How to compute softmax and its gradient?
17 views (last 30 days)
I am creating a simple two layer neural network where the activation function of the output layer will be softmax.
I have this for creating softmax in a numerically stable way
function g = softmax(z)
dim = 1;
s = ones(1, ndims(z));
s(dim) = size(z, dim);
maxz = max(z, , dim);
expz = exp(z-repmat(maxz, s));
g = expz ./ repmat(sum(expz, dim), s);
z is a matrix that contains all of the data calculated by the previous layer one row at a time.
In order to compute the derivative of this though I will need to use the Kronecker delta but I am not sure how to do it.
Can someone provide me with a vectorized implementation for computing it in Matlab?