abhimanyu08 / blog Goto Github PK

Hi Abhimanyu,

thanks for the great explanation of label smoothing. I found a little typo. At the end of section 1 you write:

Now suppose our model didn't output ${p_{1}}$ but ${p_{2}=[0.01,0.79,0.15,0.5]}$. In this case it is less sure that the image has label 2. Loss will be ${L(y^{l},p_{2})= -(0.05\log0.01+0.85\log0.79+0.05\log0.15+0.05\log0.5)=0.56}$ which is less than the loss with ${p_{1}}$ ! This goes on to show that smooth labels want the model to be confident about it's predictions but not over-confident.

The problem with p_2 is that the entries don't sum to 1. The last probability should likely be 0.05 instead of 0.5. In this case the resulting loss is 0.68. This in turn is slightly more than the loss for p_1 (0.61) which matches my expectation.

Thanks again for a great tutorial!