Closed CodexForster closed 4 years ago
The calculation is implicitly including the bias term, so the weight matrix is of shape (10, 3073)
for a total of 30,730 parameters. Refer to the bias trick section here.
Ok, that makes sense. My bad.
I hope you haven't missed the comment on the typo.
Under Module 1 of the "Optimisation: Stochastic Gradient Descent" chapter, there are multiple instances where it is mentioned that W has 30730 parameters. But according to the previous chapter and the fact that they are 32x32 pixel images in RGB scheme to be classified in 10 classes (which means 32323*10=30720) it should be 30720 instead of 30730. There are multiple instances where this mistake is made. Please look into them. Under section "Mini-batch gradient descent" of the Optimisation chapter in Module 1, there is an incorrect sentence: "the training data can have on order of millions of examples" instead of "the training data can have an order of millions of examples".