusion models

Lorenz Richter (PhD)

July 24th, 2024

At the International Conference on Machine Learning (ICML) 2024, Lorenz Richter (PhD) together with two colleagues Ludwig Winkler and Manfred Opper presented their latest paper. Abstract: Generative modeling via stochastic processes has led to remarkable empirical results as well as to recent advances in their theoretical understanding. In principle, both space and time of the processes can be discrete or continuous. In this work, we study time-continuous Markov jump processes on discrete state spaces and investigate their correspondence to state-continuous diffusion processes given by SDEs. In particular, we revisit the Ehrenfest processEhrenfest process, which converges to an Ornstein-Uhlenbeck process in the infinite state space limit. Likewise, we can show that the time-reversal of the Ehrenfest process converges to the time-reversed Ornstein-Uhlenbeck process. This observation bridges discrete and continuous state spaces and allows to carry over methods from one to the respective other setting, such as for instance loss functions that lead to improved convergence. Additionally, we suggest an algorithm for training the time-reversal of Markov jump processes which relies on conditional expectations and can thus be directly related to denoising score matching. We demonstrate our methods in multiple convincing numerical experiments.
dida talks
Lorenz Richter (PhD)
A dynamical systems perspective on measure transport and generative modeling

Lorenz Richter (PhD)

July 8th, 2024

In July 2024, Lorenz Richter (PhD) followed the invitation of the Fourth Symposium on Machine Learning and Dynamical Systems at the Fields Institute for Research in Mathematical Sciences (University of Toronto) and gave a talk about "A dynamical systems perspective on measure transport and generative modeling".
dida talks
Jan Macdonald (PhD)
Pretraining AI models for earth observation: transfer-learning and meta-learning

Jan Macdonald (PhD)

May 31st, 2024

Pretraining involves training an AI model on a large dataset to learn general features, which can then be finetuned on specific tasks with smaller datasets. This decreases the need for time intensive dataset acquisition and training efforts for each new use case, reducing the costs of application development. While pretrained models are widely used in computer vision and natural language processing, their adoption for satellite data and earth observation applications remains limited. Our investigation focuses on comparing the capabilities of transfer-learning and meta-learning approaches for the pretraining of AI models in earth observation tasks, particularly crop type classification, and their potential to generalize insights across different geographical regions.
dida talks
Fabian Dechent
Decision process automation with Large Language Models

Fabian Dechent

May 31st, 2024

Large Language Models impress with their adeptness in context-aware text generation, logic, as well as reasoning. Typically, downstream models fine tuned on chat data possess the remarkable ability to be directed towards solving tasks described in natural language without explicit further weight adaptation. In relevant applications, interesting use cases often relate multiple external data sources with each other and are characterized by a complex multistep decision process. In this talk, we discuss how predefining decision steps and integrating external data filtering can break down multifaceted problems into manageable, self-contained language processing tasks, which can readily be solved by LLMs.
dida talks
Jakob Scharlau
Applications of machine learning in the public sector

Jakob Scharlau

May 31st, 2024

This presentation dives into the transformative potential of machine learning (ML) technologies in the public sector, highlighting opportunities for efficiency, transparency, and improved service delivery. Using a case example from practice, it illustrates how ML applications already have found their way into public processes.
dida talks
Maximilian Trescher (PhD)
Anomaly detection in track scenes

Maximilian Trescher (PhD)

May 31st, 2024

Within the sector initiative “Digitale Schiene Deutschland”, our client Deutsche Bahn is developing an automated driving system for trains. As a part of the efforts towards such a system we developed, together with Deutsche Bahn, a machine learning solution to detect anomalous and hazardous objects on and around the tracks using onboard RGB cameras. It is intentionally required that this system does not simply detect objects within a given collection of classes (such as people, signals or vehicles), but rather has the ability to detect any object and rank them by how anomalous they are. This presentation explains the challenges encountered, presents several approaches explored, and provides an overview of the final solution: In order to detect objects of possibly unkown classes we developed a unique pipeline containing multiple machine learning components, including a monocular depth estimation model, a segmentation stage, image embedding models and an anomaly detection model. As dataset, Digitale Schiene Deutschland provides us with OSDAR23, an open dataset that contains 45 scenes. Each scene contains images taken by several RGB cameras and infrared cameras, together with radar and lidar data. This dataset contains annotations for twenty classes of objects, which we use both for finetuning our model and for evaluating the final results. Besides, we were also granted access to a larger amount of unannotated data, which were used for self-supervised learning.
dida talks
Axel Besinger

Augusto Stoffel (PhD)
Data extraction in the age of LLMs

Axel Besinger and Augusto Stoffel (PhD)

May 31st, 2024

In recent years, the advent of Large Language Models (LLMs) has changed the landscape of data extraction. These LLMs boast unparalleled text processing capabilities and come pre-trained on vast amounts of data, rendering them effective for information retrieval tasks. However, traditional methods such as graph neural networks and extractive models have historically been favored for their efficiency in resource utilization. Despite this, the question persists: how do LLMs compare with those models in practical data extraction applications? This presentation aims to delve into this inquiry, providing a comprehensive examination of LLMs' advantages and disadvantages compared to extractive models. Drawing from our project experiences and internal research, we aim to elucidate the practical implications of utilizing LLMs for data extraction, offering insights into their efficacy, resource requirements, and overall performance in real-world scenarios. Through this exploration, attendees will gain a deeper understanding of the role of LLMs in modern data extraction workflows and the considerations involved in their implementation. Link to the information extraction software: smartextract ( https://smartextract.ai )
dida talks
Lorenz Richter (PhD)
An Optimal Control Perspective on Diffusion-Based Generative Modeling

Lorenz Richter (PhD)

May 16th, 2024

In May 2024 Lorenz Richter (PhD) presented his work on diffusion-based generative models at the Machine Learning and Dynamical Systems Seminar of the Alan Turing Institute London. Abstract: This seminar will delve into the intersection of generative modeling via Stochastic Differential Equations (SDEs) and three pivotal areas of mathematics: stochastic optimal control, Partial Differential Equations (PDEs), and path space measures. This integration is foundational for both theoretical advancements and practical applications, such as transferring methods across fields or developing innovative algorithms for sampling from unnormalized densities. We introduce a variational framework that employs divergences between path space measures of time-reversed diffusion processes, drawing parallels to the classic Schrödinger bridge problem. This framework enables the use of novel divergence forms like the log-variance divergence, which avoids the pitfalls of the reverse Kullback-Leibler divergence and significantly enhances algorithmic performance across various methodologies.
dida talks
Jona Welsch
Information extraction with BERT from free-form text

Jona Welsch

April 28th, 2023

Jona Welsch's talk centers on using Deep Learning methods like BERT to extract information from unstructured text. A project with idealo serves as a case study, showcasing how rule-based algorithms and Deep Learning can be combined to turn product descriptions into structured data. The talk also touches on creating weakly labeled training data to ease the labeling process.
dida talks
Jakob Scharlau
Domain-specific semantic search: finding the fitting document

Jakob Scharlau

April 28th, 2023

The talk by Jakob Scharlau delves into how machine learning and pre-trained language models can be used to create search systems with a semantic understanding of language. It covers recent techniques and considerations for building such tools, and includes an example of a system that finds the right technical document based on a casual search query.
dida talks
Augusto Stoffel (PhD)
Information extraction: from graph neural networks to transformers

Augusto Stoffel (PhD)

April 28th, 2023

This talk aims to compare two prominent classes of models used in information extraction from semi-structured documents: Graph Neural Networks (GNNs) and specialized transformer-based architectures. While transformers are renowned for their text processing capabilities and come with pretrained weights, GNNs have the benefit of requiring much less computational power. The objective is to evaluate how these two types of models perform in practical scenarios, based on both project experience and internal research.
dida talks
Angela Maennel
Understanding Customer Needs with NLP

Angela Maennel

January 19th, 2023

This talk is about the benefits of Natural Language Processing (NLP) in providing flexibility over traditional restrictive online input methods. It hints at the freedom NLP offers, allowing for free-form text instead of limited set phrases or checkboxes.
dida talks
Marty Oelschläger (PhD)

Sara Zanzottera
Open NLP meetup: Ethics in Natural Language Processing

Marty Oelschläger (PhD) and Sara Zanzottera

December 1st, 2022

This talk covers two main topics. The first part delves into the ethical considerations in Natural Language Processing (NLP), discussing how language models are developed and used responsibly, addressing issues such as data privacy, algorithmic bias, and the societal impacts of automated language systems. The second segment provides a hands-on introduction to image retrieval, explaining the techniques and algorithms that enable the searching and finding of images based on content, metadata, or descriptive tags. This could include demonstrations of indexing images, feature extraction, and the use of search queries to navigate large image databases effectively.
dida talks
William Clemens (PhD)
LaserSKI: Object Detection for Defect Detection in Semiconductors

William Clemens (PhD)

May 11th, 2022

In this talk, William Clemens (PhD) presented our work about detecting defects in manufacturing processes of semiconductors. By utilizing convolutional neural networks, the system inspects images of semiconductors to identify and classify defects, enhancing the efficiency and reliability of quality control, with the goal of reducing monotonous manual inspections. LaserSKI was a joint project of three industrial manufacturers of laser diodes as well as the Ferdinand-Braun-Institut and Leibniz-Institut für Höchstfrequenztechnik (FBH). The talk was given at the Applied Machine Learning Days (AMLD) 2022.
dida talks
Moritz Besser

Jona Welsch
ML for Remote Sensing: Analyse satellite data automatically

Moritz Besser and Jona Welsch

December 6th, 2021

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.
dida talks
Petar Tomov

Philipp Jackmuth
Real Added Value from ML Projects - Our Success Factors

Petar Tomov and Philipp Jackmuth

October 26th, 2021

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.
dida talks
Augusto Stoffel (PhD)
Graph neural networks for information extraction with PyTorch

Augusto Stoffel (PhD)

July 30th, 2021

In Augusto Stoffel's talk, he introduces graph neural networks (GNNs) by comparing them to convolutional neural networks (CNNs). He describes how an image can be represented as a graph to naturally transition into the basics of GNN architecture. The talk then covers Python implementations, particularly in the PyTorch framework, and focuses on GNN applications in information extraction from tabular documents in the field of NLP.
dida talks
Konrad Mundinger
NLP trend: profitably integrating neural language models into business practice

Konrad Mundinger

July 1st, 2021

After an overview of how neural language models work and examples of their possible applications we will show things to be taken into account in order to use natural language models effectively in one's own business practice.
dida talks
Mattes Mollenhauer
Automated answering of questions with neural networks: BERT

Mattes Mollenhauer

May 26th, 2021

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.
dida talks
Fabian Gringel
Recurrent neural networks: How computers learn to read

Fabian Gringel

May 26th, 2021

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.
dida talks
Ewelina Fiebig

Fabian Gringel
Text recognition (OCR) - The first step on the way to a successful implementation of an NLP project

Ewelina Fiebig and Fabian Gringel

May 26th, 2021

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?
dida talks
Ewelina Fiebig

Fabian Gringel
Labeling Tools - The second step on the way to the successful implementation of an NLP project

Ewelina Fiebig and Fabian Gringel

May 26th, 2021

The success of an NLP project consists of a series of steps from data preparation to modeling and deployment. Since the input data are often scanned documents, the data preparation step initially involves the use of text recognition tools (OCR for short) and later on also the use of so-called labeling tools. In this webinar we will deal with the topic of selecting a suitable labeling tool.
dida talks
Konrad Schultka (PhD)

Jona Welsch
Semantic search and understanding of natural text with neural networks: BERT

Konrad Schultka (PhD) and Jona Welsch

May 26th, 2021

In this webinar you will get an introduction to the application of BERT for Semantic Search using a real case study: Every year millions of citizens interact with public authorities and are regularly overwhelmed by the technical language used there. We have successfully used BERT to deliver the right answer from government documents with the help of colloquial queries - without having to use technical terms in the queries.
dida talks
William Clemens (PhD)
Detecting Convective Clouds in Geostationary Satellite

William Clemens (PhD)

February 26th, 2020

Detecting convective clouds is crucial for weather forecasting and climate studies. In his work, William Clemens (PhD), a Machine Learning Scientist at dida, leverages Convolutional Neural Networks (CNNs) to analyze geostationary satellite data for this purpose. CNNs are particularly adept at image recognition tasks, making them suitable for identifying the complex patterns and structures characteristic of convective clouds. Clemens's approach likely involves training the CNNs on large datasets of satellite imagery labeled with the presence of convective clouds, enabling the model to learn the distinguishing features of these clouds.

Recorded talks

Building a real time license plate matching service

From OCR to LLMs: The journey to reliable data extraction from complex retail documents

Detecting subtle events in videos with YOLO-in-time

Bridging discrete and continuous state spaces: Exploring the Ehrenfest process in diffusion models

A dynamical systems perspective on measure transport and generative modeling

Pretraining AI models for earth observation: transfer-learning and meta-learning

Decision process automation with Large Language Models

Applications of machine learning in the public sector

Anomaly detection in track scenes

Data extraction in the age of LLMs

An Optimal Control Perspective on Diffusion-Based Generative Modeling

Information extraction with BERT from free-form text

Domain-specific semantic search: finding the fitting document

Information extraction: from graph neural networks to transformers

Understanding Customer Needs with NLP

Open NLP meetup: Ethics in Natural Language Processing

LaserSKI: Object Detection for Defect Detection in Semiconductors

ML for Remote Sensing: Analyse satellite data automatically

Real Added Value from ML Projects - Our Success Factors

Graph neural networks for information extraction with PyTorch

NLP trend: profitably integrating neural language models into business practice

Automated answering of questions with neural networks: BERT

Recurrent neural networks: How computers learn to read

Text recognition (OCR) - The first step on the way to a successful implementation of an NLP project

Labeling Tools - The second step on the way to the successful implementation of an NLP project

Semantic search and understanding of natural text with neural networks: BERT

Detecting Convective Clouds in Geostationary Satellite