061: Interpolation, Extrapolation and Linearisation (Prof. Yann LeCun, Dr. Randall Balestriero)
Machine Learning Street Talk (MLST)
Understanding Neural Networks through Geometry
This chapter investigates how neural networks represent input data as compositions of linear functions within polyhedral structures, highlighting the relationship between learning decision boundaries and latent spaces. It critiques the notion that neural networks inherently eliminate the need for feature engineering while exploring the complexities of optimization and generalization in high-dimensional spaces. The discussion emphasizes the significance of piecewise linear functions and their role in modeling, questioning traditional intuitions about smoothness in natural phenomena.
00:00
Transcript
Play full episode
Remember Everything You Learn from Podcasts
Save insights instantly, chat with episodes, and build lasting knowledge - all powered by AI.
