INRC Forum: Wolfgang Maass, Christoph Stoeckl & Yukun Yang

Local prediction-learning in high-dimensional spaces enables neural networks to plan

Abstract: Being able to plan a sequence of actions in order to reach a goal, or more generally to solve a problem, is a cornerstone of higher brain function. But compelling models which explain how the brain can achieve that are missing. We show that local synaptic plasticity enables a neural network to create high-dimensional representations of actions and sensory inputs so that they encode salient information about their relationship. In fact, it can create a cognitive map that reduces planning to a simple geometric problem in a high-dimensional space that can easily be solved by a neural network. This method also explains how self-supervised learning enables a neural network to control a complex muscle system so that it can handle locomotion challenges that never occurred during learning. The underlying learning strategy bears some similarity to self-attention networks (Transformers). But it does not require non-local learning rules or very large datasets. Hence it is suitable for implementation in highly energy-efficient neuromorphic hardware, in particular for on-chip learning on Loihi 2.
One goal of our presentation will be to initiate discussions about the relation of this learning-based use of large vectors to other VSA approaches, its relation to Transformers, and possible applications in robotics.

Bio: Wolfgang Maass is a Professor of Computer Science at Technische Universität Graz. He received his PhD (1974) and Habilitation (1978) in Mathematics from Ludwig-Maximilians-Universität in Munich. He conducted research at MIT, the University of Chicago, and UC Berkeley, as a Heisenberg Fellow of the Deutsche Forschungsgemeinschaft. He has been the Editor of Machine Learning (1995-1997), Archive for Mathematical Logic (1987-2000), and Biological Cybernetics (2006-present). He was also a Sloan Fellow at the Computational Neurobiology Lab of the Salk Institute in La Jolla, California from 1997-1998. Since 2005, he has been an Adjunct Fellow of the Frankfurt Institute of Advanced Studies (FIAS).
Christoph Stoeckl is a Postdoc researcher at Technische Universität Graz working in the intersection between computational neuroscience and AI. His research interests include neuromorphic hardware as well as exploring connections between Transformers and neural networks. Before joining the research lab of Prof. Maass, he obtained a Master’s degree in Computer Science also at TU Graz.
Yukun Yang is a 1st-year Doctoral Student at Technische Universität Graz, supervised by Prof. Wolfgang Maass. His primary research interest is at the intersection of AI and neuroscience, with a focus on discovering the learning principles of the brain and its neuromorphic applications. Before joining TU Graz, he earned M.S. in the ECE Department at Duke University in 2020. Earlier, he received B.E. in Information Engineering from Xi’an Jiaotong University in 2018.

For the meeting link, see the full INRC Forum Spring 2023 Schedule (accessible only to INRC Affiliates and Fully Engaged Members).