Research

Little strokes fell great oaks.

— Benjamin Franklin

My research has wandered from machine learning applications in neuroscience toward the theory and methodology of deep learning. My current projects also ask about fairness and AI oversight. Across these areas, I’m interested in how learning systems pick up useful patterns, and when those patterns carry over to new people, data, or tasks.

See my publication list for papers, preprints, presentations, and datasets, with filters by topic and type.

At Penn, I worked with electroencephalography (EEG)—electrical recordings of brain activity—and used machine learning to ask whether brain signals could help predict behavior, like whether someone would remember an item. We’re still a ways off from mind-reading, but it’s pretty cool stuff. These days I’m interested in why deep learning works so well, where it runs into trouble, and whether a better theoretical understanding can help us build models that are more reliable and interpretable.

Theory of deep learning

Right now I’m especially interested in how training procedures shape the solutions neural networks learn, and what those solutions can tell us about generalization.

Inductive bias / implicit regularization

Paper: Estimating Implicit Regularization in Deep Learning · Blog post

One possible explanation for why neural networks generalize well is that they have some kind of inductive bias that encourages them to learn generalizing solutions. I’m interested in using shifts away from the original loss minimum to reverse-engineer the inductive biases that training introduces.

We use gradients of the loss and a proposed regularizer to estimate its unknown parameters.

In the animation, the learned weights stay fixed while the regularization strength changes. The red curve is the empirical loss, the dashed blue curve is the regularizer, and the gray curve is their sum. At a minimum of the combined objective, the loss and regularizer gradients cancel.

Matching loss and regularizer gradients at the learned weights.

Activity descent and learning rules

Paper: Weight Space Gradient Descent Induces Approximate Activity Descent, Making Learning Rules Testable

A central question in neuroscience is whether biological neural networks optimise a loss function by adjusting synaptic weights—and if so, by what rule. Candidate rules are difficult to test because synaptic weights cannot be measured at scale in behaving animals, whereas neural activities can be recorded across learning.

We show that weight-space gradient descent induces kernel-filtered activity descent; when the kernel is approximately diagonal, each neuron follows its own loss gradient. We validate this prediction in artificial neural networks. The same relationship can and should be tested in biological neural networks with existing experimental techniques.

AI safety

Fairness in the training data

Paper: Targeted Learning for Inference on Data Fairness

This project uses targeted learning to quantify uncertainty in fairness properties inherited from training data. We evaluate data fairness—fairness in the data generating process used for training—and derive estimators for demographic parity, equal opportunity, and conditional mutual information.

Scalable oversight with co-trained monitors

Paper: Improving Scalable Oversight with Co-Trained Monitors

Training workers against fixed monitors can incentivize monitor evasion. This project studies how co-training monitors alongside workers can make AI oversight more robust to monitor evasion. We explore supervised and self-supervised approaches, with theory and code-security experiments in which a worker is trained to fool its monitor.

EEG analysis and machine learning applications

Working towards foundation models for neural data

For my master’s thesis, I trained deep neural networks to predict behavior (especially memory) from neural data collected across different people. This is a kind of transfer learning called domain adaptation: the hope is that a model can pick up patterns shared across brains and use them to make predictions for someone it hasn’t seen before.

The work received a departmental award, though we didn’t end up with a paper. Since then, similar approaches have shown promise for tasks with clearer neural signals than memory, like motor activity, sleep stages, and stress. I’m not working on this now, but I still think it’s a really promising direction.

Decoding brain states and improving memory

Paper: Decoding EEG for optimizing naturalistic memory, Journal of Neuroscience Methods

Oscillatory biomarkers of memory

Paper: Hippocampal theta and episodic memory, Journal of Neuroscience

Sports analytics

In my spare time, I like to dabble in sports analytics. In 2022, my team of Penn grad students was named a finalist for the NFL Big Data Bowl! You can check out our Kaggle notebook as well as the NFL’s press release announcing the finalists and my team’s video presentation of our project.

Our project used high-resolution player-tracking data to predict kick-return outcomes, then built a framework for finding optimal return paths and evaluating players’ decisions.