Technical Program

Paper Detail

Paper: PS-1A.6
Session: Poster Session 1A
Location: Symphony/Overture
Session Time: Thursday, September 6, 16:30 - 18:30
Presentation Time:Thursday, September 6, 16:30 - 18:30
Presentation: Poster
Publication: 2018 Conference on Cognitive Computational Neuroscience, 5-8 September 2018, Philadelphia, Pennsylvania
Paper Title: The many directions of feedback alignment
Manuscript:  Click here to view manuscript
Authors: Brian Cheung, Daniel Jiang, UC Berkeley, United States
Abstract: Backpropagation is a key component in training neural network models which have been successfully applied to many perceptual tasks including vision and audio. Despite this success, an analogous learning mechanism has yet to be discovered in biology. The feedback alignment algorithm relaxes the constraints imposed by the backpropagation procedure while still demonstrating successful learning in feedforward neural networks. In this work, we further loosen the constraints of these learning algorithms by removing the directionality constraint of the forward and backward paths. We show these paths can be operated in a bidirectional manner to train multiple networks without interference even when the networks are solving different tasks.