How to utilize physics to enhance artificial intelligence
Bahram Jalali, Achuta Kadambi, Vwani Roychowdhuri · 2021
Deep neural networks exploit millions or more free parameters that are tuned to a requisite large and curated dataset. The black-box nature of these models masks interpretability and the ability to diagnose failures. Although astonishing performance gains are being achieved, these come at the expense of exponential rise in computation and memory utilization. This talk will review how the emerging convergence of physics and neural networks will confront these challenges, extend the rise of artificial intelligence, and create opportunities for scientific discoveries.