Check out my MSR internship work for an automatic way to uncover skill-level differences of models like GPT-4o, Claude Sonnet, and Gemini that overall accuracy hides. 

Will give a lightening talk at Google's Theory and Practice of Foundation Models workshop in November.

ArtSavant accepted to the EvalEval workshop and the RBFM workshop.

A practical tool to protect artistic copyright, inspired by legal literature (co-authored by a law professor!) and a method to improve T2I compositionality.  

Large collaboration, primarily with folks from Meta FAIR, provides an extensive introduction to vision-language modeling. [ arxiv ]

[ Read more ]

Transparent zero-shot classification (Meta internship work), and uncovering geographic disparities in LLMs with WorldBench. 

Will work in the AI Frontiers team, led by Besmira Nushi, to study responsible ways to use the most cutting edge AI models.

Interpretable failure model extraction with PRIME, and debut of WorldBench @Global AI cultures and Robust Foundation Models WS.

Proposed a PhD thesis entitled Towards Reliable Deep Learning via Efficient Interpretation and Intervention. [ View slides ]. Hoping to graduate around December '24. 

A method and evaluation framework for how to add new data to debug a model without overfitting to errors. [ Read the paper ] 

Two works on model and data interpretability accepted to be presented as Spotlights at Neurips '23 in New Orleans. 

Worked with Diane Bouchacourt, Mark Ibrahim, and Mike Rabbat on Fair Zero-shot Classification using VLMs and LLMs. 

"Text-to-Concept (and back) via Cross Model Alignment" [ Paper + Video ] [ Twitter Thread ] 

"Text2Concept: Concept Activation Vectors Directly from Text" [ Read more ] [ Watch video ]

Presented to the Serre Lab of Carney Institute of Brain Science.

Presented as part of the CMU AI Seminar Series. 

Received endowment award from Metropolitan Washington Chapter of the ARCS Foundation.

Recognized as among top 2% of Graduate Assistants in the University of Maryland. 

Check out the posters and presentation videos 

[ Watch Tradeoffs Video ] [ Watch Hard ImageNet Video ]

Two new works, both focused on developing data solutions to model problems. [ Read Data-Centric Debugging paper ] [ Read Spuriosity Rankings paper ]

Facilitated project assessing biases in text-to-image models at largest hackathon for underrepresented genders. [ About Technica ] [ My group's project ]

First time with multiple first author acceptances :) [ Read HardImageNet paper] [ Read Tradeoffs paper ]

Deep Learning crash course for high schoolers from underrepresented groups; a project on synthesis with robust models. [ Read more ]

Launched a challenge dataset with segmentations for objects with strong spurious cues. [ Download Hard ImageNet Dataset ] 

We discover adversarial training can increase spurious feature reliance, decreasing distributional robustness. [ Read the paper ] 

View at salient-imagenet.cs.umd.edu/home.html

Selected for Oral presentation (top 4.2%) [ Read the paper ]   [ Download RIVAL10 Dataset ]

Led data science projects about climate economics and deep learning in music & art.

[ Read the paper ] 

Had the privilege of introducing deep learning to  6 talented highschool girls  [ Read more ]

I will be studying under Prof Soheil Feizi on deep learning foundational theory. 

High distinction, and honors from both math and computer science departments.

Coded in python, java, scala, javascript, and spark.

Music Machine Learning Workshop. Work done while abroad at ETH Zurich. [ Read the paper ] 

First Mudder to study abroad at ETHZ. Recognized for work done at NIH. [ Read more ]

Given to one freshman annually for top academic performance and community contribution.