Richard Lanas Phillips


This is a picture of me smiling
My research interests are in the ways we make decisions around ML algorithms. How do we calibrate around ML predictions? How do model decisions affect our community development on social media platforms? How might algorithms amplify or augment our existing biases? I'm currently using the tools of network science, algorithmic transparency and interpretability, and mechanism design towards exploring these questions.



Figure from upcoming work showing a lower bound example

One for One, or All for All: Equilibria and Optimality of Collaboration in Federated Learning

We study fundamental issues related to equity and collaborative incentive in federated learning. Rather than assuming that one benefits from learning over as much data as possible, we consider what burdens individual populations should shoulder given their performance demands and the structure of their data as related to other collaborators. We build models for PAC learning, random sampling, and resource sharing and compare these; we consider the existence of equilibria, feasibility, and lower bounds for the gap between global optimality and optimality under incentive constraints.



WILDS benchmark logo

WILDS: A Benchmark of in-the-Wild Distribution Shifts

I studied several datasets on the WILDS team. We looked for the existence of problematic covariate shift in real-world datasets in different domains. The initial release of the data contains 7 bechmark datasets with specified domains, metrics, and baselines. Across the project, however, many other datasets were also explored. For my contribution, I studied covariate shift through issues of fairness and geography in several algorithmic fairness-related datasets.



Encoder and decoder diagram. Shows input and protected feature being encoded, decoded and re-fed through the model. At the same time a discrimantory model is meant to train the encoder to obscure sensitive information in the nonsensitive data.

Disentangling Influence: Using Disentangled Representations to Audit Model Predictions

We showed how disentangled representations might be used for auditing. This is useful for finding proxy features and searching for potential introduction of bias in correlated features. Our solution allowed for the explicit computation of feature influence on either individual or aggregate-level outcomes.



Scatterplot showing that Black people in the dataset face greater relative uncertainty both before and after active learning, even though balanced accuracy went up.

Interpretable Active Learning

This project proposed a way we might explain why active learning queries are being suggested to a domain expert. We built this to help batch active learning queries in a way that let chemists design natural experiments. We also explored some auditing and fairness applications of the proposed framework.



Dark Reactions Project Logo

Dark Reaction Project

At Haverford College, I worked on the Dark Reaction Project team. The goal of this project was to take thousands of unused, "failed" reactions and leverage them into a machine learning model. This was successful in speeding up exploration and the system helped expose human biases in the exploratory synthesis pipeline.