Project Scratch

How does the LM implement the following tasks? (Do they work? When do they break?)

Face Generation GAN

Teachers have a hard time getting to know students by face, especially when students are wearing masks. Flashcard apps help, but the teacher can easily “overfit” to quirks of the student photo (background, clothing, etc.).

Potential resources:

Code Analysis for Intro Programming Classes

AI models of (programming) code have improved markedly in recent years (see, e.g., Unified Pre-training for Program Understanding and Generation), but intro programming classes haven’t yet been able to benefit from them. Could you figure out a way to use program understanding methods to give good feedback to CS learners and their instructors? (e.g., help the instructor see patterns in students’ code)

Some code and pre-trained models you might play with:

Learned Multimedia Decoder

Many existing images/videos/audio are locked in poor quality low-efficiency codecs (old personal pictures, audio Bible recordings, video, music, graphics, etc.). If we could invert the poor-quality encoder, we could both recover a more faithful representation of the original and also re-encode the result in a high-efficiency codec.

Deepfake Detection

Make some deepfakes. Try to detect them.

Miscellaneous ideas

Research Projects
Interpretability Initiative