Position: PhD Student
Current Institution: Cornell University and Google Brain
Abstract: Insights from Deep Representations with Healthcare Applications
To continue the successes of deep learning, it becomes increasingly important to better understand the phenomena exhibited by these models, ideally through a combination of systematic experiments and theory. Central to this challenge is a better understanding of deep representations. I overview adapting Canonical Correlation Analysis (SVCCA) as a tool to directly compare latent representations, across layers, training steps, and even different networks. The results highlight the difference between signal and noise stability in the representations and show us how we can representationally compress networks. We also gain insights into per-layer and per-sequence step convergence, along with differences between generalizing and memorizing networks. I then overview some of the applications of this style of analysis in helping us design better, more tractable tasks for uncertainty estimation with noisy healthcare data. Finally, I highlight some applications in the context of reinforcement learning — a new testbed that lets us us study different RL algorithms, single agent, multiagent, and self-play settings and evaluate generalization in a systematic way.
Maithra Raghu is a PhD student in computer science at Cornell University, working with Jon Kleinberg. She also collaborates extensively with Google Brain, particularly with Samy Bengio. She is interested in developing a science of deep learning and deep representations and applying these insights to healthcare.