Machine Learning Street Talk (MLST)

SWaV: Unsupervised Learning of Visual Features by Contrasting Cluster Assignments (Mathilde Caron)

Sep 14, 2020
In a fascinating discussion, Mathilde Caron, a research scientist at Facebook AI Research, dives into her groundbreaking work on the SWaV algorithm for unsupervised visual learning. Joined by Sayak Paul, a machine learning expert, they explore innovative techniques such as online clustering and multi-crop data augmentation. The conversation highlights challenges in reproducing algorithms and the evolving landscape of self-supervised learning. They also discuss the implications of clustering strategies on image recognition and the balance of data versus inductive priors in machine learning.
Ask episode
AI Snips
Chapters
Transcript
Episode notes
INSIGHT

Contrastive Learning and Clustering

  • Contrastive learning methods in image recognition compare pairs of images.
  • This comparison helps differentiate dissimilar images while grouping similar ones.
INSIGHT

Clustering in SWaV

  • SWaV clusters images based on shared semantic content, not individual pixels or colors.
  • This is ensured by using various image distortions and crops, forcing the model to learn consistent features.
ANECDOTE

Reproducing SWaV

  • Sayak Paul and Ayush Thakur reproduced SWaV's results, overcoming implementation challenges.
  • They rewrote some PyTorch functionalities into TensorFlow and clarified doubts with Mathilde Caron.
Get the Snipd Podcast app to discover more snips from this episode
Get the app