Biology-Inspired Neural Networks with Multi-Directional Propagation

Summary: While biological axons can propagate in both directions, current ANNs are focused on unidirectional propagation. Also usually they only propagate values, while uncertainty is shown also crucial for making decisions of biological organisms - suggesting to propagate also variance, probability distributions. I will talk about novel KAN-like approach to ANNs repairing these lack by using neurons containing inexpensive local joint distribution model as polynomial, allowing to freely change propagation direction by just switching indexes, also propagate entire probability distributions represented by vectors of moments. Beside backpropagation, it also allows many additional training approaches, like direct estimation, tensor decomposition, and through information bottleneck.

Speaker: Dr Jarek Duda is an assistant professor at Jagiellonian University. He holds degrees in computer science (PhD), mathematics (MSc) and physics (PhD). He is mainly focused on physics foundations, information theory, statistical analysis, and is known for introduction of asymmetric numeral systems.Summary: While biological axons can propagate in both directions, current ANNs are focused on unidirectional propagation. Also usually they only propagate values, while uncertainty is shown also crucial for making decisions of biological organisms - suggesting to propagate also variance, probability distributions. I will talk about novel KAN-like approach to ANNs repairing these lack by using neurons containing inexpensive local joint distribution model as polynomial, allowing to freely change propagation direction by just switching indexes, also propagate entire probability distributions represented by vectors of moments. Beside backpropagation, it also allows many additional training approaches, like direct estimation, tensor decomposition, and through information bottleneck. Speaker: Dr Jarek Duda is an assistant professor at Jagiellonian University. He holds degrees in computer science (PhD), mathematics (MSc) and physics (PhD). He is mainly focused on physics foundations, information theory, statistical analysis, and is known for introduction of asymmetric numeral systems.

Moderators: Dr. Pawel Gora, CEO of Quantum AI Foundation (https://www.qaif.org)

and Dr. Sebastian Zajac , member of QPoland (https://qworld.net/qpoland/)