Cynthia Tianqing Liu

Cynthia Tianqing Liu

Scholar Title

MIT EECS | Advanced Micro Devices Undergraduate Research and Innovation Scholar

Research Title

Alignment as an Indicator for Two Distinct Training Regimes

Cohort

2018–2019

Department

EECS

Research Areas
  • Artificial Intelligence & Machine Learning
Supervisor

Aleksander Madry

Abstract

Deep learning is quickly becoming more prevalent in both research and industry. To make use of deep learning models, we must train a model on some task. While there are many optimizers and schedulers designed to improve training results or increase convergence speed, there is little understanding as to why these methods work. To fill this gap, we hypothesize that the training process of deep nets can be separated into two distinct phases: a generalization phase and a convergence phase. To provide evidence for this hypothesis, we observe a metric called “alignment,” which measures whether our training steps are moving us towards the solution minima during training. We hope to use this metric to demonstrate that there are two distinct phases of training, and potentially suggest a new training scheduler that takes advantage of this fact.

Quote

Through this SuperUROP project, I would like to explore deep learning while applying machine learning and mathematics to real-life problems. I previously conducted research in applied machine learning and reinforcement learning. This SuperUROP project will let me expand on that experience and tackle machine learning from a more theoretical angle, while allowing me to utilize principled reasoning skills learned in classes such as 18.100B (Real Analysis) and 6.046 (Design and Analysis of Algorithms).

Back to Scholars