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
Predictive Coding LightJochen Triesch Thu, 13. 3. 2025, 16:30, 1/336 Current machine learning systems consume vastly more energy than biological brains. Neuromorphic systems aim to overcome this difference by mimicking the brain?s information coding via discrete voltage spikes. However, it remains unclear how both artificial and natural networks of spiking neurons can learn energy-efficient information processing strategies. Here we propose Predictive Coding Light (PCL), a recurrent hierarchical spiking neural network for unsupervised representation learning. In contrast to previous predictive coding approaches, PCL does not transmit prediction errors to higher processing stages. Instead it suppresses the most predictable spikes and transmits a compressed representation of the input. Using only biologically plausible spike-timing based learning rules, PCL reproduces a wealth of findings on information processing in visual cortex and permits strong performance in downstream classification tasks. Overall, PCL offers a new approach to predictive coding and its implementation in natural and artificial spiking neural networks. |
Replicating physiological data with a new basal ganglia - prefrontal cortex modelSusanne Holtzsch Wed, 26. 3. 2025, 10:00, 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. |
Vergangene Veranstaltungen
Implicit Neural Representations for Spatio-Temporal Modeling of Satellite-Derived Surface ReflectanceVladyslav Shapran Mon, 17. 2. 2025, https://webroom.hrz.tu-chemnitz.de/gl/jul-2tw-4nz Implicit Neural Representations (INRs) have demonstrated tremendous success in solving various inverse problems, spearheaded by Neural Radiance Fields (NeRFs) in the field of 3D scene reconstruction. Recently INRs have also found applications in Satellite Remote Sensing, where the research has been mostly focused on satellite photogrammetry with the goal to recover the true surface reflectance of the scene instead of radiance. These works have mostly focused on modeling of shadows, illumination, and on accounting for other environmental effects; however, they have given little attention to the INR model itself, and how the inductive biases of its architecture regularize the reconstructed scene. In addition, the adoption of INRs has been rather slow for more conventional satellite remote sensing applications other then photogrammetry, which are mostly focused on two-dimensional data and on spatio-temporal modeling. In this work we attempt to bridge this gap by studying how INRs can be used to model the surface reflectance in two-dimensional case, using atmosphere-corrected reflectance images from the Sentinel-2 satellite, as well as an additional synthetic data set that we create. Most importantly, we decide to forego the modeling of environmental effects and instead study the differences between various INR architectures and how suitable they are for reconstructing geospatial data. We further extend our research to the spatio-temporal case by studying how a trained INR model generalizes to different time-steps of the same scene. To this end, we run an extensive hyperparameter optimization for NeRF, Fourier-NeRF, SIREN, Gaussian-based INR, and wavelet-based WIRE architectures. We further propose to slightly modify the WIRE architecture by adding a sigmoid activation right before the last layer. This new architecture, which we call S-WIRE, produces impressive reconstructions of surface reflectance scenes, which far surpass all other models that we test. Most spectacularly, S-WIRE can almost perfectly reconstruct a previously unseen time-step of the training scene using only a few sample measurements. We hope that our work can facilitate the use of INRs for other remote sensing tasks in general, and for spatial or spatio-temporal geospatial data in particular. |
Time-Dynamic Cherry Blossom Prediction Using FORCE-Trained Echo State NetworksGantogoo Oyunbat Thu, 13. 2. 2025, 1/346 and https://webroom.hrz.tu-chemnitz.de/gl/jul-2tw-4nz The human brain performs complex predictive tasks while consuming minimal energy, highlighting the vast efficiency gap between biological and artificial computing systems. While modern computers offer impressive computational capabilities, their energy requirements limit widespread deployment of predictive models. Neural systems inspire reservoir computing approaches that aim to bridge this gap, combining computational power with energy efficiency. In this study, we explore this potential by applying reservoir computing to a real-world time-dynamic forecasting challenge: cherry blossom date prediction, a task of cultural, ecological, and economic importance in Japan. This study uses an Echo State Network (ESN) with FORCE (First-Order Reduced and Controlled Error) recursive least squares training to model the time-dynamic relationship between temperature, humidity, geographical coordinates, and cherry blossom dates. Public datasets, including meteorological data and historical bloom records, are processed to align with the temporal dynamics of ESN, ensuring structured input for prediction. FORCE training adapts the readout layer to non-linear, dynamic relation of the inputs and outputs. Compared to transformer-based models, reservoir computing requires significantly fewer learnable parameters while maintaining relatively accurate predictions, highlighting the potential of reservoir computing in real-world predictive challenges. With fixed internal weights and simplified training, ESNs are well-suited for neuromorphic hardware, offering a scalable and energy-efficient alternative to conventional forecasting methods. |
Adaptive Horizontal Pod Autoscaling (AHPA) based on Reinforcement Learning in Kubernetes for Machine LearningNatnicha Rodtong Tue, 11. 2. 2025, 1/367 and https://webroom.hrz.tu-chemnitz.de/gl/jul-2tw-4nz The field of artificial intelligence (AI) is rapidly expanding across various sectors, with many industries still in the early stages of adoption. Machine learning (ML) inference tasks, in particular, require substantial computational power, especially in terms of CPU and memory. Effective resource allocation for applications requires expertise to optimize system performance, throughput, and inference types, among other factors. Both under- and over-allocation can negatively impact application performance. Over-allocation can strain the scheduler, disrupt other active nodes, and lead to higher costs, reduced responsiveness, and decreased node capacity. Kubernetes, with its robust architecture, is a preferred choice for AI systems, as it efficiently handles AI workloads and dynamically scales applications based on data processing needs. This presents a valuable opportunity to leverage ML for assessing the computing resource requirements of such systems, with the goal of improving both service quality and resource efficiency. We thereby study, firstly, how can ML, especially reinforcement learning (RL), be utilized to predict future computing resource needs and dynamically adjust the number of Kubernetes Pods in a horizontal manner, enabling scaling in and out, to handle fluctuating demand of requests, secondly, what are the most suitable RL algorithms for implementing a Kubernetes horizontal autoscaler, and, finally, how does the reliability of an RL-based autoscaler for ML inference compare to a basic Kubernetes Horizontal Pod Autoscaler (HPA) in terms of average response time and packet loss. This study introduces the Adaptive Horizontal Pod Autoscaler (AHPA), which utilizes RL with a Deep Q-Network (DQN) to dynamically adjust the number of Kubernetes Pods for horizontal scaling, enabling both scaling in and scaling out. We evaluate the performance and reliability of AHPA in image classification tasks, comparing its effectiveness against a traditional horizontal autoscaler in Kubernetes. The results showed that RL, particularly the model-free, value-based approach using DQN, is an effective method for addressing the challenges of adaptive horizontal scaling in Kubernetes environments, efficiently accommodating the fluctuating volume of incoming requests for ML applications with impressive managing the dynamic nature of resource allocation. Furthermore, our proposed AHPA outperforms the standard Kubernetes HPA in terms of average latency, packet loss, and Pod utilization under most experimental conditions. However, under high-load conditions, Kubernetes HPA, which relies on fixed thresholds, outperformed AHPA. |
Contrastive learning - part 2René Larisch Thu, 30. 1. 2025, 1/346 and https://webroom.hrz.tu-chemnitz.de/gl/jul-2tw-4nz Deep neural networks have shown very good performance in object recognition and object detection tasks. To perform such a task, a good representation of different input features is necessary. It has been shown that representation learning methods, such as contrastive learning, provide good feature representation when used as a pretext task. In addition, recent contrastive learning methods are trained in an unsupervised manner, allowing the network to be trained on a large corpus of data without the need for labels. By providing a useful representation of input features, contrastive learning methods are also interesting as simple models of the visual cortex, providing input to higher cortical areas. In my talk, I will give a brief overview of the basic principles of contrastive learning, present different concepts for the further development, and discuss open problems in contrastive learning. |
Evolving Neuromorphic Micro-Architectures for Motor ControlMaximilian Titz Thu, 23. 1. 2025, 1/346 and https://webroom.hrz.tu-chemnitz.de/gl/jul-2tw-4nz Artificial intelligence (AI) research has predominantly focused on large-scale models like deep learning and large language models, which require significant computational resources. However, numerous AI applications exist that operate under strict constraints in terms of power consumption, memory, and form factor. Microrobotics represents one such application, in which these constraints are particularly amplified. This presents unique challenges for the architecture and algorithms used to control the robots. Spiking neural networks (RSNNs) implemented on specialized neuromorphic chips offer a higher power efficiency than conventional computing methods. Through emulating biological principles, such as in-memory computation, they enable bypassing the limits of conventional computing models. However, training RSNNs in resource-constrained environments remains a significant challenge. This thesis proposes a novel methodology that combines reservoir computing with evolutionary optimization techniques to minimize the memory overhead of training. The approach is evaluated on a reinforcement learning task using a software simulation of a neuromorphic chip and then extended by optimizing the input configurations to increase its effectiveness. The results highlight the potential of evolutionary optimization in training RSNNs for applications constrained by power and memory. |
Contrastive learning - part 1René Larisch Thu, 19. 12. 2024, 1/346 and https://webroom.hrz.tu-chemnitz.de/gl/jul-2tw-4nz Deep neural networks have shown very good performance in object recognition and object detection tasks. To perform such a task, a good representation of different input features is necessary. It has been shown that representation learning methods, such as contrastive learning, provide good feature representation when used as a pretext task. In addition, recent contrastive learning methods are trained in an unsupervised manner, allowing the network to be trained on a large corpus of data without the need for labels. By providing a useful representation of input features, contrastive learning methods are also interesting as simple models of the visual cortex, providing input to higher cortical areas. In my talk, I will give a brief overview of the basic principles of contrastive learning, present different concepts for the further development, and discuss open problems in contrastive learning. |
A Survey of Reservoir Computing for Control TasksMaximilian Titz Thu, 17. 10. 2024, 1/346 and https://webroom.hrz.tu-chemnitz.de/gl/jul-2tw-4nz Reservoir Computing has seen a noticeable increase in popularity in recent years. Reservoirs aim to leverage the complex behaviors that naturally emerge in high-dimensional, dynamical systems in response to an input. In doing this, they constitute a simple way to circumvent the vanishing/exploding gradient problem that arises when training recurrent neural networks. This makes them effective for tasks that require time series processing. Additionally, they significantly reduce the number of parameters that need training. Despite these advantages, the training of reservoirs presents unique challenges. These include, for example, finding good hyperparameters used to initialize the network. This seminar delves into recent research papers that utilize reservoir computing for control tasks. The primary goal is to showcase and compare different approaches, highlighting how reinforcement learning techniques, such as Q-learning and Proximal Policy Optimization, are used with reservoir computing. Furthermore, it explores novel methodologies, such as optimizing the hyperparameters of a reservoir using Meta-Learning and the integration of random weight Convolutional Neural Networks with reservoir computing to process video data. By examining these diverse strategies, the seminar aims to provide an overview of the current state of RC and its potential for future advancements in the field. |
Bridging Consciousness and AI: Evaluating the Global Latent Workspace in Embodied AgentsNicolas Kuske Wed, 9. 10. 2024, 1/368a In the wake of the success of attention-based transformer networks, the discussion about consciousness in artificial systems has intensified. The global neuronal workspace theory (GNWT) models consciousness computationally, suggesting the brain has specialized modules connected by long distance neuronal links. Depending on context, inputs, and tasks, content from one module is broadcasted to others, forming the global neuronal workspace representing conscious awareness. The global latent workspace (GLW) model introduces a central latent representation around which multiple modules are built. A semi-supervised training process ensures cycle consistency, enabling content translation between modules with minimal loss. The central representation integrates necessary information from each module, with access determined by transformer-like attention mechanisms. We examine the dynamics of a virtual embodied reinforcement learning agent with a minimal GLW setup, involving deep visual sensory and motor modules. The augmented PPO agent exhibits complex goal-directed behavior in the Obstacle Tower Challenge 3D environment. Latent space representations cluster into sensorimotor affordance groups. This study links GNWT with sensorimotor contingency theory, suggesting that changes in sensory input relative to motor output constitute the neuronal correlates of conscious experience. This convergence in a machine learning setup raises the question: Can such in silico representation suffice for phenomenal spatial perception? |
Development of a Hierarchical Approach for Multi-Agent Reinforcement Learning in Cooperative and Competitive EnvironmentsRobert Pfeifer Wed, 11. 9. 2024, 1/368 and https://webroom.hrz.tu-chemnitz.de/gl/jul-2tw-4nz Multi-Agent Reinforcement Learning (MARL) enables multiple agents to learn purposeful interactions in a shared environment by maximizing their own cumulative rewards. As a subject of current research MARL is getting applied in various fields, such as robotics, traffic control, autonomous vehicles and power grid management. However, these algorithms faces challenges including partial observability, non-stationarity of the environment, coordination among agents and scalability issues when incorporating a multitude of agents. This thesis explores Hierarchical Reinforcement Learning (HRL) as a potential solution to address these challenges, using the Goal-Conditional Framework to decompose complex tasks into simpler sub-tasks, which facilitates better coordination and interpretability of behavior. The Goal-Conditional Framework learns this decomposition automatically in an end-to-end fashion by assigning the respective tasks to a hierarchy of policies. The higher-level policy proposes a goal vector either in a latent state representation or directly in the state space while operating on a potentially slower time scale. The lower-level policy receives the goal as part of its observation space and obtains an intrinsic reward based on how well it achieves the goal state. Only the top-level policy receives the reward signal from the environment. The thesis implements the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm proposed by (Lowe et al., 2020) and extends it for hierarchical learning using either a centralized or decentralized manager policy. It investigates the performance of this approach in various environments, focusing on diverse learning aspects such as agent coordination, fine-grained agent control, effective exploration and directed long-term behavior. Additionally, the thesis explores the influence of architecture, time horizon and intrinsic reward function on final performance, aiming to gain a deeper understanding of the possibilities and challenges associated with Goal-Conditional hierarchical algorithms. |
Entwicklung von KI-Modellen zur Adressdatenextraktion: Herausforderungen, Methoden und Empfehlungen zur Bewältigung von Vielfalt und KontextfehlernErik Seidel Tue, 27. 8. 2024, 1/368 and https://webroom.hrz.tu-chemnitz.de/gl/jul-2tw-4nz Die präzise Erkennung postalischer Adressen stellt eine zentrale Herausforderung dar, insbesondere, wenn Modelle auf die Erkennung anderer Entitäten, wie beispielsweise Personen, trainiert werden. Die dabei entstehenden Konflikte haben Einfluss auf deren Identifikation. Ein Beispiel hierfür ist die Verwechslung bei Namen wie "Professor Richard Paulik" mit "Professor Richard Paulik Ring". Aber auch die Unterscheidung zwischen Personen- und Ortsnamen sind eine Herausforderung. Zentrale Probleme stellen Komplexität, Diversität (Vielfalt) und Ambiguität (Mehrdeutigkeit) der Entitäten dar. Ausreichend annotierte Daten von Adressen oder Personen sind nur schwer zu finden. Die Arbeit untersucht, welche Ansätze sich am besten für die Erkennung von Adressen eignen. Im Fokus stehen Transformer Modelle. Für das Problem der Datenbeschaffung wird der Einsatz synthetisch erstellter Daten betrachtet. Das Ziel ist eine möglichst breite Vielfalt an Beispielsätzen für das Training und die Evaluation zu erreichen. Es werden verschiedene Metriken und Werkzeuge vorgestellt, mit dessen Hilfe sich Modelle und Daten bewerten und optimieren lassen. |
