Phillip Isola teaches this session of MIT's 6.7960 Deep Learning, covering reconstruction-based approaches to representation learning. He works through autoencoders, clustering, and vector quantization as methods for compressing data into useful internal representations, then connects these to self-supervised learning setups that train on reconstruction losses rather than labels. Isola also discusses how learned representations can be analyzed and compared, drawing parallels between representations that emerge in artificial networks and those observed in biological brains. At 81 minutes, the lecture builds on prior sessions in the course and assumes familiarity with basic neural network training, making it best suited to students already following the 6.7960 sequence rather than complete newcomers to deep learning.