Tracing Model Behavior Back to Training Data

In this talk, I will share my research on how artificial neural networks learn from data. I will discuss my past work on developing algorithms to identify which training examples influence particular model behaviors, evaluating these algorithms with controlled benchmarks, and extending them to scalable approaches for larger models. I will briefly close by describing how this work motivates my future research on recovering reusable knowledge units that a model learns from training data and studying how these units develop and depend on one another during learning.