News | Neuro-Mechanistic Modeling

Our Paper On Effective Human Oversight Of AI at ACM FAccT 2024

Thu, 06 Jun 2024 00:00:00 +0000

The ACM FAccT in Rio de Janeiro, is a conference that addresses some of the most important questions of our time.

Kevin Baum, accompanied by Sarah Sterz and Philip Meinel, were privileged to have had the opportunity to present their interdisciplinary paper “On the Quest for Effectiveness in Human Oversight” at the ACM Conference on Fairness, Accountability, and Transparency ACM FAccT.

The paper highlights challenges and solutions for effective human oversight of AI systems and their legal implications for the EU AI Act.

This is work by Sarah Sterz, Sebastian Biewer, Holger Hermanns, Anne Lauber-Rönsberg, Philip Meinel and Markus Langer.

The paper involved researchers from accross Germany, from Universität des Saarlandes, Technische Universität Dresden, Universität Freiburg, and the Deutsches Forschungszentrum für Künstliche Intelligenz (DFKI)/the Center for European Research in Trusted Artificial Intelligence (CERTAIN).

Click here for the final Paper

Tutorial accepted for IJCAI 2024 on Machine Ethics

Sat, 01 Jun 2024 00:00:00 +0000

Machine ethics is a multi-disciplinary effort focused on enabling ethical behavior from machines towards people and (in a sense) towards other machines.

This tutorial aims to present the state of the art in machine ethics, helping researchers enter this field or reflect critically on the ethical dimensions of their own work.

More info can be found here.

Successful Workshop on Safe and Causally Informed Deep Reinforcement Learning

Sun, 26 May 2024 00:00:00 +0000

On May 23 and 24, Verena Wolf’s research department for Neuro-Mechanistic Modeling (NMM) at the German Research Center for Artificial Intelligence (DFKI) Saarbrücken organized a scientific workshop on Safe and Causally Informed Deep Reinforcement Learning.

With guests from Kaiserslautern, Berlin, Trieste, and Trento, among others, the participants learned about recent advances in the field of Deep Reinforcement Learning with a particular focus on approaches to Safe RL and the incorporation of causality into the RL pipeline (Causal RL). Building on this, different ideas for new projects could be developed and promising lines of future work were discussed, where also interesting synergies between Safe and Causal RL were found.

Keynote Presentation by Prof. Dr. Verena Wolf at the ZF Data & AI Conference 2024

Tue, 16 Apr 2024 00:00:00 +0000

Prof. Dr. Verena Wolf was a keynote speaker at ZF’s internal annual 3rd Data & AI conference, which took place from 8th to 10th April 2024.

She presented her ideas on the topic of Unlocking the Potential of Hybrid Approaches by Integrating Machine Learning and Mechanistic Models.

You can find the topics presented in detail here.

New Journal Article: Algorithmic Fairness, XAI, Effective Human Oversight — and How These Topics Can Interact

Tue, 09 Apr 2024 00:00:00 +0000

How can we identify unfair assessments of AI models on an individual level and communicate them to a human-in-the-loop? And what does this have to do with the diesel emissions scandal aka. software doping?

Building on the brilliant preliminary work of our first author Sebastian Biewer on the topic of software doping and on the notation of individual fairness from Fairness through Awareness (by Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Richard Zemel), the team formally specified a runtime monitor resulting in a global, model-agnostic XAI method and investigated how such a system can help to make human oversight (also, but not restricted to, the spirit of the EU AI Act) effective and meaningful.

The journal article Software Doping Analysis for Human Oversight has been published in Formal Methods in System Design.

Paper on Effective Human Oversight accepted for ACM FAccT in Rio de Janeiro

Tue, 09 Apr 2024 00:00:00 +0000

Congratulations to Kevin Baum and his interdisciplinary team led by Sarah Sterz!

The paper identifies what distinguishes nominal from meaningful or effective human oversight — primarily understood as a means of risk mitigation. This allows us to systematically classify facilitators and inhibitors of effective human supervision. We pay particular attention to the application of our findings to Article 14 of the AI Act, which focuses on risk minimization.

The preprint is available here.

NextAID takes off!

Mon, 10 Jan 2022 00:00:00 +0000

The NextAID (Neuro-explicit AI for Drug Discovery) project officially starts at Saarland University.

The process of discovering a drug follows a time-consuming and expensive pipeline that explores the chemical space of potential drugs today mainly based on wet-lab experiments and database searches. High expectations are placed on deep learning methods to simplify this process.

We believe that such neuro-explicit approaches are a key concept for substantial advances in drug discovery. Our research focuses on hybrid approaches where domain knowledge is integrated into neural learning models in various forms. This can significantly improve the generalization capabilities of neural models, allow extrapolation beyond training data, and thus need less data to learn from.

Joschka Groß joins the chair as a PhD student

Thu, 01 Apr 2021 00:00:00 +0000

He will be working on reinforcement learning. Welcome Joschka!

Seminar on Monte Carlo Methods

Thu, 22 Oct 2020 00:00:00 +0000

Next semester we will offer a seminar on Monte Carlo Methods.

Monte Carlo Methods

The core idea of Monte Carlo (MC) methods is to perform computer simulations of a real-world system based on pseudo-random numbers. MC is applied in a large variety of research areas: physical sciences, computer sciences, engineering, statistics, finance, etc. Moreover, ideas that have originally been developed in the context of MC are meanwhile also in use for sampling problems in the area of machine learning.

Although being a simple and very direct approach to analyze a system, MC methods come with a lot of challenges, in particular when sampling rare events or when systems have multiple time scales. In this seminar, we take a computer science perspective on Monte Carlo and cover different MC algorithms to tackle these challenges.

More information can be found on the course page.

Seminars on Reinforcement Learning and Complex Networks

Sun, 29 Sep 2019 00:00:00 +0000

Next semester we will offer two seminars!

Exploring Complex Networks

Many real-world phenomena like rumor spreading, contagions in financial markets, epidemic outbreaks, or cognitive processes can be expressed in terms of complex networks. The seminar explores dynamical processes which are linked to networked structures in intriguing ways. In particular, we aim at understanding the complex interplay between the topology and the emerging dynamics of a network. We analyze brain networks or online social networks and explore various techniques for their analysis and classification ranging from diffusion models and stochastic simulations to deep learning approaches.

More information can be found on the course page.

Reinforcement Learning

Reinforcement Learning is — besides Supervised and Unsupervised Learning — one of the outstanding variants of Machine Learning. It has evolved to an effective tool, being able to create AI systems that are able to beat humans in almost every game, e.g. the famous strategy board game Go.

This seminar will try to do both: teach the basics of Reinforcement Learning in theory and also provide the ability to apply it. We will therefore first have a look at the actual state-of-the-art algorithms of Reinforcement Learning and afterwards apply these to the well-known game Connect 4.

More information can be found on the course page.

Timo Gros joins the chair as a PhD student

Sat, 01 Jun 2019 00:00:00 +0000

He will be working on reinforcement learning. Welcome Timo!

New Preprint on Control Variates in Stochastic Chemical Reaction Networks

Thu, 02 May 2019 00:00:00 +0000

In this work we show how the estimation of quantities can be made much more efficient by using information on the moments of the process.

This is done via the technique of control variates, which is essentially a random variable that is correlated with the variable of interest. The control variate technique allows to leverage this correlation by adjusting the final estimate according to the sampled value of the control variate.

In a chemical reaction network the moment $E(X_t^m)$ over time is given by

$$ \frac{d}{dt}E(X_t^m)=\sum_j \left( \left( {(X_t - v_j)}^m - X_t^m \right) \alpha_j(X_t) \right) $$

From this one can, by integration, derive

$$ 0 = E({f(X_{T})}) - e^{\lambda T}E({f(X_{t_0})}) + \sum_{k}c_k\int_{t_0}^{T}e^{\lambda(T-t)}E({X_t^{m_k}}),dt,. $$

The random variable implied by the expectation above is correlated with the expected value of the process at time $T$. Using it as a control variate we often can reduce the variance of the final estimator considerably:

The pre-print is available on arXiv.

Charalampos Kyriakopoulos successfully defends his PhD Thesis

Tue, 30 Apr 2019 00:00:00 +0000

Congratulations to Charis!

Today Charis successfully defended his PhD thesis “Stochastic modeling of DNA demethylation dynamics in ESCs”.

He will soon move on to Sevilla to continue his work as a data scientist.

Lecture: Probabilistic Models and Data Analysis

Fri, 15 Mar 2019 00:00:00 +0000

Next semester we will again offer the advanced lecture Probabilistic Models and Data Analysis.

The aim of this course is to give students of bioinformatics and computer science a solid background in probability and statistics as well as detailed knowledge about computer simulation of stochastic models, parameter estimation and hypothesis testing. Students will also learn how to use software tools such as R and Matlab to analyze data and probabilistic models.

This course is a flipped classroom course, which means that the two lecture/tutorial slots will be used for exercises, examples, questions and discussion of the material — participants prepare for these slots by either watching videos or reading the lecture notes.

More information can be found on the lecture page.