SimCLR Training Objective

Explore the SimCLR training objective by understanding how two augmented image views are processed through a neural network backbone and MLP projection head to produce feature embeddings. Learn how the contrastive loss uses similarity maximization between positive pairs and minimizes similarity among negative pairs by computing a similarity matrix. This lesson guides you step-by-step in implementing the SimCLR contrastive loss, preparing you to build self-supervised learning models for unlabeled data.

We'll cover the following...

Network architecture
Contrastive loss
- Similarity matrix
- Positive and negative pairs

Now that we have two augmented versions of the input batch, $T_1(B)$ and $T_2(B)$ , we'll look into other components of the SimCLR training pipeline.

Network architecture

As shown in the figure below, the two augmented versions of an image, $X_i$ (i.e., $T_1(X_i)$ and $T_2(X_i)$ ), are passed through the neural network $f(.)$ to get the penultimate feature representations, $h_{i1}$ , and $h_{i2}$ , respectively. These feature representations are passed again through a multilayer perceptron (MLP) projection head $g(.)$ to get the feature embeddings $z_{i1}$ and $z_{i2}$ ...

1.Introduction to Self-Supervised Learning

2.Pretext Tasks

3.Similarity Maximization and Redundancy Reduction

4.Masked Image Modeling

5.Appendix

SimCLR Training Objective

Network architecture