Projects

Here are some projects I’ve worked on:

Publications

In Preparation

Estimating the Effect of Introducing Modern Plant-Based Analogs in Restaurants
J. Winslow, J. Peacock, J. Dhargawe, H. Sherman, S. Green, M. B. Mathur

Probing Adaptive Decision-Making Under Uncertainty Using Extended Hidden Markov Models
R. Singha, J. Winslow, R. Kim, J. T. Serences, N. Rungratsameetaweemana

Under Review

Asymmetric Social Representations in the Prefrontal Cortex for Cooperative Behavior
Y. Cheng, Y. Chen, M. Kwak, R. P. Kempner, R. Singha, J. Winslow, R. Liu, U. Khan, T. Spangler, A. Khan, T. Pereira, M. Whiteway, E. S. Schaffer, N. Rungratsameetaweemana, N. Yang, H. Z. Wu

2025

Reanalysis of “A Longitudinal Study of Meat Reduction Over Time in the UK”
J. Winslow, M. B. Mathur

RL & Explainable AI

Leader-Follower Dynamics in Multi-Agent RL

Does one agent have what it takes to become a leader? The real research questions here are: Can RL agents cooperate? Yes. Do they adopt leader-follower dynamics in our environment? Somewhat. To the extent that mice do? Not at all.

Layerwise Relevance Propagation for Feature Networks

Unfortunately, many black-box saliency maps seem to be edge detectors. If a confused model calls a picture of an elephant a giraffe (through adversarial training), and everything in the interpretability method says the model still thinks it’s an elephant, it’s not a very good interpretability method now is it.

Statistics & Applications

ITS with Multilevel Integer-Valued GARCHs

Aka the ITSMLZINBINGARCHX. Thank you Stan for allowing this large, unwieldy model to converge. A win for partial pooling, and a success for classical time series forecasting in balancing interpretability and prediction capability.

RAG for Stan Modeling

I’m sure people have tried using RAG with LLMs on more domains and in more ways than I can count, but I still think there may be room on the table for forcing LLMs to use best practices when those standards have already been exhaustively detailed somewhere.

Remember that French “whale” who made a fortune on polymarket betting on neighbor polling? Well, let’s see whether he was on to something in guessing political affiliations or just on the right side of variance. From our analysis, it seems social ties seem to reveal at least a small amount information about individuals’ politics hidden from standard polling.

Cost-Effective Lead Sampling with GPs

I’d say it this is a perfect use case for high-dimensional Bayesian Gaussian processes, especially when you have few samples. If we can guess where the lead isn’t in a containment site (and be confident about it) then we don’t need to dig it all up.

Malaria Drug Effects

A clinical trial in Africa. Looks promising but can’t tell more until it’s published.

Nephrology Survival Analysis

Similar to the malaria trial. Nephrology (from nephrons) is the study of kidney disease.

Language Application Effectiveness

Another yet-to-be-published secret…

Textbook Parser

Textbooks can sometimes be incredibly uninteractive, and I wanted to change that. So far, breaking down Elements of Statistical Learning into knowledge graphs and interactive visuals seems to be working, albeit with much misplaced emphasis by Claude. It works well as an example since 1) it’s very popular, and 2) I know it, so I can validate.

Fun Stuff

Auto Anki (Chinese Proof-of-Concept)

I’ve been wanting to make something to auto-generate anki cards for a long time, and it works perfectly. Upload a show and it will create single-word cards based on your current vocabulary including video, audio, full text context, and translation.

Pokemon TCG Scanner

Pokemon TCG scanner! I’m surprised by how much other scanners are missing, so that was my focus: multi-card identification, variant detection (e.g., 1st edition), multiple language support, and individual-sales condition-level pricing.