Forschungsseminar
Das Forschungsseminar richtet sich an interessierte Studierende des Master- oder Bachelorstudiums. Andere Interessenten sind jedoch jederzeit herzlich willkommen! Die vortragenden Studenten und Mitarbeiter der Professur KI stellen aktuelle forschungsorientierte Themen vor. Vorträge werden in der Regel in Englisch gehalten. Den genauen Termin einzelner Veranstaltungen entnehmen Sie bitte den Ankündigungen auf dieser Seite.
Informationen für Bachelor- und Masterstudenten
Die im Studium enthaltenen Seminarvorträge (das "Hauptseminar" im Studiengang Bachelor-IF/AIF bzw. das "Forschungsseminar" im Master) können im Rahmen dieser Veranstaltung durchgeführt werden. Beide Lehrveranstaltungen (Bachelor-Hauptseminar und Master-Forschungsseminar) haben das Ziel, dass die Teilnehmer selbstständig forschungsrelevantes Wissen erarbeiten und es anschließend im Rahmen eines Vortrages präsentieren. Von den Kandidaten wird ausreichendes Hintergrundwissen erwartet, das in der Regel durch die Teilnahme an den Vorlesungen Neurocomputing (ehem. Maschinelles Lernen) oder Neurokognition (I+II) erworben wird. Die Forschungsthemen stammen typischerweise aus den Bereichen Künstliche Intelligenz, Neurocomputing, Deep Reinforcement Learning, Neurokognition, Neurorobotische und intelligente Agenten in der virtuellen Realität. Andere Themenvorschläge sind aber ebenso herzlich willkommen!Das Seminar wird nach individueller Absprache durchgeführt. Interessierte Studenten können unverbindlich Prof. Hamker kontaktieren, wenn sie ein Interesse haben, bei uns eine der beiden Seminarveranstaltungen abzulegen.
Kommende Veranstaltungen
Evaluating Speculative Decoding in Large Language Models: A Comparative Study of MEDUSA and EAGLE Architectures on Hallucination RisksAlaa Alshroukh Tue, 14. 10. 2025, 13:45, TBA and https://webroom.hrz.tu-chemnitz.de/gl/mic-cv7-ptw Large Language Models (LLMs) deliver strong performance on tasks such as summarization and question answering, but autoregressive decoding remains a latency bottleneck in real-time applications due to its sequential, compute-intensive nature. This bottleneck arises in part because each decoding step requires transferring the entire set of model parameters from high-bandwidth memory (HBM) to the accelerator's cache, which limits throughput and underutilizes computational resources. To address this challenge, two acceleration methods are examined: EAGLE (Extrapolation Algorithm for Greater Language-model Efficiency) - a speculative sampling framework, and MEDUSA - a multi-head inference approach. Both techniques reduce latency by generating multiple tokens per step. Performance is evaluated by comparing each accelerated model to its baseline on summarization tasks, enabling controlled within-model analyses of speed and output quality. Particular attention is given to hallucination - fluent yet factually unsupported output - since generating multiple tokens at once can amplify small inaccuracies into extended errors. Results indicate a clear speed - factuality trade-off: EAGLE achieves the lowest latency and highest throughput while broadly preserving source meaning, but exhibits reduced factual stability on dense inputs. Medusa is slower but demonstrates stronger adherence to the source and more consistent factuality. Overall, the findings emphasize that accelerated LLM decoding requires careful balancing of efficiency and reliability, especially in fact-sensitive applications. |
Part 2: Key Concepts from the Nengo Summer School 2025Erik Syniawa Tue, 21. 10. 2025, 13:45, TBA This talk presents key concepts from the Nengo Summer School, focusing on the Neural Engineering Framework (NEF) and its implementation through Nengo. The NEF provides three core principles: representation (how populations encode information), computation (how connections compute functions), and dynamics (how recurrent networks implement differential equations). These principles enable the construction of large-scale spiking neural networks for cognitive modeling (SPAUN). The talk covers the Semantic Pointer Architecture (SPA), a cognitive architecture that uses compressed neural representations to connect low-level neural activity with high-level symbolic reasoning. It also examines Legendre Memory Units (LMUs), an efficient method for processing temporal information in neural networks. Additionally, I will present my project work: a biologically inspired architecture that combines a visual pathway with the basal ganglia to play different Atari games. |
Vergangene Veranstaltungen
Erweiterung eines Aufmerksamkeitsmodells um eine gelernte IT-SchichtLucas Berthold Thu, 11. 9. 2025, 1/336 and https://webroom.hrz.tu-chemnitz.de/gl/mic-cv7-ptw Es gehört noch zu den zentralen offenen Fragen der Neurowissenschaften, warum die Repräsentationen von Objekten im visuellen System des Menschen stabil bleiben, auch wenn sich die Augen ständig bewegen. Es wird angenommen, dass der inferotemporale Kortex (IT) eine entscheidende Rolle bei der Objekterkennung spielt, indem dort Objekte in Form von stabilen Repräsentationen kodiert werden (Li und DiCarlo, 2008). Trotz der Robustheit dieser Repräsentationen gegenüber Veränderungen der Position eines Objekts auf der Retina haben Li und DiCarlo (2008) gezeigt, dass durch gezielte Manipulation an einer bestimmten retinalen Position einzelne IT-Neuronen, die zuvor selektiv auf bestimmte Objekte reagierten, nach der Manipulation an Selektivität verloren haben. Die zentrale Frage dieser Arbeit ist, ob sich die von Li und DiCarlo (2008) beobachteten Effekte auch in einem künstlichen neuronalen Modell nachbilden lassen, wenn allein die Verbindungen eines höheren visuellen Areals von einem bestehenden Aufmerksamkeitsmodell (Beuth, 2019) zu einer IT-Schicht, unter Berücksichtigung von lateraler Inhibition, gelernt werden. Dazu ist erforderlich, dass die IT-Neuronen sowohl selektiv für bestimmte Objekte als auch positionsinvariant sind. Die Ergebnisse der Arbeit zeigen, dass die Neuronen der gelernten IT-Schicht objektspezifisch und positionsinvariant selektieren können. Die Beobachtungen aus dem Manipulationsexperiment von Li und DiCarlo (2008) konnten jedoch nicht repliziert werden. |
Part 1: Key Concepts from the Nengo Summer School 2025Erik Syniawa Thu, 17. 7. 2025, 1/336 and https://webroom.hrz.tu-chemnitz.de/gl/mic-cv7-ptw This talk presents key concepts from the Nengo Summer School, focusing on the Neural Engineering Framework (NEF) and its implementation through Nengo. The NEF provides three core principles: representation (how populations encode information), computation (how connections compute functions), and dynamics (how recurrent networks implement differential equations). These principles enable the construction of large-scale spiking neural networks for cognitive modeling (SPAUN). The talk covers the Semantic Pointer Architecture (SPA), a cognitive architecture that uses compressed neural representations to connect low-level neural activity with high-level symbolic reasoning. It also examines Legendre Memory Units (LMUs), an efficient method for processing temporal information in neural networks. Additionally, I will present my project work: a biologically inspired architecture that combines a visual pathway with the basal ganglia to play different Atari games. |
Influence of Tokenization Strategies on the Prediction of CAD Model DescriptionsSayeda Hadisa Habibi Thu, 3. 7. 2025, 1/336 and https://webroom.hrz.tu-chemnitz.de/gl/mic-cv7-ptw Computer-aided design (CAD) significantly impacts manufacturing, architecture, and engineering industries. Automating the generation and modification of 3D CAD models helps reduce repetitive tasks, saves resources, and allows engineers to focus on innovation and complex design challenges. Modern CAD tools like Fusion 360 support programmatic design modifications, enabling AI-based automation in design workflows. As CAD systems become more advanced, there is a growing for deep learning methods that understand and manipulate CAD data. However, applying deep learning to CAD remains a challenge, primarily due to the difficulty of finding effective representations of 3D models. Traditional formats like voxels and point clouds are memory-intensive and lose geometric detail. The B-Rep format, while precise, is too complex to feed directly into neural networks. Moreover, tokenization methods like WordPiece, originally developed for text, struggle with the structured, continuous nature of CAD data. This research investigates CAD-specific tokenization approaches by comparing the standard WordPiece method with the sketch-extrude-based technique introduced by DeepCAD, which converts CAD models into text-based representations of design steps, making them suitable for neural network architectures. A masked token prediction task using a TinyBERT-style model is used to evaluate both methods in the context of CAD understanding and representation. |
Parameter-Efficient Fine-tuning of LLMs for Domain-Specific Programming LanguagesSupriya Dravid Tue, 3. 6. 2025, https://meet.google.com/xvk-gnhi-vxh This thesis investigated how Large Language Models (LLMs) could be adapted to generate reliable code for domain-specific languages, with a focus on ABAP, the core programming language used in SAP systems. Unlike widely used general-purpose languages, domain-specific languages like ABAP are significantly underrepresented in public datasets. This lack of representation poses a challenge, as zero-shot code generation often results in outputs that are inaccurate, incomplete, or misaligned with enterprise coding standards. To address this issue, the study explored parameter-efficient fine-tuning techniques, LoRA (Low-Rank Adaptation) and QLoRA (Quantized LoRA). These methods selectively fine-tuned small subsets of parameters while keeping the original pretrained model weights mostly unchanged. This approach enabled effective domain adaptation with much lower computational and memory overhead compared to full-model retraining. The methods were applied to small and mid-sized open-source models such as Phi-3-mini and Codegen-2B, and their performance was evaluated using metrics including Pass@k, CodeBLEU, and perplexity. In addition to this, the research included qualitative assessments from experienced ABAP developers, who rated the generated code based on functional correctness, performance, robustness, and maintainability. The results demonstrated that fine-tuned models significantly outperformed zero-shot baselines across all evaluation criteria. This work provides practical insights into how LLMs could be adapted for specialized programming environments, offering a cost-effective pathway for integrating generative AI into enterprise software development workflows. |
Developing a Topology Optimization Tool for Human Experts - Implementing Machine Learning Approaches into the Hybrid Cellular Automata MethodChristoph Ruff Thu, 17. 4. 2025, 1/336 and https://webroom.hrz.tu-chemnitz.de/gl/mic-cv7-ptw Due to the large nonlinearities that occur in crash simulations, topology optimization of crashworthiness structures poses a challenging field. To date, the global minima and, consequently, the optimal design of crashworthiness structures remain often unknown. Therefore, specialized methods, such as the Hybrid Cellular Automata (HCA) method, have been implemented. The HCA method is a heuristic-based method with the goal to homogenize the Internal Energy Density (IED) of the design. While this approach is suitable for certain crashworthiness structures, it has several limitations that make this method unsuitable for the use in an industrial context. Therefore, in this work, the HCA method has been adapted using Machine Learning (ML) algorithms to address these challenges. K-Means clustering was employed to reduce the design variables and enhance the interpretability and manufacturability of the optimized design. eXplainable Artificial Intelligence (XAI) methods were then applied to obtain the importance of each feature for each cluster. Furthermore, dimensionality reduction algorithms, namely Principal Component Analysis (PCA) and AutoEncoder (AE), were implemented to extract and utilize additional information from the crash analysis. This information was combined into a single feature, which was subsequently homogenized instead of the IED. The developed algorithms were investigated using an application example. While no method has surpassed the results of the state-of-the-art HCA method, numerous promising approaches have been proposed and were analyzed. A proof-of-concept has been demonstrated for the combined feature homogenization, whereas further research is needed to achieve good results for the reduction of design variables. Nonetheless, this thesis presents valuable results that can serve as a foundation for future research. |
Modellierung von dynamischer Aufmerksamkeitsmodulation in der visuellen SucheIna Lin Thu, 27. 3. 2025, 1/367 and https://webroom.hrz.tu-chemnitz.de/gl/jul-2tw-4nz Um ein gesuchtes Objekt in einer Umgebung ausfindig zu machen, wird eine interne Repräsentation der Merkmale des Objekts gebildet. Diese interne Repräsentation, auch Attention Template genannt (Duncan & Humphreys, 1989; Witkowski & Geng, 2019), erhöht durch die neuronale Gain-Modulation die Aktivität der relevanten, merkmalspezifischen Neuronen des gesuchten Objekts (Maunsell & Treue, 2006; Reynolds & Heeger, 2009; Treue & Trujillo, 1999). Merkmale wie zum Beispiel Farbtöne, Orientierung etc. werden in der Suche zwischen relevanten Reizen (Targets) und irrelevanten Reizen (Nontargets) unterschieden. Die Differenz des neuronalen Hervorhebens zwischen den beiden Reizarten, das Signal-Rausch-Verhältnis (SNR), ermöglicht eine effiziente Auswahl eines Objekts (Peltier & Becker, 2016). Die optimale Aufmerksamkeitsmodulation (optimal tuning) bezieht sich auf die Steuerung der Suche durch Neuronen, die Merkmale kodieren, welche sich vom Distraktor unterscheiden. Nach den Erkenntnissen von Maith et al. (2021), dass sich das Attention Template durch Lernprozesse in den Basalganglien bildet, wird in der Neurocomputing- Studie ?How the Basal Ganglia Could Contribute to Optimal Tuning of the Attentional Template?, ein Modell aus dem visuellen System und den Basalganglien zusammengeführt, um diesen Ansatz auf das Cueing-Paradigma von Kerzel (2020) zu erweitern. Das Modell ist in der Lage, ein farbiges Target unter Distraktoren durch Aufmerksamkeitsmodulation ausfindig zu machen. Allerdings zeigen Kerzel und Cong (2021), dass die Form des Attention Templates von der Art der Suchaufgabe abhängt: Bei Singleton Search (SiS) (farbiges Target, einfarbige oder nicht farbige Nontargets) ist die Aufmerksamkeitskarte eher breit gefächert, während bei Feature Search (FeS) (farbiges Target, farbig variierende Nontargets) das Attention Template schmaler ist. In FeS müssen aktiv mehr nichtrelevante Merkmale der Nontargets unterdrückt werden, was zu einer verstärkten Aktivierung des indirekten Pfades führt, während bei SiS weniger Unterdrückung erforderlich ist. Dadurch ist die Aktivierung des indirekten Pfades bei SiS geringer. Zudem kann eine Wiederholung der gleichen Target-Nontarget-Beziehung das Attention Template beeinflussen (Kerzel & Huynh Cong, 2024). |
Replicating physiological data with a new basal ganglia - prefrontal cortex modelSusanne Holtzsch Wed, 26. 3. 2025, 1/367 and https://webroom.hrz.tu-chemnitz.de/gl/jul-2tw-4nz The basal ganglia play a fundamental role in category learning and action selection. They learn in a supervised way, with their connections rapidly adapting based on a reward prediction error regulated by dopamine levels. By learning stimulus-response associations, they teach the cortico-cortical connection from the inferior temporal cortex (IT) to the prefrontal cortex (PFC), so that the PFC acquires stable, abstract category knowledge. In the published model (Villagrasa, 2018), this dynamic was simulated when the model performed a Prototype Distortion Task (PDT). In this task, previously done with macaque monkeys, the goal is to classify visual dot stimuli of two categories. The model was able to replicate the development of category selectivity recorded in the monkeys' PFC. However in an online learning task, catastrophic forgetting occurred. So the model was updated by including inhibitory interneurons to make the PFC more sparse. Additionally a certainty signal was added which controls, that when the model classifies a stimulus with high confidence, the PFC directly determines the response. The thesis investigates whether the updated model can still replicate the physiological data and perform well in the PDT, as well as how increased sparsity affects PFC category selectivity. Furthermore, it explores whether specific model parameters or adjustments can help the new model replicate behavioral data more accurately. |
