Learning One-Hidden-Layer Neural Networks with Landscape Design

Explore the intricacies of learning one-hidden-layer neural networks through landscape design in this 32-minute conference talk by Tengyu Ma from Stanford University. Delve into the challenges of optimization in machine learning and discover a new objective for training neural networks. Examine why straightforward objectives fail and learn about an innovative analytic formula. Investigate provable non-convex optimization algorithms and gain insights into potential paradigms for optimization theory in machine learning. This Simons Institute presentation, part of the "Optimization, Statistics and Uncertainty" series, offers a deep dive into the interfaces between users and optimizers, providing valuable knowledge for researchers and practitioners in the field of neural networks and machine learning optimization.

Intro
Interfaces Between Users and Optimizers?
Optimization in Machine Learning: New Interfaces?
Possible Paradigm for Optimization Theory in ML?
This Talk: New Objective for Learning One-hidden-layer Neural Networks
The Straightforward Objective Fails
An Analytic Formula
Provable Non-convex Optimization Algorithms?
Conclusion