Video

Video by Kevin D Schmidt

Dr. Yubei Chen

Air Force Research Laboratory

March 22, 2024 | 01:18:31

In this edition of QuEST, Dr. Yubei Chen discusses his work on Principles of Unsupervised Representation Learning

Key Moments in the video include:
Introduction to Dr. Chen’s lab, mentors, and collaborators
Current Machine Learning Paradigm
Natural intelligence learns with intrinsic objectives
Future machine learning paradigm and unsupervised representation learning
Defining Unsupervised representation learning
Supervision and similarities - spatial co-occurrence, temporal co-occurrence, Euclidean neighborhoods

Main points:
- derive unsupervised representation transform from neural and statistical principles
- simplification and unification of deep unsupervised learning
- the convergence
Neural principle: sparse coding
Statistical principle: manifold learning
Manifold learning, local linear embedding
Sparse manifold transform
Encoding of a natural video sequence
Recap of Main points

Audience questions:
On the three sources of similarity, do you think there is a way to map semantic similarities from crowdsourcing kinds of things like Concept Net?
Are there equivalencies here with cryo-EM analyses?
One of the things that made deep learning what it is their performance on ImageNet and AlexNet, right? Same thing with transformers and language translation, so how are you going to demonstrate that this impressive body of work is better than whatever state-of-the-art is out there? How are you going to demonstrate that it’s useful?
Follow-up: Is there a benchmark or standard data set, which you might produce, that establishes something about representation learning?
Co-occurrence is great for a lot of things but a poor choice for comparison when you have different dimensions for valuation you might want to do? Are you thinking about extending your ideas beyond things that are co-occurring or similar along one dimension and further away?
Is there any sort of procedure for pruning vestigial actions that are no longer necessary for the interpolated tasks that won’t just propagate down for future interpolations?