Kurtis Evan David

I graduated with an MS in Computer Science at UT Austin, where I worked on interpretability and bias of deep learning. I had the pleasure of being advised by Professor Qiang Liu and Dr. Ruth Fong. I also completed my undergraduate studies at UT, double majoring in Computer Science and Mathematics, and acted as a competition organizer for UT Machine Learning and Data Science.

Previously a Software Engineer at Facebook AI Applied Research working on algorithmic fairness.

I am currently a Research Scientist at Protopia AI working on data privacy.

Email  /  CV  /  GitHub  /  LinkedIn  /  Google Scholar

Broadly, I'm interested in deep learning, computer vision, and NLP. In school I worked on methods for interpretability and understanding bias of convolutional neural networks. I am also interested in adversarial robustness and model pruning.

We show that we can recreate early work in debiasing word embeddings within convolutional neural networks using an extra meta step. We evaluate on a multitude of conditions and show it to be competitive to adversarial debiasing methods.

To increase trust from model agnostic anchors, we instead conditionally sample backgrounds generated from a generative adversarial network, rather than random images from the dataset. We address additional complexity of our method through a diverse encoder, and show that our explanations can lead to smaller and higher precision anchors.

Developed tools for measuring bias on large scale ML models. I supported novel research on our platforms, as well as privacy preserving methodologies.

Primarily worked on adversarial robustness of neural networks and understanding Adversarial Examples Are Not Bugs, They Are Features. Investigated promising defenses using neural network pruning and Fourier space analysis.

Developed new Instagram Stories ranking models, balancing Direct and Stories metrics. Deployed IG share sheet rankings with significant gains, and tested Direct caching. Lastly, implemented Lottery Ticket Hypothesis on production ranking models.

Incorporated new user side NLP features into the feed ads ranking model. Explored possible connections between user and ads side features to increase metrics. Implemented a new pooling layer into their models, pushed to open source.

Built an anomaly detection model for internal audits using PCA and Isolation Forest. Running an automated proof-of-concept supervised learning model for future continuous audits that achieved an F1-score of .93 with .08% anomalies (100 of 120K). Additionally supported a document analysis tool that found similarities between two specification documents using a bag-of-words implementation.

Analyzed the relationship between cancer drug sensitivity data and their reported side effects, scraped from the FDA’s Adverse Event Reporting System database. Developed an organ level side effect classifier for novel cancer drugs.

Legendary website template from Jon Barron.