Week 4: The Hidden Geometry of Data & Self-Supervised Learning

Stop assuming your data is perfect. Real world datasets are corrupted and neural networks fail spectacularly if initialized blindly. You will uncover the "dark matter" of Machine Learning by training on massive unlabeled datasets using Self-Supervised Learning. We bend your mind with high-dimensional geometry, unpack the architecture behind Apple's FaceID, and reveal how Kaiming initialization prevents deep networks from collapsing. Core Topics: Cross-Validation & Hyperparameter Tuning, Receptive Fields, Weight Initialization (Xavier/Glorot & Kaiming/He), The Manifold Hypothesis & High-Dimensional Geometry (The "Sea Urchin" Effect), Metric Learning (Siamese Networks, Triplet Loss, Hard Negative Mining), and the Holy Trinity of Self-Supervised Learning (SimCLR, CLIP, and Masked Autoencoders). Week 4 Lab: Real world data is garbage. We hand you a corrupted dataset so you can let your model's own mistakes flag the bad images. Purge the fakes in Voxel51, augment what is left, and crush your original score.

8 weeks · 59 lectures · free to watch