Direct Feedback Alignment Scales to Modern Deep Learning Tasks and Architectures

Despite being the workhorse of deep learning, the backpropagation algorithm is no panacea. It enforces sequential layer updates, thus preventing efficient paral-lelization of the training process.