Neural Architecture Search and Meta Pseudo Labels
Wednesday, April 14, 2021
4:00 – 5:00 pm
Zoom webinar: https://berkeley.zoom.us/s/99825370819
Principal Research Scientist
I will talk about our recent works on computer vision:
- Smart architecture search methods to discover a highly efficient model called EfficientNet, and EfficientNetV2 which are much more efficient than previous image classification models.
- Using unlabeled data to improve accuracy and robustness of vision models. The method, named “Noisy Student”, requires generating pseudo labels on unlabeled data and using them to train a student. The method significantly improves the robustness of the student on hard test sets, ImageNet-A, ImageNet-C, and ImageNet-P.
- An improvement of the Noisy Student method called Meta Pseudo Labels that creates a feedback loop from the student network to the teacher network to improve the teaching of the teacher network. This method combined with EfficientNets reaches the top-1 accuracy of 90% on ImageNet.
Quoc Le is a principal research scientist at Google Brain. He obtained his PhD at Stanford, and was advised by Andrew Ng. Quoc started out as an intern when Google Brain was founded. Together with Google Brain team members, Quoc worked on seq2seq, pretrained language models, neural architecture search etc. His current interests are self-improving machine learning (e.g., AutoML), and semi-supervised learning, which are basically the two themes of the talk above.