Replacing the count matrix with a trainable neural network — one-hot encoding, a weight matrix, softmax, and gradient descent.
This lesson requires an active subscription.