PhD defense Madison Cotteret: Vector-symbolic attractor networks

Neuromorphic computing aims to match the efficiency and adaptability of human intelligence by emulating the computational principles of biological cognitive function. In doing so, it promises to overcome the limitations of conventional AI systems which, despite impressive performance, often lack generalisable capabilities and require vast computational resources. More faithfully emulating biological processing elements has yielded success in small-scale sensory tasks, but scaling up this approach to more difficult tasks has proven challenging. As the size and demands of neuromorphic systems grow, issues of reliability and programming complexity emerge, prohibiting the development of extended neuromorphic systems for real-world application.

In his thesis, Madison Cotteret addresses these challenges by combining two biologically-inspired areas of study: Vector-Symbolic Architectures (VSAs) and attractor networks. VSAs are a systematic framework for representing data structures with high-dimensional vectors, such that the encoded information is distributed equally across all neurons. Attractor networks endow neuromorphic and biological systems with emergent stability and have found widespread use as state-holding autoassociative memories, but their potential for broader computational use remains underexplored. Both paradigms are particularly appealing for neuromorphic hardware, where scale and parallelism are cheap, but the reliability of individual components is far from assured.

The combination of VSAs and attractor networks enables a scalable approach to directly map symbolic data structures and computational primitives onto large unreliable networks. By leveraging the insensitivity to component failure of attractor networks, this approach achieves a previously unattainable degree of simultaneous flexibility and robustness in neuromorphic systems. This is a necessary advancement for neuromorphic systems to mature into more general and reliable computing platforms, enabling the development of low-power neuromorphic hardware capable of performing extended cognitive tasks.

Dissertation: https://hdl.handle.net/(...)97-a122-0d1ca772a9d2