Weight Matrix Headcount

You train a linear softmax classifier on d-dimensional input features for k classes, giving every class its own weight vector and bias term. How many trainable parameters does the model have in total?

(k - 1)(d + 1), because softmax models are always fit with one class fixed as a reference

k(d + 1)

d(k + 1)

kd, since biases are not needed for multiclass models