Backpropagation and the brain

Abstract

During learning the brain modifies synapses to improve behaviour. In the cortex synapses are embedded within multi-layered networks, making it difficult to determine the effect of an individual synaptic modification on the behaviour of the system. The backpropagation algorithm solves this problem in deep artificial neural networks, but has historically been viewed as biologically problematic. Nonetheless, recent developments in neuroscience and the successes of artificial neural networks have reinvigorated interest in whether backpropagation offers insights for understanding learning in the cortex. The backpropagation algorithm learns quickly by computing synaptic updates using feedback connections to deliver error signals. While feedback connections are ubiquitous in the cortex, it is difficult to see how they could deliver the error signals required by strict formulations of backpropagation. Here we build on past and recent developments to argue that feedback connections may instead induce neural activities whose differences can be used to locally approximate these signals, and hence drive effective learning in deep networks in the brain.

Explaining face representation in the primate brain using different computational models

“…These results suggest that the face subspace portion of the representation learned by CORnet-Z may be interpreted in much simpler terms, as a shape appearance model. The results provide an important counter-example to the increasingly popular view that only distributed representations learned by multi-layer networks can well explain IT activity (Kietzmann et al, 2018;Lillicrap et al, 2020). Why a network trained on object classification should learn an approximation to a generative model of faces is an interesting question for future research.…”

Integrated World Modeling Theory (IWMT) Revisited

“…This information flow may also be viewed as Bayesian belief propagation or (marginal) message passing (Friston et al, 2017c;Parr et al, 2019b). In contrast to variational autoencoders in which training proceeds via backpropagation with separable forward and backward passes-where cost functions both minimize reconstruction loss and deviations between posterior latent distributions and priors (usually taking the form of a unit Gaussian)-training is suggested to occur (largely) continuously in predictive processing (via folded autoencoders), similarly to recent proposals of target propagation (Hinton, 2017;Lillicrap et al, 2020). Note: Folded autoencoders could potentially be elaborated to include attention mechanisms, wherein higher-level nodes may increase the information gain on ascending prediction-errors, corresponding to precision-weighting (i.e., inverse variance over implicit Bayesian beliefs) for selected feature representations.…”

An Integrated World Modeling Theory (IWMT) of consciousness: Combining Integrated Information and Global Neuronal Workspace Theories with the Free Energy Principle and Active Inference Framework; towards solving the Hard problem and characterizing agentic causation

“…This information flow may also be viewed as Bayesian belief propagation or (marginal) message passing (Friston et al, 2017;Parr et al, 2019). In contrast to variational autoencoders in which training proceeds via backpropagation with separable forward and backward passes-where cost functions both minimize reconstruction loss and deviations between posterior latent distributions and priors (usually taking the form of a unit Gaussian)-training is suggested to occur (largely) continuously in predictive processing (via folded autoencoders), similarly to recent proposals of target propagation (Hinton, 2017;Lillicrap et al, 2020). Note: Folded autoencoders could potentially be elaborated to include attention mechanisms, wherein higher-level nodes may increase the information gain on ascending prediction-errors, corresponding to precision-weighting (i.e., inverse variance over implicit Bayesian beliefs) over selected feature representations.…”