Self-Training With Noisy Student Improves ImageNet Classification

Explore an innovative approach to improving ImageNet classification through self-training with Noisy Student in this 22-minute Launchpad video. Delve into key concepts such as knowledge distillation, soft and hard labels, and the interplay between self-training and distillation. Examine the Noisy Student training algorithm, understand the effects of noise, and learn about pseudo labels. Discover the architecture behind this method and analyze experimental results, including its impact on robustness. Gain valuable insights into this cutting-edge technique for enhancing image classification performance.

Introduction
Knowledge Distillation
Soft and Hard labels
Self-training and distillation/ Noisy student
Algorithm: Noisy student training
Noise effect
Pseudo labels
Architecture
Experimental results
Robustness results