Learning Dynamics, Task Technology and Information Plasticity for Weakly- Supervised Representation Learning

Abstract

This project is intended to explore (Task 1) defining a topology in the space of tasks, and computing ""static"" distances, or diverg"ence measures, between tasks, which help determine what resources (time and capacity) are needed to transfer a trained model to a f"ine-tuned task; (Task 2) studying the role of the transient dynamics of learning, and define a ""dynamic"" distance, or discrepancy ," between tasks, to enable analyzing critical learning periods in biological and artificial systems. and (Task 3) exploit information plasticity to overcome critical learning period and facilitate learning with few or no supervision (semi supervised learning). We have discovered that artificial systems exhibit critical learning periods, much like biological systems, which prompts to the critical role of the transient dynamics in learning, whereas almost the entirety of the analysis and theoretical results in machine learning and optimization are asymptotic. Since the transient is, literally, critical, we believe it is urgent to investigate the fundamental issues underlying critical periods, and we have preliminary evidence that such understanding can be leveraged for improving semi-supervised and transfer learning algorithms. The fact that critical period phenomena involve information plasticity, rather than (biological) neural plasticity, suggests that we can conduct Artificial Neuroscience experiments, far simpler and cheaper than psychophysical experiments, before validating the observations in biology.

Document Details

Document Type
DoD Grant Award
Publication Date
Apr 25, 2019
Source ID
N000141912229

Entities

People

  • Stefano Soatto

Organizations

  • Office of Naval Research
  • United States Navy
  • University of California, Los Angeles

Tags

Readers

  • Control Systems Engineering.
  • Neural Network Machine Learning.
  • Systems Analysis and Design

Technology Areas

  • AI & ML
  • AI & ML - Neural Networks
  • Space