Week 8

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

Lecture part A

In this section, we focused on the introduction of contrastive methods in Energy-Based Models in several aspects. First, we discuss the advantage brought by applying contrastive methods in self-supervised learning. Second, we discussed the architecture of denoising autoencoders and their weakness in image reconstruction tasks. We also talked about other contrastive methods, like contrastive divergence and persistent contrastive divergence.

Lecture part B

In this section, we discussed regularized latent variable EBMs in detail covering concepts of conditional and unconditional versions of these models. We then discussed the algorithms of ISTA, FISTA and LISTA and look at examples of sparse coding and filters learned from convolutional sparse encoders. Finally we talked about Variational Auto-Encoders and the underlying concepts involved.

Practicum

In this section, we discussed a specific type of generative model called Variational Autoencoders and compared their functionalities and advantages over Classic Autoencoders. We explored the objective function of VAE in detail, understanding how it enforced some structure in the latent space. Finally, we implemented and trained a VAE on the MNIST dataset and used it to generate new samples.


📝