Phillip Isola teaches this session of MIT's 6.7960 Deep Learning course, covering the point where representation learning and generative modeling intersect. The lecture centers on variational autoencoders and the role of latent variables in learning compressed representations of data that also support generation of new samples. Isola works through the mathematical framing of VAEs, including how encoding data into a latent space and decoding back out lets a network learn structure without explicit labels. The session runs 81 minutes and fits within the broader deep learning curriculum, building on earlier lectures about representation learning before extending into generative approaches. Expect blackboard or slide-based derivation rather than live coding, aimed at students who already have a grounding in neural network basics and are ready to see how generative modeling reframes what a learned representation is for.