Reviews: Assessing the Scalability of Biologically-Motivated Deep Learning Algorithms and Architectures

The authors provide a clear and succinct introduction to the problems and approaches of biologically plausible forms of backprop in the brain. They argue for behavioural realism apart from physiological realism and undertake a detailed comparison of backprop versus difference target prop and its variants (some of which they newly propose) and also direct feedback alignment. In the end though, they find that all proposed forms of bio-plausible alternatives to backprop fall quite short on complex image recognition tasks. Despite the negative results, I find such a comparison very timely to consolidate results and push the community to search for better and more diverse alternatives. Overall I find the work impressive. The authors claim that weight sharing is not plausible in the brain.