Week5

Week 5 #

Topic: The Deep Metric Learning Family in Self-Supervised Learning

Keynote Speaker: Xinyao Li, Yiming Zhang, Shengqi Fang

Time: Jul 20, 19:30 - 21:30 pm

Venue: Lecture Hall 3, 302 (SUSTech)

Online Link: TencentMeeting

Compendium #

I. Contrastive Predictive Coding：

  • Introduce Contrastive Predictive Coding, which learns self-supervised representations by predicting the future in latent space by using powerful autoregressive models.
  • Introduce the concept of InfoNCE loss, which is a widely used loss function.

II. The Contrastive Loss Function

  • Introduce the core ideas of Deep Metric Learning and discuss the shortcomings of traditional Metric Learning methods.
  • Discuss DrLIM (Dimensionality Reduction by Learning an Invariant Mapping).
  • Introduce Contrastive Loss and its spring model analogy: Attract-only spring and m-Repulse-only spring.
  • Show the experiments of DrLIM along with traditional Metric Learning methods.

III. The Triplet Loss Function

  • Introduce Triplet Loss: anchor, positive and negative; easy, hard and semi-hard triplets.
  • Discuss the importance of triplet selection and two methods: choosing semi-hard triplets (applied in FaceNet) and Batch Hard.

VI. A brief review of SimCLR:

  • Introduce structure of contrastive learning.
  • emphasize the importance of data augmentation(random cropping and color distortion).

Material #

Slides 1, 2 and 3 for Deep Metric Learning Family from Xinyao Li, Shengqi Fang and Yiming Zhang.

Reference #

  1. Balestriero, R et al. A Cookbook of Self-supervised Learning

  2. R Hadsell et al, Dimensionality Reduction by Learning an Invariant Mapping

  3. F Schroff et al, FaceNet: A unified embedding for face recognition and clustering

  4. A Hermans et al, In Defense of the Triplet Loss for Person Re-Identification

  5. J Goldberger et al， Neighbourhood Components Analysis

  6. K Sohn et al, Improved Deep Metric Learning with Multi-class N-pair Loss Objective

  7. A Oord et al， Representation Learning with Contrastive Predictive Coding

  8. T Chen et al, A Simple Framework for Contrastive Learning of Visual Representations

  9. A Dosovitskiy et al, Discriminative Unsupervised Feature Learning with Convolutional Neural

  10. Z Wu et al, Unsupervised Feature Learning via Non-Parametric Instance-level Discrimination