The goal of this project is to develop and mathematically analyze simple models of empirical phenomena observed in deep learning.

Over the past decade, deep learning has emerged as a dominant paradigm for machine learning, but many aspects of it remain poorly understood. An example of such a question is the following: How does the training process used for deep neural networks uncover features of data that are useful for prediction tasks? Such a question is difficult to answer in full generality, since the “feature learning” phenomenon varies greatly from one application to another. But by restricting to a narrow-enough regime, one may hope to develop a useful theory for understanding the phenomenon.

This project will begin with a literature review to identify robust empirical phenomena of interest, followed by iterative development of probabilistic/mechanistic models that are amenable to theoretical analysis.

This project is eligible for a stipend, with matching funds from the faculty advisor and the Data Science Institute. This is not a guarantee of payment, and the total amount is subject to available funding.

Faculty Advisor

  • Professor: Daniel, Hsu
  • Center/Lab: Computer Science
  • Location: 426 mudd
  • I work on algorithmic statistics and machine learning, with the goals of designing efficient algorithms for learning and data analysis, and understanding the limits of efficient computation for these tasks.

Project Timeline

  • Earliest starting date: 6/1/2023
  • End date: 8/31/2023
  • Number of hours per week of research expected during Spring-Summer 2023: ~20

Candidate requirements

  • Skill sets: Strong mathematical and analytic skills are required. In particular, fluency in multivariable calculus, probability, linear algebra, and algorithmic analysis is a must. A solid background in machine learning and statistics is highly desirable.
  • Student eligibility: freshman, sophomore, junior, senior, master’s
  • International students on F1 or J1 visa: eligible
  • Academic Credit Possible: Yes