Abstract:
The relevant limit for machine learning is not N → infinity but instead N → 0. The human visual system is proof that it is possible to learn categories with extremely few samples. This talk will discuss steps towards building such systems and it is structured in three parts. The first part will discuss algorithms to adapt representations of deep networks to new categories with few labeled data. The second part will discuss when such adaptation works well and when it does not. It will develop a method to compute the optimal distance between two learning tasks and algorithmic tools to learn tasks that are far away from each other. The third part will discuss how make the optimal use of unlabeled data to learn a task.

This talk will discuss results from the following papers.
1. An Information-Geometric Distance on the Space of Tasks. Yansong
Gao, Pratik Chaudhari. ICML 2021.
Paper: https://arxiv.org/abs/2011.00613, Code: https://github.com/Yansongga/An-Information-Geometric-Distance-on-the-Space-of-Tasks
2. Model Zoo: A Growing “Brain” That Learns Continually. Rahul
Ramesh, Pratik Chaudhari. ICLR 22.
Paper: https://arxiv.org/abs/2106.03027. Code:
https://github.com/rahul13ramesh/MultitTask_ModelZoo
3. Deep Reference Priors: What is the best way to pretrain a model?. Yansong Gao, Rahul Ramesh, and Pratik Chaudhari. ICML 22.
Paper: https://arxiv.org/abs/2202.00187, Code: https://github.com/grasp-lyrl/deep_reference_priors