Spiking Neural Networks for Predictive Coding with a Feedforward Gist Pathway
Kwangjun Lee, Jorge Mejias, Cyriel Pennartz, University of Amsterdam, Netherlands; Shirin Dora, Loughborough University, United Kingdom; Sander Bohte, Centrum Wiskunde & Informatica, Netherlands
Posters 1 Poster
Pacific Ballroom H-O
Thu, 25 Aug, 19:30 - 21:30 Pacific Time (UTC -8)
How does the brain seamlessly perceive the world without having a direct access to the source of sensations in the physical world? Rather than passively relaying information that sensory organs pick up from the external world, it actively gathers statistical regularities from sensory inputs to represent (i.e., predict) the subject's current situation. Predictive coding (PC) offers a biologically plausible scheme for such a generative model via hierarchical prediction error minimization. To advance the biological realism of previous PC models based on artificial neural networks, we developed a spiking neural network for predictive coding (SNN-PC), which introduces spiking neurons and proposes structural and algorithmic modifications required to overcome challenges from spiking dynamics. In addition, we present the feedforward gist pathway, which models a neural code for the gist of an image and provides a neurobiological alternative to arbitrary choices of a prior. After training with simple visual images, SNN-PC developed hierarchical latent representations and reconstructed input images. SNN-PC suggests biologically plausible mechanisms by which the brain performs perceptual inference and learning in an unsupervised manner.