Self-Supervised Learning for Spiking Neural Networks: Predictive Coding and Forward-Only Learning

Open

https://www.rit.edu/directory/agovcs-alexander-ororbia

Alexander Ororbia

• Director of the Neural Adaptive Computing Laboratory at the Rochester Institute of Technology.

• Assistant Professor in Computer Science

• Affiliate Professor in Psychology

• Affiliate faculty in RIT's Center for Applied Neuroscience

Friday, July 21st, 2023
Start: 10:00 PT / 1:00 ET / 19:00 CET
Duration: 30 minutes

In this talk, I will discuss self-supervised credit assignment for spiking neural systems, an important direction for research in brain-inspired computing that has emerged from my lab, the Neural Adaptive Computing laboratory, over the last several years. Specifically, this talk will cover two key computational frameworks - spiking predictive coding and contrastive signal dependent plasticity. With respect to the first, I will present how the underlying local cellular dynamics of predictive coding can be generalized to naturally work with arbitrary spike-response functions and present results for an adaptive multi-layer generative model composed of leaky integrate-and-fire neurons using parallel, iteratively learned feedback. For the second, more recent framework, contrastive signal dependent plasticity (CSDP), I will present how a parallel recurrent spiking network, which embodies many of the favorable structural properties of spiking predictive coding, such as top-down and bottom-up neural signal integration, can be learned online without feedback synapses, representing the first spike-level generalization of the forward-forward and predictive forward-forward algorithms. Finally, I will close with issues, challenges, and future directions for event-driven self-supervised learning.

Not Yet Available