Beyond Lazy Training for Over-parameterized Tensor Decomposition
Offered By: Fields Institute via YouTube
Course Description
Overview
Explore tensor decomposition and over-parameterization in this 37-minute conference talk from the Fields Institute's Mini-symposium on Low-Rank Models and Applications. Delve into the comparison between lazy training regimes and gradient descent techniques for finding approximate tensors. Examine the challenges of analyzing gradient descent, the failures of lazy training, and the existence of local minima. Learn about a novel algorithm that escapes local minima through random correlation and amplifies initial correlation using tensor power methods. Gain insights into the importance of over-parameterization in training neural networks and its implications for avoiding bad local optimal solutions.
Syllabus
Intro
Low rank models and implicit regularizati
Regimes of over-parametrization
Tensor (CP) decomposition
Why naïve algorithm fails
Why gradient descent?
Two-Layer Neural Network
Form of the objective
Difficulties of analyzing gradient descent
Lazy training fails
O is a high order saddle point
There are local minima away from 0
Our (high level) algorithm
Proof ideas
Escaping local minima by random correla
Amplify initial correlation by tensor power man
Conclusions and Open Problems
Taught by
Fields Institute
Related Courses
Applied Linear Algebra in AI and MLIndian Institute of Technology, Kharagpur via Swayam AlphaTensor by DeepMind Explained
Yannic Kilcher via YouTube Constraining 3D Fields for Reconstruction and View Synthesis
Andreas Geiger via YouTube Neural Nets for NLP 2018 - Learning from-for Knowledge Graphs
Graham Neubig via YouTube Smoothed Analysis in Unsupervised Learning via Decoupling
IEEE via YouTube