DSAI Seminar

TBD

Bio: TBD. (Jingjing is a postdoc at DSAI.)

TBD

Bio: TBD. (Richard is a postdoc at DSAI.)

TBD

Bio: Dr. Martin Trapp is an Assistant Professor in Machine Learning at KTH Royal Institute of Technology, WASP fellow, and a member of the ELLIS society working on probabilistic machine learning. Previously, he was an Academy of Finland-funded independent postdoctoral researcher at Aalto University. His research interests are in scalable and principled methods in probabilistic machine learning with a focus on tractable models and Bayesian statistics.

TBD

Bio: TBD. (Nikolai is a postdoc at DSAI.)

TBD

Bio: TBD. (Frederik is an incoming postdoc at DSAI.)

Causal discovery is the process of learning structural information that allows for the inference of causal effects using observational data. Standard methods postulate that causal structure can be described by a directed acyclic graph (DAG), which can be partially reconstructed by using the faithfulness principle. According to this principle, conditional independencies in the observational distribution correspond exactly to those encoded in the unknown DAG.
However, causal discovery algorithms are challenged by the fact that, without even stronger assumptions, approximate failures of faithfulness are empirically undistinguishable from exact ones. This leaves an inferential gap, as exact failures of faithfulness can lead to major implications to causal effect identification. Moreover, in many applications, it is not realistic to expect that exact conditional independencies hold among most recorded variables: they are often just proxies to unobservable manipulable factors, so that there is no structural justification to expect such independencies to emerge.
The above suggests a different perspective, reliant neither on exact independence nor on the distinction between approximate and exact faithfulness. By considering the specialized problem of discovering whether unmeasured confounding can be ruled out for a particular cause-effect pair, we introduce a principle we call naturalness, in which the strength of observational associations bounds the strength of causal dependencies. The bounding function is unknown, with degrees of freedom not present in faithfulness. We model uncertainty about this function using a Bayesian framing.
To go beyond uncertainty which is purely theory-driven, we present an algorithm that grounds it using experimental data from past studies, resulting on adaptive causal discovery principle. We demonstrate its usefulness in providing partial identification of causal effects compared to other methods that relax the assumption of no unmeasured confounding, and discuss how the promising results can motivate new principles for general causal discovery.

Bio: Ricardo is a Professor of Statistical Machine Learning and Data Science at University College London, department of Statistical Science, and a member of the UCL AI Centre. He holds an Open Fellowship from the Engineering and Physical Sciences Research Council (EPSRC, UK), as well as being a co-lead of the multi-institutional Causality in Healthcare AI Hub (CHAI) funded by EPSRC, and a co-investigator in the UCL Science of Fundamental AI Research (SOFAIR) Lab. Ricardo’s research has touched many aspects of uncertainty in causal modelling, including issues of measurement error, uncertainty in identifiability, and Bayesian inference for integrating multiple sources of evidence. He was one of the Programme co-Chairs of the Uncertainty in Artificial Intelligence Conference (UAI) in 2018, and Conference co-Chair of UAI 2019, besides being a co-organizer of multiple workshops along the years in conferences such as NeurIPS, UAI and ICML.

Scientific machine learning has emerged as a promising research paradigm at the intersection of machine learning (ML) and scientific computing. Scientific knowledge can be incorporated into ML models, for example, via architectural changes or physics-based loss terms, improving their accuracy and their ability to generalize. Conversely, ML methods can enable new applications in the natural sciences that would otherwise be too computationally expensive or intractable. In this talk, I will give an overview of several research projects in this area from my PhD research and my ongoing work in Simon Olsson's group at Chalmers. Physics-informed neural networks (PINNs) constitute a highly flexible framework for solving differential equations, with particular promise for inverse problems in non-standard settings. In my research, I focused on training PINNs with noisy data. Specifically, we will see how the noise shape can be identified together with the solution to the differential equation, and how uncertainty intervals can be estimated. The design of particle detectors typically relies on scientific intuition and expensive simulations. There is the possibility to make the exploration of different detector configurations more efficient by using optimization methods from ML. To this end, we developed a generative surrogate model for signal generation in the IceCube-Gen2 detector, to serve as part of a fully differentiable optimization pipeline. While it is possible to perform molecular simulations to high accuracy, they can be extremely computationally expensive. Implicit transfer operators (ITOs) seek to remedy this issue by training an ML surrogate model capable of taking large time steps, and with the potential to generalize to new molecules. In our work, we develop a framework for the systematic evaluation of ITOs, in particular, how accurately different physical properties of the underlying dynamical processes are reproduced. In a separate branch of research, we study the effects that the masking of the training data has on ITOs. We show how this could, in principle, imbue the ITO with additional capabilities, and we highlight connections to coarse-graining.

Bio: Philipp Pilar joined the Artificial Intelligence in the Natural Sciences (AIMLeNS) group at Chalmers University of Technology as a postdoctoral researcher in 2026. He holds a PhD in physics-informed machine learning from Uppsala University and an MSc in Technical Physics from TU Wien. His research interests lie in scientific machine learning, with a particular focus on developing machine learning methods for scientific applications. He is currently working on generative machine learning models for molecular dynamics.

Real-world datasets are inherently heterogeneous, yet how per-class structural differences and sampling imbalance shape the training dynamics of diffusion models—and potentially exacerbate disparities—remains poorly understood. While models typically transition from an initial phase of generalization to memorizing the training set, existing theory assumes homogeneous data, leaving open how class imbalance and heterogeneity reshape these dynamics. In this work, we develop a high-dimensional analytical framework to study class-dependent learning in score-based diffusion models. Analyzing a random-features model trained on Gaussian mixtures, we derive the feature-covariance spectrum to characterize per-class generalization and memorization times. We reveal the explicit hierarchy governing these dynamics: class variance is the primary determinant of learning order—consistently favoring higher-variance classes—while centroid geometry plays a secondary role. Sampling imbalance acts as a modulator that can reverse this ordering and, under strong imbalance, forces minority classes to acquire distinct, delayed speciation times during backward diffusion. Together, these results suggest that diffusion models can memorize some classes while others remain insufficiently learned. We validate our theoretical predictions empirically using U-Net models trained on Fashion MNIST.

Bio: Flavio worked from 2020 to 2023 as a PhD student with Prof. Federico Ricci-Tersenghi (Sapienza University of Rome) and Prof. Silvio Franz (Université Paris-Saclay) in an international cotutelle program, studying the low-temperature physics of vector spin glasses. From July 2023 to March 2025, he worked as a postdoc with Prof. Federico Ricci-Tersenghi, studying generalizations of vector spin glasses to neural network models. Flavio is now a postdoc at DSAI working with Stefano Sarao Mannelli. His current research focuses on understanding the impact of bias in machine learning and animal behaviour, building on simplified and solvable neural network models.

In this talk several types of biosignals will be covered explaining the challenges of collecting, processing, extracting information and making sense of the physiological dynamics. The application of machine learning techniques to biosignals will also be covered to give a perspective on how to create automated decision mechanisms based on biosignals capture. During the several topics of the presentation, research examples conducted at Nova University of Lisbon in collaboration with PLUX and Fraunhofer Portugal will be given.

Bio: Hugo Gamboa is a Full Professor at the Physics Department of the Nova School of Science and Technology of the Universidade Nova de Lisboa and member of LIBPHYS. PhD in Electrical and Computer Engineering from Instituto Superior Técnico, University of Lisbon. As a Senior Scientist at Fraunhofer Portugal, he coordinates the Lisbon Office research group focusing on Intelligent Systems. He is a founder and President of PLUX, a technology-based innovative startup in the field wireless medical sensors, focused on microelectronics, biosignal processing and software development. He is the Director of LIBPHys and leads a researcher team with expertise on medical instrumentation, biosignal processing and machine learning applied to biosignals. Published more than 80 Journal Papers; 15 Book Chapters; 10 books (selected best papers); 130 Conference Papers.

Antisense oligonucleotides (ASOs) are short synthetic nucleic acids designed to bind specific RNA targets and modulate their expression. In the cellular environment, however, ASOs can also bind to other mRNAs, potentially sequestering them away from their intended targets. Mechanistic modeling of these interactions leads to large reaction networks spanning multiple timescales and involving species with widely varying molecular abundances. These properties make exact stochastic simulation using the Gillespie Stochastic Simulation Algorithm (GSSA) computationally demanding. Approximate approaches, such as τ-leaping and hybrid methods, can reduce computational cost, but provided insufficient speedup for our models. To address this bottleneck, I developed a GPU-accelerated stochastic simulator implementing a parallelized version of the GSSA in CUDA. In this talk, I will present the design and implementation of the solver, discuss the challenges of mapping the GSSA to GPU architectures, and evaluate its performance on biochemical reaction networks of varying size and complexity. (Based on joint work with Alexander Schliep.)

Bio: Aleksandra (Sasha) Khatova is a PhD student in Medical Bioinformatics at Brandenburg University of Technology Cottbus-Senftenberg. Her research focuses on kinetic modeling of oligonucleotide action, with a particular interest in how intracellular RNA dynamics influence the efficacy of oligonucleotide therapeutics. Her work uses stochastic simulations to study these systems and their variability at the single-cell level. To enable efficient simulation of large and computationally demanding models, she also develops GPU-accelerated methods for stochastic simulation of biochemical reaction networks.

This is a brief description of a project that has already autoformalized a large portion of the general topology from the Munkres textbook (which has in total 241 pages in 7 chapters and 39 sections). The project has been running since November 21, 2025 and has as of January 4, 2026, produced 160k lines of formalized topology. Most of it (about 130k lines) have been done in two weeks,from December 22 to January 4, for an LLM subscription cost of about $100. This includes a 3k-line proof of Urysohn's lemma, a 2k-line proof of Urysohn's Metrization theorem, over 10k-line proof of the Tietze extension theorem, and many more (in total over 1.5k lemmas/theorems). The approach is quite simple and cheap: build a long-running feedback loop between an LLM and a reasonably fast proof checker equipped with a core foundational library. The LLM is now instantiated as ChatGPT (mostly 5.2) or Claude Sonnet (4.5) run through the respective Codex or Claude Code command line interfaces. The proof checker is Chad Brown's higher-order set theory system Megalodon, and the core library is Brown's formalization of basic set theory and surreal numbers (including reals, etc). The rest is some prompt engineering and technical choices which we describe here. Based on the fast progress, low cost, virtually unknown ITP/library, and the simple setup available to everyone, we believe that (auto)formalization may become quite easy and ubiquitous in 2026, regardless of which proof assistant is used.

Contemporary dance offers a demanding yet revealing testbed for the design of interactive generative systems. It is a creative practice that treats the human body as a primary medium of expression, values experimentation over codified style, and integrates influences beyond Western dance traditions. From a human–computer interaction and creative computing perspective, it poses several challenges: embodied and tacit knowledge resists formalisation, somatic awareness is difficult to externalise and quantify through sensing technologies, and highly individual creative approaches defy generalisation and standardisation. This talk presents an artist-guided approach to developing interactive generative systems in which expert dancers actively contribute to system design, data representation, and evaluation. The approach reframes development as a co-creative process—one where dancers’ embodied knowledge informs both system design and interaction strategies, and where generative algorithms act as adaptive partners rather than fixed tools. Two case studies illustrate this methodology. Pemida (Personalised Musical Instrument for Dancers) employs machine learning to model individual improvisation strategies to music, enabling dancers to perform with highly personalised instruments responsive to full-body motion. Expressive Aliens applies simulation-based generative methods to map dancers’ distinct movement vocabularies to behavioural patterns for robotic lights, integrating dynamic light as a choreographic medium. Together, these projects demonstrate how artist-guided co-design can ground the development of generative systems in real-world creative practice, enhance their artistic validity and usability in performance contexts, and promote long-term relevance within dancers’ evolving repertoires. They highlight the potential of embodied co-creation as a productive framework for research in human–AI interaction and creative computing.

When lying in the park on the grass, thoughts jump from one to another—seemingly unpredictable yet structured. How can high-dimensional neural systems generate flexible activity, as observed in spontaneous cognition? In this talk, I will present my PhD work, which focuses on this question using coarse-grained models of cortical dynamics inspired by statistical physics. In the Potts associative network, multi-state units represent cortical patches storing categorical features, and long-range interactions encode distributed memories. While the standard model performs pattern completion through attractor dynamics, introducing slow adaptation drives the system into a qualitatively different regime: instead of converging to a fixed point, it exhibits ongoing transitions between memories. This latching dynamics provides a minimal example of self-organized, structured activity in associative networks. I will show that latching dynamics is intrinsically non-equilibrium and cannot, in general, be reduced to effective equilibrium descriptions: embedding the system into energy-based frameworks either fails to destabilize attractors or leads to dynamics that retain memory of initial conditions. I will then present results we obtained using dynamical mean-field theory, which allowed us to derive an exact effective single-unit description of the dynamics and characterize when transitions are absent, random, or structured. I will emphasize a key result: multi-state representations are essential, as binary coding schemes lack the internal degrees of freedom required for these non-trivially structured transitions. Finally, I will show how coupling subnetworks with different adaptation timescales leads to emergent hierarchical control of dynamics, suggesting a link between cortical organization and the flexible yet structured flow of spontaneous cognition. More broadly, the work presented in this talk will point to general principles by which high-dimensional systems can generate structured yet intrinsically unpredictable dynamics.

Several successful strategies in automated reasoning rely on human-supplied guidance about which term or clause shapes are interesting. In this paper we aim to discover interesting term shapes automatically. Specifically, we discover abstractions: term patterns that occur over and over again in relevant proofs. We present our tool Twitch which discovers abstractions with the help of Stitch, a tool originally developed for discovering reusable library functions in program synthesis tasks. Twitch can produce abstractions in two ways: (1) from a partial, failed proof of a conjecture; (2) from successful proofs of other theorems in the same domain. We have also extended Twee, an equational theorem prover, to use these abstractions. We evaluate Twitch on a set of unit equality (UEQ) problems from TPTP, and show that it can help prove hard problems as well as yield significant speed-ups on many other problems.

How can you create a system that can design novel, good games? And why would you want to do that? To answer the latter question first, we might want to understand how humans create or model their decision-making processes, but we may also want to create new testbeds for AI development. Or ideation tools for designers. To answer the first question, I will describe a series of systems developed my teams at IT University of Copenhagen and New York University. These systems generally build on evolutionary algorithms, and use simulated game-playing as part of an evaluation function. Accurately algorithmically estimating game quality is hard, but inspiration can be taken from game design theory, developmental psychology, and other cognitive sciences. Estimating how well a human would play or learn to play a game, a necessary precondition in many theories of game quality, is itself non-trivial, but reinforcement learning, imitation learning, and good old planning provide useful tools.

Creatively translating complex gameplay ideas into executable artifacts (e.g., games as Unity projects and code) remains a central challenge in computational game creativity. Gameplay design patterns provide a structured representation for describing gameplay phenomena, enabling designers to decompose high-level ideas into entities, constraints, and rule-driven dynamics. Among them, goal patterns formalize common player-objective relationships. Goal Playable Concepts (GPCs) operationalize these abstractions as playable Unity engine implementations, supporting experiential exploration and compositional gameplay design. We frame scalable playable pattern realization as a problem of constrained executable creative synthesis: generated artifacts must satisfy Unity's syntactic and architectural requirements while preserving the semantic gameplay meanings encoded in goal patterns. This dual constraint limits scalability. Therefore, we investigate whether contemporary large language models (LLMs) can perform such synthesis under engine-level structural constraints and generate Unity code (as games) structured and conditioned by goal playable patterns. Using 26 goal pattern instantiations, we compare a direct generation baseline (natural language -> C# -> Unity) with pipelines conditioned on a human-authored Unity-specific intermediate representation (IR), across three IR configurations and two open-source models (DeepSeek-Coder-V2-Lite-Instruct and Qwen2.5-Coder-7B-Instruct). Compilation success is evaluated via automated Unity replay. We propose grounding and hygiene failure modes, identifying structural and project-level grounding as primary bottlenecks.
https://arxiv.org/abs/2603.07101

Scientific discovery increasingly relies on mechanistic simulators to translate real observations into testable explanations of natural phenomena. Two core challenges arise: unknown simulator parameters must be inferred efficiently, and simulators are often imperfect representations of reality. This talk presents two approaches that address both challenges. First, we introduce a method leveraging tabular foundation models to perform simulation-efficient posterior parameter inference. Our method surpasses state-of-the-art methods and eliminates the need for training or tuning. Second, to close the gap between simulators and reality, we present a framework for LLM-based model discovery leveraging sequential Monte Carlo methods for probabilistic model inference. Together, we aim to enable principled and efficient simulator inference that can stimulate scientific discovery as well as the development of next-generation discovery algorithms.

This talk explores the emerging field of generative recommender systems and their foundation in large language models. We'll examine what foundation models are, when they offer advantages for recommendation tasks, and when traditional approaches may be more suitable. A key focus will be on Semantic IDs—a novel approach to representing items in ways that LLMs can understand and generate. I'll demonstrate how to leverage Semantic IDs to build LLM-based recommender systems, covering practical aspects including: Implementing models for Semantic ID generation, Fine-tuning large language models for recommendation tasks, Optimization strategies for efficient training, Evaluation methodologies for generative recommendation systems. The session will provide both conceptual foundations and practical insights for researchers and practitioners interested in applying generative AI to recommendation problems.