WYX同学的Blog

Q1:What happens to loss if car scores decrease by 0.5 for this training example?

A1:The score of the car is larger than the others.So if the score of the the car change a little bit,the margin of one will still be retained and the loss will not change.

Q2:What is the min/max possible SVM loss $L_i$?

A2: The minimun loss is zero and the maximum is infinity.

Q3:At initialization W is small so all $s \approx 0$.What is the loss $L_i$,assuming N examples and C classes?

A3:We loop over all of the incorrect classes,so it is C-1 classes.And the S is about the same,so the loss $L_i$ will be C-1.

Q4:What if the sum was over all classes?(including $j = y_i$)

A4:The result is the loss $L_i+1$,as we also loop over the correct class.

Q5:What if we used mean instead of sum?

A5:It will rescale the whole loss function by a constant,but it doesn’t matter,it won’t have any effect on the classification.

Q6:What if we used $L_i = \sum_{j \neq y_i}max(0,s_j-s_{y_i}+1)^2$

A6:This magnifies an error we encounter, and some of the nuances we encounter are even more subtle because all the results are shown as squares.