Forschungsseminar

A Neural Framework for Motivation-Driven Spatial Behavior

Seoyeon Park

Tue, 22. 9. 2026, A10.273 and https://webroom.hrz.tu-chemnitz.de/gl/mic-cv7-ptw

Autonomous, goal-directed behavior in animals emerges from the interaction between motivational regulation and spatial navigation - a coupling that remains largely absent from artificial agents, which typically pursue externally defined objectives rather than internally generated goals. This thesis develops a biologically inspired framework that integrates an allostatic model of motivational drive competition (Rosado et al., 2022, 2025) with a hippocampal sequence model of goal-directed spatial navigation (Gönner et al., 2017), aiming to link the "why" of goal selection to the "how" of spatial execution within a unified spiking neural architecture. To this end, the hippocampal sequence model was migrated to a modern simulation environment and its attractor dynamics were quantitatively evaluated and calibrated as a validated foundation; the allostatic model's drive dynamics were then translated into a spiking representation and coupled to the hippocampal network through a winner-take-all motivational layer. Component-level diagnostics confirm that the coupled system correctly performs drive competition, context activation, and motivated exploration; however, a drive-specific weight-convergence issue has so far prevented the agent from consistently reaching a learned state for all motivational contexts, leaving Task Success Rate and Navigation Efficiency outcomes for the fully integrated agent as an open result.

A Neural Network Model of Inattentional Blindness Based on SimCLR Contrastive Learning Image Encoding

Florian Karsch

Thu, 17. 9. 2026, A10.368 and https://webroom.hrz.tu-chemnitz.de/gl/mic-cv7-ptw

Phenomena of visual attention like inattentional blindness demonstrate that subjective perception can be misleading. Computational models enable the simulation of visual attention mechanisms, allowing us to visualize and further understand visual attention. This thesis simulates the attentional spotlight during the experiment of Simons & Chabris (1999) with the computational attention model of Beuth (2019). To provide the required feature representations for the model, a self-supervised SimCLR framework is trained. Several training approaches are tested and attention simulation during the experiment is evaluated. Self-supervised approaches fail to generate successful attention simulations, likely due to data scarcity as well as the framework's sensitivity to parameter tuning. Instead, an additionally trained supervised training approach achieves human-like simulation quality. However, this success proved to be highly unstable as only a single encoder training yielded the necessary feature representations. Still, a biologically plausible simulation of the attentional spotlight during the inattentional blindness experiment is achieved. Although only one completely accurate simulation is produced, multiple partly good simulations are generated. The simulated behaviour closely mirrors the experimental results of Simons & Chabris (1999).

Towards a Therapy-Informed Conversational Agent Using Large Language Models

Maren Gröne

Wed, 19. 8. 2026, A10.368 and https://webroom.hrz.tu-chemnitz.de/gl/mic-cv7-ptw

This paper presents a therapy-informed conversational agent based on Dirk de Sousa's therapeutic framework and implemented using a modular large language model (LLM) architecture. Therapy sessions are structured into phases, goals, and subgoals managed by a Stage Manager, while dedicated agents evaluate goal completion, summarize dialogue for short-term memory, extract patient information into session notes, and provide additional psychotherapeutic context through retrieval-augmented generation (RAG).The system is evaluated using automatically generated conversations between a Coach Agent and an LLM-based Patient Agent representing synthetic patients. An ablation study shows that the Stage Manager, Evaluator Agent, and Summarizer Agent are important for maintaining structured progression toward therapeutic objectives, whereas the benefits of the RAG and Note Taker Agents remain inconclusive. The automated evaluation framework supports iterative system development and comparison of architectural changes. Future work should refine module-specific evaluation, improve module integration, explore synthetic dialogues for model training, and validate the system with human participants and qualified psychotherapists.

Approaches to Reducing Draft-Verifier Misalignment in Speculative Decoding

Silvio Meyer

Thu, 16. 7. 2026, A10.368 and https://webroom.hrz.tu-chemnitz.de/gl/mic-cv7-ptw

Large language models are inherently limited by their autoregressive nature, requiring tokens to be generated sequentially. Speculative decoding addresses this limitation by using a lightweight draft model to generate candidate tokens that can be verified in parallel by the target model, thereby reducing inference latency. The effectiveness of speculative decoding largely depends on how accurately the draft model predicts the target model's output. This seminar examines three methods for improving draft-verifier alignment: DistillSpec which employs knowledge distillation to align the draft model's output distribution with the target model's, EAGLE-3 which utilizes the hidden states from the target model to generate more accurate drafts and SD² which steers the draft model using a vector derived from the target model's hidden states. The three approaches are compared with respect to their underlying mechanisms, performance and trade-offs.

Reward-Free Learning with Auto-Generated Goals from Demonstrations

Lida Taleb Safa

Thu, 9. 7. 2026, A10.368 and https://webroom.hrz.tu-chemnitz.de/gl/mic-cv7-ptw

While reinforcement learning has achieved promising results in robotic task learning, it continues to face challenges related to reward function design and sample inefficiency, particularly in multi-task settings. Goal-conditioned imitation learning (GCIL) and goal-conditioned reinforcement learning (GCRL) address these challenges by conditioning policies on the desired goals. However, these methods still rely either on large expert datasets, whose collection is costly, or on sparse rewards, which introduce challenges in accurate value learning, particularly in long-horizon tasks. To address these limitations, this work proposes a reward-free learning framework that automatically generates goals from trajectories using hindsight relabeling and learns a dense reward function from the relabeled dataset. An initial policy is trained through goal-conditioned behavioral cloning (GCBC) on the relabeled dataset and subsequently refined through online interaction using Implicit Q-Learning (IQL) guided by the learned reward function. As a result, the framework can learn from unlabeled reward-free datasets with varying trajectory quality, including both optimal and suboptimal demonstrations. Experimental evaluation on AntMaze and Franka Kitchen demonstrates that the performance of STG-based online fine-tuning is highly sensitive to hyperparameter selection and environment characteristics. Ablation studies further suggest that appropriate choices of 𝜏 and 𝜌 can help mitigate the impact of reward estimation errors and improve training stability.

Towards Explainable System State Characterization in Cobot Gripper Systems: A Multimodal Learning Approach with Transformer Based Model

Hitesh Babariya

Fri, 26. 6. 2026, https://webroom.hrz.tu-chemnitz.de/gl/mic-cv7-ptw

Collaborative robots use vacuum grippers for pick-and-place tasks, but diagnosing failures is difficult because different faults often look similar. This thesis proposes an explainable multimodal learning framework for automatic fault classification in cobot grippers. It uses self-supervised learning to extract features from sensor data, a transformer model to combine multiple signals and detect fault types, and a retrieval-based system to generate clear, human-readable explanations from technical documents. Tested on real experimental data, the approach accurately distinguishes faults, improves interpretability, and provides practical maintenance insights for industrial use.

Explicit and Implicit Measures of The Rubber Hand Illusion and An Experiment in Unsupervised Learning

Carolin Loschinski

Thu, 25. 6. 2026, A10.368 and https://webroom.hrz.tu-chemnitz.de/gl/mic-cv7-ptw

The Rubber Hand Illusion is a phenomenon of bodily perception where humans believe that a fake hand is associated with their body. Explicit measurements typically capture various aspects of this association such as ownership, agency or location perception. Robot studies and simulations however rely on implicit measurements of which proprioceptive drift is the most prominent representative. In this presentation, I summarize current research on the relation between explicit and implicit metrics. Furthermore, I present a neuron model integrating population-coded visual, proprioceptive and tactile inputs. The model learns feedforward connections from the input population to the integration population using Annealed Linear Learning. I evaluate the results that can be achieved with Annealed Linear Learning and compare them to Hebbian learning as another representative of unsupervised learning.

A GPT-Style Transformer for Next-Step Prediction in CAD Sequences

Sayeda Hadisa Habibi

Wed, 24. 6. 2026, A10.367 and https://webroom.hrz.tu-chemnitz.de/gl/mic-cv7-ptw

This work investigates the application of a lightweight GPT-style architecture, based on NanoChat, for the generation of parametric CAD models. The objective is to evaluate whether transformer-based language models can effectively learn structured geometric representations and generate valid CAD sequences from CAD data. In addition, the presentation provides an overview of the model implementation process, including data preprocessing, tokenization, training, and evaluation. To study the impact of data representation on model performance, we implement and compare two tokenization strategies. The first follows the original DeepCAD tokenization method, which represents CAD commands and their associated parameters as structured sequences of discrete tokens using a fixed-length representation. The second strategy, referred to as BPE-DC, is inspired by both DeepCAD and Byte Pair Encoding (BPE). Instead of encoding command parameters into a fixed-length vector, parameters are preserved as separate elements, after which BPE is applied to learn the vocabulary over CAD commands and parameter sequences. In our experiments, BPE-DC is used as the baseline representation, while the DeepCAD tokenization is evaluated within the same modeling framework. Both models are trained using an identical NanoChat-based architecture to ensure a fair comparison. Their performance is assessed at two levels. Token-level accuracy measure the correctness of individual next-token predictions. Operation-level accuracy measures whether an entire CAD operation, including both the command token and all associated parameter tokens, is predicted correctly.

Studie zur Performanz von LLM-Agenten mit MCP bei steigender Aufgabenkomplexität am Beispiel der automatisierten Testfallimplementierung

Victoria Nöther

Thu, 11. 6. 2026, A10.368 and https://webroom.hrz.tu-chemnitz.de/gl/mic-cv7-ptw

Mit der zunehmenden Verbreitung von Large Language Modellen (LLMs) im Arbeitskontext gewinnen auch LLM-Agenten an Bedeutung. Obwohl LLMs beeindruckende Leistungen zeigen, machen sie weiterhin Fehler, was in Kundenprojekten besonders kritisch sein kann. Ein Auslöser für solche Fehler ist unter anderem eine steigende Komplexität der Aufgaben. Was Komplexität im Einzelfall bedeutet, lässt sich jedoch unterschiedlich definieren und ist in der praktischen Anwendung bisher wenig detailliert analysiert. Diese Arbeit fokussiert sich auf die Implementierung von Testfällen im automotive Softwaretest mithilfe eines LLM-Agenten. Dazu bewerten Domänenexperten die Testfälle hinsichtlich ihrer Komplexität. Anschließend implementiert ein eigens entwickelter LLM-Agent diese Testfälle im Softwaretestprogramm ecu.test. Die dafür benötigten Funktionen (Tools) werden dem Agenten über das Model Context Protocol (MCP) bereitgestellt. Die Task Success Rate des Agenten liegt bei 75% für leichte Testfälle und nimmt mit steigender Komplexität ab. Darüber hinaus werden die spezifischen Fehlerursachen des Agenten untersucht. Die Auswertung zeigt, dass Fehler sehr robust an bestimmten Strukturen im Testfall auftreten. Die verwendete Methode zur Einteilung in Komplexitätsstufen erweist sich als geeignet, um die Komplexität einer Aufgabe für einen LLM-Agenten in einer praktischen Anwendungsdomäne intuitiv einzuschätzen. Die zugrunde liegenden Faktoren der Komplexität erweisen sich als vielschichtig und werden im Rahmen dieser Arbeit diskutiert.

Bayesian Without a Cause: How Unsupervised Learning Shapes Multisensory Space Representations

Valentin Forch

Thu, 4. 6. 2026, A10.368 and https://webroom.hrz.tu-chemnitz.de/gl/mic-cv7-ptw

Multisensory spatial perception refers to how the brain combines signals from different senses, such as vision, touch, and proprioception, to make sense of the relative positions of our body and other objects in space. A classic example is the rubber hand illusion, in which seeing a fake hand being touched in synchrony with one?s real, hidden hand can shift perceived hand location and even induce a sense of ownership over the fake hand. In this talk, I present a neurocomputational model of multisensory integration in a rubber-hand-illusion-like setting. The model uses population-coded visual, proprioceptive, and tactile inputs, and learns feedforward and recurrent connections through Hebbian and anti-Hebbian plasticity. I connect this to theories of population coding and show that, under this framework, receptive field width needs to scale with sensory reliability for optimal decoding. I also highlight a problem created by classical Oja normalization: while it stabilizes Hebbian learning, it also forces different sensory inputs to compete for a shared synaptic weight budget, which can impair neural coding of information. I show that applying normalization separately to modality-specific input groups preserves sensory information and improves integration. The final model reproduces Bayesian-causal-inference-like behavior and generates proprioceptive drift patterns under touch and no-touch conditions. More broadly, the work links local synaptic learning rules to probabilistic models of body ownership and multisensory perception.

The Long Tail of Meaning: Compositionality in Machine Learning Models vs. Humans

Valentin Forch

Thu, 28. 5. 2026, A10.368 and https://webroom.hrz.tu-chemnitz.de/gl/mic-cv7-ptw

Natural language and visual scenes are full of rare combinations: unusual object-attribute pairings, uncommon spatial relations, and novel arrangements of familiar parts. Human cognition handles these long tails efficiently. For example, we can understand a description like ?A dog is riding on a pink horse,? even if we have never encountered this situation before. This flexibility is often linked to grammar-like, generative symbolic representations: complex meanings can be built by combining simpler parts according to rules. Machine ?cognition? in current vision-language models works very differently. These models perform impressively on frequent patterns learned from large datasets, but they are much more likely to fail when familiar elements appear in unfamiliar combinations. Such failures suggest limits in compositional generalization: the ability to recombine known concepts systematically in new situations. This seminar focuses on how compositional generalization is evaluated in vision and vision-language models, and how this relates to different cognitive skills and perception. Thus, I aim to shed light on how machine learning models could move beyond pattern matching toward more systematic, continual, and human-like recombination of knowledge.

Themenfindung für das Forschungsseminar

Fred Hamker

Thu, 7. 5. 2026, A10.368

Beratung über Themen für interessierte Studierende.

Introduction to Deep Reinforcement Learning and Successor Representations

Erik Syniawa

Thu, 30. 4. 2026, A10.368 and https://webroom.hrz.tu-chemnitz.de/gl/mic-cv7-ptw

In this talk I introduce basic deep reinforcement learning (RL) concepts from the ground up, covering imitation learning, policy gradients, and actor-critic methods, before asking what these approaches get wrong. Standard model-free RL conflates two distinct problems: learning where a policy goes and how rewarding those places are. This coupling makes reward adaptation expensive and transfer across tasks structurally difficult. I motivate the Successor Representation (Dayan, 1993) as a principled decoupling of dynamics from reward, and show how Successor Features (Barreto et al., 2017) can scale this idea to large state spaces. I close with a brief outlook on latent self-predictive representations as a modern version of these principles. All concepts are developed intuitively from first principles, (hopefully) making them accessible to everyone. But a small disclaimer: there will be math and there is no elegant way to fully hide a Bellman equation ;)

Bridging the Latency Gap: A Neurocomputational Model of Predictive Remapping

Nikolai Stocks

Thu, 23. 4. 2026, A10.368 and https://webroom.hrz.tu-chemnitz.de/gl/mic-cv7-ptw

Visual perception remains remarkably stable across saccadic eye movements despite rapid and large displacements of the retinal image. This stability is thought to rely, in part, on predictive remapping processes that update spatial representations in anticipation of eye movements. However, the precise functional role and temporal characteristics of these mechanisms remain debated. In this study, we investigate predictive remapping within a neurocomputational model of spatial perception based on quasi-multiplicative gain fields. The model integrates retinal input with corollary discharge (CD) and proprioceptive eye-position (PC) signals to generate distributed spatial representations. We emulate a visual foraging paradigm and reproduce receptive field dynamics reported in electrophysiological recordings from lateral intraparietal area (LIP). Our results show that CD-driven predictive remapping is sufficient to account for the rapid transition between pre- and postsaccadic representations observed experimentally, particularly during the early postsaccadic interval. By systematically manipulating model components, we demonstrate that the latency period following saccade offset - during which bottom-up visual input is not yet available - imposes a fundamental computational constraint that necessitates predictive mechanisms. In contrast, presaccadic activity patterns are not well explained by CD-based remapping and are more consistent with attentional or top-down influences. Together, these findings suggest a functional dissociation between presaccadic modulation and postsaccadic continuity. Predictive remapping, driven by corollary discharge, appears to primarily serve to bridge the latency-induced gap in visual processing, ensuring a seamless transition between perceptual states across saccades.

Reframing Continual Motor Learning in Robotics: Hypernetwork-Driven Control of Spiking Reservoir Dynamics

Johannes Hofele

Thu, 16. 4. 2026, A10.368 and https://webroom.hrz.tu-chemnitz.de/gl/mic-cv7-ptw

Der zunehmende Einsatz von Robotersystemen in dynamischen Umgebungen erfordert Regelungsstrategien, die nicht nur präzise, sondern auch anpassungsfähig sind und kontinuierlich lernen können. Herkömmliche Regelungsmethoden und rein datengesteuerte Ansätze haben oft Schwierigkeiten, auf neue Aufgaben zu verallgemeinern, und neigen bei sequenziellem Training zu katastrophalem Vergessen. Inspiriert von biologischen Lernmechanismen untersucht diese Arbeit ein Steuerungsframework, das kontextabhängige Hypernetzwerke mit Spiking-Reservoirs für die Echtzeit-Motorsteuerung kombiniert. Der Ansatz nutzt ein Hypernetzwerk, um die rekurrenten Gewichte eines Spiking-Reservoirs zu generieren, was eine aufgabenabhängige Modulation der Netzwerkdynamik ermöglicht. Die Architektur wird an einem Roboterarm mit zwei Gelenken evaluiert, der mehrere Trajektorienabbildungsaufgaben ausführt, darunter lineare, kreisförmige und Acht-förmige Bewegungen. Die Ergebnisse zeigen, dass das Modell einfache Trajektorien zuverlässig mit hoher Genauigkeit lernen und reproduzieren kann, während komplexere Bewegungen aufgrund der limitierten Ausdruckskraft des Reservoirs weiterhin eine Herausforderung darstellen. Darüber hinaus zeigt das System eine teilweise Fähigkeit, katastrophales Vergessen zu mildern, indem es unter kontinuierlichen Lernbedingungen Schlüsselmerkmale zuvor gelernter Aufgaben beibehält. Insgesamt unterstreicht diese Arbeit das Potenzial von hypernetzwerkgesteuerten Spiking-Systemen für adaptive Motorsteuerung und kontinuierliches Lernen, während sie gleichzeitig aktuelle Einschränkungen hinsichtlich Ausdruckskraft und Skalierbarkeit transparent aufzeigt.

Informationsveranstaltung für Studenten für das Sommersemester

Fred Hamker

Thu, 9. 4. 2026, A10.368

Camera Simulation Credibility Assessment for Automotive Perception Approaches

Arya Anup

Tue, 31. 3. 2026, https://webroom.hrz.tu-chemnitz.de/gl/mic-cv7-ptw

Synthetic data is widely used for testing and validation for Autonomous driving tasks. This thesis investigates different methods to assess credibility of using synthetic data for real-world perception tasks in autonomous driving. A structured multi-level evaluation framework is conducted on object detection, multi-object tracking and semantic segmentation tasks using the dataset pairs KITTI-VKITTI 2 and Cityscapes-Synscapes. Leave-One-Out (LOO) training splits are used to analyze robustness across varying scene compositions. For assessing input-level similarity between real and synthetic training splits, Frechet Inception Distance (FID) and Kernel Inception Distance (KID) is used. Task-level performance differences are assessed using bootstrap-based statistical equivalence testing. Distribution-level and scene-level deviations are further analyzed to characterize the magnitude and structure of performance shifts. Featurelevel representation similarity between models trained on real and synthetic data is assessed using Centered Kernel Alignment (CKA) metrics. For the selected dataset pairs, the results indicate measurable feature distribution differences between real and synthetic training data that exceed intra-domain baselines. At the task level, aggregate performance metrics such as mAP@0.5, MOTA, and mIoU frequently fail statistical equivalence under stricter tolerance margins, whereas dominant object categories remain comparatively stable in some settings. Performance deviations are consistent across splits, and representation analysis reveals strong alignment in early network layers with increasing divergence in higher task-specific layers. However, no consistent split-wise relationship is observed between input-level similarity, representation similarity, and task-level performance differences. The findings demonstrate that credibility assessment requires integrating multiple complementary evaluation levels, as no single metric fully explains sim-to-real performance behavior.