Hey, I'm Evan! ππΌ
![Evan](/assets/images/me.jpg)
I am a computer science Ph.D. student at MIT. My research focuses on understanding what deep neural networks learn from large quantities of data: what concepts are encoded in their features, and how can we communicate the network's learned algorithm to humans? I am extremely fortunate to be advised by Jacob Andreas and to work closely with Antonio Torralba.
Before graduate school, I earned my B.S. in computer science and mathematics from the University of Wisconsin-Madison. There, I developed a natural language interface for the educational programming environment Blockly and studied how to synthesize example programs for students using a noisy, hand-drawn sketch as a specification.
I also worked for Google as a software engineer and contributed to Chromium.
Research
![neuro descriptions](assets/images/2022-milan.gif)
We present a procedure to automatically generate natural language descriptions of neurons in computer vision models. These generated descriptions support important interpretability applications: we use them to analyze neuron importance, identify adversarial vulnerabilities, audit for unexpected features, and edit out spurious correlations.
![visual concept vocabulary](assets/images/2021-visual-vocab.gif)
GANs sometimes encode visual concepts in their latent space as linear directions. We construct a visual concept vocabulary for pretrained GANs, consisting of latent directions and free-form language descriptions of the changes they induce. We then distil the vocabulary into simpler, one-word visual concepts (e.g., snow or clouds).
![low-dimensional subspace](assets/images/2021-low-dim-probes.png)
How do word representations geometrically encode linguistic abstractions like part of speech? We find that many linguistic features are encoded in low-dimensional subspaces of contextual word representation spaces, and these subspaces can causally influence model predictions.
Teaching
I had the pleasure of mentoring an MSRP summer intern on a research project. She developed a language-based image editing tool for images generated by GANs.
MITβs primary NLP course, typically taken after a first course in ML. I wrote homework assignments, planned recitations, and led weekly office hours.
For three years, I tutored underrepresented students in UW-Madison engineering programs on introductory computer science and math classes. I also developed tutoring software to support the tutoring by request and drop-in tutoring services.