Event Information

CS Machine Learning Seminar: Faster Neural Network Training, Algorithmically
Thursday, September 29, 2022
3:30 p.m.
Online presentation
For More Information:
Soheil Feizi

CS Machine Learning Seminar
Faster Neural Network Training, Algorithmically

Jonathan Frankle
Chief Scientist

Zoom link:

Password: 828w

Training modern neural networks is time-consuming, expensive, and energy-intensive. As neural network architectures double in size every few months, it is difficult for researchers and businesses without immense budgets to keep up. In this talk, I will describe one approach for managing this challenge: changing the training algorithm itself. While many companies and researchers are focused on building hardware and systems to allow existing algorithms to run faster in a mathematically equivalent fashion, there is nothing sacred about this math. To the contrary, training neural networks is inherently approximate, relying on noisy data, convex optimizers in nonconvex regimes, and ad hoc tricks and hacks that seem to work well in practice for reasons that elude us.

I will argue that we should exploit the approximate nature of neural network training to change the math of training in order to improve efficiency. I will discuss how we have put this approach into practice at MosaicML, including the empirical approach to research that we take, the dozens of algorithmic changes we have studied (which are freely available open source), the science behind how these changes interact with each other (the composition problem), and how we evaluate whether these changes have been effective. I will also detail several surprises we have encountered and lessons we have learned along the way. In the year since we began this work in earnest, we have reduced the training times of standard computer vision benchmarks (ResNet-50 on ImageNet, DeepLab-v3 on ADE-20K) by 5-7x and standard language models (BERT and GPT on C4) by 2-4x, and we believe we are just scratching the surface.

Jonathan Frankle is Chief Scientist at MosaicML, where he leads the company's research team toward the goal of developing more efficient algorithms for training neural networks. In his PhD at MIT, he empirically studied deep learning with Prof. Michael Carbin, specifically the properties of sparse networks that allow them to train effectively (his "Lottery Ticket Hypothesis"). He will be joining the faculty of the School of Engineering and Applied Sciences at Harvard in the fall of 2023 as an assistant professor. He earned his BSE and MSE in computer science at Princeton and has previously spent time at Google Brain, Facebook AI Research, Microsoft, and Georgetown Law.

Relevant references:


Browse Events By Calendar

Calendar Home

« Previous Month    Next Month »

May 2024
1 2 3 4 w
5 6 7 8 9 10 11 w
12 13 14 15 16 17 18 w
19 20 21 22 23 24 25 w
26 27 28 29 30 31 w

Search Events

Campus Map
Engineering Building Map






Back to top          
CORE Home Clark School Home UMD Home Aerospace Engineering