Special Maths Colloquium: Meta Learning: Predicting Physical and Biological Systems with Little Data

Speaker: Prof. Ying-Cheng Lai

Affiliation: Arizona State University

Abstract

Modern artificial intelligence has the power to revolutionize scientific discovery, but conventional machine learning usually requires massive amounts of data, presenting a major obstacle in fields such as ecology and quantum chaos where observational data can be exceptionally scarce. To address this challenge, the speaker will present a “meta-learning” approach to demonstrate how the data gap can be bridged by pre-training machine-learning models on abundant, unrelated datasets before applying it to situations where only small datasets are available. In ecology, synthetic data from general mathematical models is utilized to train reservoir computers to predict biological population evolution, achieving high accuracy with five to seven times less real-world data than conventional reservoir computing requires. Similarly, in physics, deep convolutional neural networks pre-trained on standard image datasets can successfully detect elusive quantum scars, rare wave patterns in quantum systems exhibiting chaos in the classical limit, from thousands of images without any human intervention. The successful application to both ecological forecasting and quantum state detection highlights meta-learning as a versatile tool for solving challenging, data-limited problems in different fields.

Bio

Dr. Lai is a Regents Professor, the ISS Endowed Professor of Electrical Engineering, and Professor of Physics at ASU. He is a Vannevar Bush Faculty Fellow (U.S. Department of Defense), a Fellow of the American Physical Society and the American Association for the Advancement of Science, a Corresponding Fellow of the Royal Society of Edinburgh (National Academy of Science and Letters of Scotland), and a foreign member of Academia Europaea. His research interests include machine learning for nonlinear and complex dynamical systems, computational biology, and theoretical ecology.