Figure 3.
The diagram begins with x a. Delta is added to x a to create adversarial x a. Adversarial x a and x b pass through f theta to produce latent representations. The contrastive loss L C L is calculated using adversarial x a, x b and theta. The representation z 1 a adversarial serves as an anchor in one branch, while z 1 b serves as an anchor in another branch. Arrows labelled attract connect each anchor with its corresponding paired representation. Arrows labelled repel connect each anchor with the other representations. The gradient of L C L with respect to delta is used to update delta.Self-supervised adversarial training adopts the perturbation estimated according to contrastive learning loss where instance discrimination is performed