Backprop from Scratch

The backward pass, one layer at a time. Activation gradients, linear and embedding layer gradients, softmax cross-entropy, and a numeric gradient check to keep everything honest.

12 problems