To add, there is a whole field of biologically-plausible learning. Predictive coding is one such method. It is isomorphic to SGD[1], and doesn't require global computation. ie it's backprop-free optimization. Although slower than SGD on silicon, it's massively parallelizible which ends up making it more biologically plausible rather than less.
Comments
To add, there is a whole field of biologically-plausible learning. Predictive coding is one such method. It is isomorphic to SGD[1], and doesn't require global computation. ie it's backprop-free optimization. Although slower than SGD on silicon, it's massively parallelizible which ends up making it more biologically plausible rather than less.
1. 2020 PREDICTIVE CODING APPROXIMATES BACKPROP ALONG ARBITRARY COMPUTATION GRAPHS https://arxiv.org/pdf/2006.04182.pdf