
CS330: Deep Multi-Task and Meta Learning - What is Multi-Task Learning? (Lecture 1)
Chelsea Finn opens Stanford's CS330 by laying out why multi-task and meta learning matter once a single model needs to handle many problems rather than one. She defines a task formally in terms of a data distribution and loss function, then distinguishes multi-task learning from transfer learning and meta learning, showing how each handles sharing information across related problems differently. The lecture covers conditioning strategies for multi-task architectures, how task identity gets fed into a network, and basic optimization approaches like gradient surgery for cases where tasks interfere with each other. Finn grounds the discussion in PyTorch implementation details students will use in course assignments, and previews the quarter's arc from supervised multi-task setups toward meta learning and reinforcement learning. As the first lecture in the course, it functions as orientation and foundation, setting vocabulary and motivating examples that later sessions build on.