Embedded Kolmogorov-Arnold Networks

​

​A part of our research deals with novel neural network architectures based on the Kolmogorov-Arnold representation theorem. The focus is on Kolmogorov-Arnold Networks (KANs), a promising approach in which the weights of a neural network are replaced by learnable nonlinear functions. This structure enables models that are more interpretable, flexible, and expressive than classical multi-layer perceptrons (MLPs).

As part of our research, we evaluate the areas of application in which the KAN architecture offers a practical advantage and how the resource requirements of KAN models can be reduced, e.g., by applying quantization or the replacement of learned nonlinear functions. Due to the explainability of KAN models, this method can offer advantages in safety-critical areas such as medicine or the automotive sector.

In line with our research, we are developing the KANLib framework . This is a modular and easily extensible KAN implementation for PyTorch. The aim is to evaluate and improve the practical applicability of KANs in embedded systems.

Contact: Julian Hoever, M.Sc. | Recent Publication

Memory-Constrained Training of Deep Neural Networks on Microcontrollers

​

​​Training deep neural networks is predominantly performed on powerful GPUs or specialized accelerators, while microcontroller-based systems are typically limited to inference. In our research, we challenge this assumption and investigate the direct training of neural networks on microcontrollers under severe memory constraints.

Our focus is on classical Cortex-M systems with only a few hundred kilobytes of RAM, no external memory, and no hardware accelerators for matrix operations. In this setting, training is not compute-bound but memory-bound: activations, gradients, and optimizer states quickly exceed the available resources, making existing training methods and optimization algorithms difficult to transfer. The goal of this research is the systematic analysis and reduction of memory consumption during training. To this end, analytical memory models are developed, and techniques such as quantization, unstructured pruning, sparse updates, and lossy compression are selectively combined and dynamically adapted during training.

A particular focus is placed on quantized parameter updates (QPU). This novel approach enables the training of neural networks directly with quantized parameters and stochastically rounded updates, allowing floating-point weights to be completely eliminated and significantly reducing the memory footprint during training. There is special interest in utilizing the freed-up memory resources to improve convergence speed and final model accuracy, for example through adaptive batch sizes, dynamic bit widths, and the gradual extension of simple optimizers toward more memory-intensive methods.

The proposed methods are evaluated on time-series and image classification tasks and implemented on real MCU hardware. In the long term, this work contributes to autonomous, privacy-preserving, and energy-efficient adaptation of neural networks directly on embedded systems.

Contact: Leo Buron, M.Sc. | Recent Publication

Deep Reinforcement Learning

Deep Reinforcement Learning (DRL) has established itself as a powerful approach for complex decision-making and control problems. Despite its considerable empirical success, fundamental theoretical challenges remain at the interface of Deep Learning (DL) and Reinforcement Learning (RL), which are the focus of this research.

A central theoretical issue in Reinforcement Learning is the so-called deadly triad: the combination of function approximation, bootstrapping, and off-policy learning, which can lead to divergence of the model parameters. Using an eigenvalue analysis of classical counterexamples, we investigate the underlying update matrices. We show that, in the counterexamples considered, the matrix governing the parameter updates possesses at least one positive eigenvalue. This provides a theoretical explanation for why DRL methods often converge stably in practice: for certain problem settings, suitable function approximations exist under which stable parameter updates are possible.

A further focus of our work is the IID assumption, a fundamental prerequisite for the use of stochastic gradient-based optimization methods in Deep Learning. In Reinforcement Learning, however, this assumption is violated, as training data is generated through sequential interaction and is neither independent nor stationary. Our research demonstrates that the success of modern DRL algorithms relies heavily on techniques such as experience replay and target networks, which deliberately alter the data structure. In this way, our work contributes to a deeper understanding of how deep learning methods can be adapted to the specific requirements of experience-based learning.

Contact: Fatih Özgan, M.Sc. | Recent Publication