ASTRONOMY COLLOQUIUM

Artificial Intelligence is transforming science at many levels. Neural networks have changed the ways we interact with data and how we think about statistics. What is fundamentally different? What can we expect to learn from them? Do they have to be opaque? This talk will demystify the "black box" of neural networks by examining their fundamental learning mechanisms and internal representations. I will show how to reveal their encoded features and how to compare networks trained on different datasets. I will then discuss how to use them for astrophysical explorations while maintaining the level of control of more traditional approaches.