Development of spiking architectures using Bayesian deep learning

Objectives

1. Identifying tractable network architectures to capture spatio-temporal patterns in spiking and non-spiking data
2. Investigation of computational methods to solve supervised and unsupervised learning tasks
3. Investigation of scalable learning method for online learning
4. Implementation of the methods above on dedicated hardware

Description

Spiking neural networks (SNNs) can offer increased processing speed and reduced power consumption, especially when implemented on dedicated hardware (neuromorphic chips or FPGAs). Standard SNNs are typically fed with spiking data (e.g., streams of binary values), and the output of each layer remains spiking. They are particularly well suited for several new sensors, such as event-cameras, or single-photon detectors (used for single-photon Lidar), that natively produce event-like data compatible with spiking networks. While data from current sensors can be manually converted into events for fast processing, it is also possible to develop hybrid structures where some layers (such as the input and/or output layer) are traditional (continuous) layers.

Bayesian methods have recently been proposed to train SNNs, either to train deterministic SNNs (with deterministic activation functions) or probabilistic SNNs. They can also handle quantized weights using the reparameterization trick. Still, while methodological tools (first principles) exist, their application to solve real world problems still need to be demonstrated. This is what we intend to achieve in this project.