The Deep Learning Channel

In this chapter, we continue to study the local learning principle and the nature of the information I i that must be sent over the DLC. At the same time, we tackle the problem that backpropagation requires perfectly symmetric weights in the forward and backward pass, something that is very easy to implement in digitally simulated neural networks, but very hard to implement in biological and other physical neural networks. This question is part of a broader set of concerns regarding backpropagation in relation to learning in the machine, and in particular to learning in biological networks. These include: