|
|
February 19, 2026
|
|
|
Faculty of Computer Science, AGH University of Krakow
|
Biological neural networks differ fundamentally from modern artificial neural networks. In particular, biological axons allow signal propagation in both directions, while contemporary artificial neural networks are almost exclusively based on unidirectional information flow. Moreover, biological organisms make decisions under uncertainty, suggesting that neural computation should propagate not only single values but also entire probability distributions. In this talk, Prof. Jarosław Duda will introduce a new class of biology-inspired neural networks in which neurons contain local joint distribution models represented as polynomials. This approach enables multi-directional propagation of information by simply switching indices, and allows the propagation of full probability distributions described by vectors of moments, rather than just expected values.
The proposed framework, inspired by Kolmogorov–Arnold Networks (KANs), provides new learning possibilities beyond classical backpropagation, including direct parameter estimation, tensor decomposition methods, and training based on the information bottleneck principle.
Jarosław Duda, PhD, is a mathematician and computer scientist at the Jagiellonian University. His research interests include data compression, information theory, probability theory, and biology-inspired computational models. He is the inventor of Asymmetric Numeral Systems (ANS), a breakthrough entropy coding method that forms the basis of modern compression standards such as Zstandard and is widely used in industrial applications worldwide.
In recent years, his research has focused on developing novel neural network architectures that more closely reflect the properties of biological nervous systems, particularly in terms of uncertainty representation and multi-directional information processing.
Online streaming:
Streaming link: https://agh-mche.webex.com/meet/informatyka
The seminar will be held in English. We warmly invite everyone interested!