Is backpropagation biologically plausible?

Abstract

The author searches for neurobiologically plausible implementations of the backpropagation gradient descent algorithm. Any such implementation must be consistent with physical constraints such as locality (i.e., that the behavior of any component can be influenced solely by components in physical contact with it) and contingent facts of biology, and must also preserve global network properties such as fault tolerance, stability, and graceful degradation to hardware errors. The authors finds that in several posited implementations these design considerations imply that a finely structured neural connectivity is needed as well as a number of neurons and synapses beyond those inferred from the algorithmic network presentations of backpropagation. Gating synapses (Sigma-Pi units) are present while Hebbian (or pseudo-Hebbian) synapses are absent from all his posited implementations. Although backpropagation can in principle be implemented in neurobiology, such high network structure and the organizational principles required for its generation at the level of individual neurons will require more support from experimental neurobiology.<>