Combining various types of spike timing-dependent plasticity to learn efficient neural codes
Authors: Antony W. N’dri, Céline Teulière and Jochen Triesch
Presentation type: Short talk at SNUFA 2023 online workshop (7-8 Nov 2023)
Abstract
Biological spiking neural networks learn efficient representations of sensory inputs with local learning rules. Mimicking such abilities in brain-inspired computing approaches is still a grand challenge. Here we present a learning framework for hierarchical spiking neural networks called predictive coding light (PCL). PCL uses both excitatory and inhibitory spike timing-dependent plasticity (STDP) rules in a recurrent hierarchical network architecture to learn an efficient redundancy-reducing representation of sensory inputs. When exposed to naturalistic visual input, our PCL network learns simple and complex cell-like receptive fields and exhibits surround suppression effects as observed biologically. We systematically compare different STDP-rules (causal, anti-causal, symmetric) for different connection types within the PCL network. Our results indicate that the variety of biologically observed STDP rules enhances the learning abilities of hierarchical spiking neural networks.