Week 11

$$\gdef \sam #1 {\mathrm{softargmax}(#1)}$$ $$\gdef \vect #1 {\boldsymbol{#1}} $$ $$\gdef \matr #1 {\boldsymbol{#1}} $$ $$\gdef \E {\mathbb{E}} $$ $$\gdef \V {\mathbb{V}} $$ $$\gdef \R {\mathbb{R}} $$ $$\gdef \N {\mathbb{N}} $$ $$\gdef \relu #1 {\texttt{ReLU}(#1)} $$ $$\gdef \D {\,\mathrm{d}} $$ $$\gdef \deriv #1 #2 {\frac{\D #1}{\D #2}}$$ $$\gdef \pd #1 #2 {\frac{\partial #1}{\partial #2}}$$ $$\gdef \set #1 {\left\lbrace #1 \right\rbrace} $$

Lecture part A

In this section, we discussed about the common activation functions in Pytorch. In particular, we compared activations with kink(s) versus smooth activations - the former is preferred in a deep neural network as the latter might suffer with gradient vanishing problem. We then learned about the common loss functions in Pytorch.

Lecture part B

In this section, we continued to learn about loss functions - in particular, margin-based losses and their applications. We then discussed how to design a good loss function for EBMs as well as examples of well-known EBM loss functions. We gave particular attention to margin-based loss function here, as well as explaining the idea of “most offending incorrect answer.


This practicum proposed effective policy learning for driving in dense traffic. We trained multiple policies by unrolling a learned model of the real world dynamics by optimizing different cost functions. The idea is to minimize the uncertainty in the model’s prediction by introducing a cost term that represents the model’s divergence from the states it is trained on.