IPAB Workshop - 03/11/2022

Title: Locality-sensitive hashing using an insect-inspired architecture

Abstract: The Mushroom Body (MB), a center of learning and memory in the insect brain, encodes sensory signals in a form that is hypothesised to be locality-sensitive, i.e. it preserves the similarity of the inputs. This is an important step that precedes reinforcement, allowing insects to generalize their behavior to input patterns similar to those that they previously learned. Along with preserving similarity, the MB has characteristics that make it an attractive model to study for Locality-Sensitive Hashing (LSH): it forms sparse and random connections with the previous layer and exhibits a sparse code. Due to these characteristics, models inspired by the MB can be shown to be less computationally expensive in terms of training, inference, and search speed. In this talk, I will compare different MB-inspired architectures tested on a nearest neighbours task and explore how they can be used for visual place recognition.