Fit Happens: The Paradox of Generalization in Overparameterized Models

Deep Learning has revolutionized fields like computer vision, natural language understanding, and data analysis by solving problems that were previously considered unsolvable. According to conventional wisdom in machine learning, overparameterized models are expected to overfit the training data. However, deep learning models often defy this expectation by generalizing well to new data. Numerous researchers have explored why these overparameterized models can still generalize and have proposed various theoretical frameworks to better understand this phenomenon.

In this talk, I will review the initial experimental evidence showing that deep learning models can generalize despite being overparameterized. I will then discuss modern approaches to analyzing these models, providing a more robust theoretical foundation.

Next, I will introduce my own project, which involves implementing one of these theoretical methods in PyTorch. Finally, I will demonstrate how these insights and methods can be applied to large language models, highlighting their practical implications.

Does your application need computer vision support? Contact our sponsor Applied Photons for a free consultation.

About the speaker:
Elliott Miller has been a Machine Learning practitioner for over ten years. Initially working in geospatial, he now spearheads efforts in applying machine learning to next gen wireless networks at Digital Global Systems. He is always interested in what is making an impact in the world of understanding data. Outside of work he attends several local DC meetups and loves to engage with and learn from others in the field.

This is an IN-PERSON event:
Greencourt Innovation Center
12358 Parklawn Dr. Rockville, MD 20852
(ground floor classroom, Wilkins Ave. side)
Can't make the talk? Look for the recording on our Youtube channel.

Agenda

• 6:30 Presentation
• 7:00 Q+A
• 7:30 Pizza & Networking

Parking

• There is free parking on-site.

Public transit

• We are conveniently located near the Twinbrook Metro station.

If you haven't done so yet, please fill out this brief, anonymous survey
so we can keep presenting topics that match your interests!