A question to follow, a project to make your own. Browse these possibilities, or bring an idea of your own.
What can a learning system safely forget? Compare simple predictors and compression methods on sequences with hidden patterns, then change the patterns or ask a different question about the same data. Investigate which summaries remain useful, what information has been lost, and whether good prediction alone is enough to support further learning.
Link to this idea
Can changing how a problem is represented make it easier to learn and explain? Create a small shape-recognition task and compare the same simple learning model using raw pixels or features designed to capture geometric structure. Vary training data, noise, and orientation; test which representations improve predictions and whether the apparent structure actually matters when altered or removed.
Link to this idea
Build a movie recommendation engine that learns people's tastes and predicts what they'll enjoy. Explore collaborative filtering, machine learning, and the tradeoffs between recommending familiar favorites and helping people discover something unexpected.
Link to this idea
How can a system verify work and coordinate decisions when producing answers is easy? Build a prototype in which AI-generated answers or code changes must pass explicit checks before acceptance. Use known-good and deliberately flawed submissions to compare self-checking, independent checking, and agreement among checkers. Measure errors and effort, including when the same participant both proposes and approves work.
Link to this ideaAll 25 project ideas, the mentoring invitation, and next steps in one illustrated lookbook.
Mike Andrews P'28 · Independent ISP mentorship