Recent Advances in Resonator Networks for Neurosymbolic Computing

Alpha Renner, Christopher Kymn, Edward Paxon Frady, Friedrich Sommer

Keywords: Resonator, Vector Symbolic Architectures, Hyperdimensional Computing, distributed representations, compositional reasoning, neuromorphic algorithms, visual odometry, cognitive maps

TL;DR: Resonator Networks decompose structured representations into their constituent parts offering a promising path toward compositional reasoning in neurosymbolic AI systems

Abstract:

As an alternative to hybrid neuro-symbolic approaches for combining the strengths of Neural Networks and symbolic AI, Vector Symbolic Architectures (VSAs) provide a seamless framework for robust parallel computing with transparency. A product-like dyadic vector operation in VSAs enables variable binding and the encoding of data structures by decomposable distributed representations. One critical, long-standing problem in VSA-style neuro-symbolic computing was how to decompose a representation of bound variables. Given the encoding schemes of the variables, pattern matching and unbinding involves vector factorization, that is, searching large combinatorial spaces of vector products. By interleaving binding operations with attractor network dynamics, Resonator Networks can efficiently solve this problem.
Recent advances demonstrate the versatility and potential of Resonator Networks. Here, we present four key developments published in the past year: a hierarchical resonator network handling non-commutative transformations in composed visual scenes, a neuromorphic hardware implementation, applications to visual odometry, and insights into cognitive map formation in hippocampal-entorhinal circuits.
These studies offer theoretical insights and practical applications for neuro-symbolic parallel computing. In the future, incorporating learning mechanisms for the underlying generative models in Resonator Networks offers a promising path toward compositional reasoning in neuro-symbolic AI systems.

Track: Main Track

Paper Type: Extended Abstract

Resubmission: No

Software: https://doi.org/10.24433/CO.1543398.v1

Submission Number: 72