Figure 6.
The diagram begins with x 1 and additional samples. Weak transformation t w produces anchor samples, including x 1 anch. Strong transformation t s produces class samples, including x 1 cls. Both sets pass through f theta. The anchor samples produce anchor features, including z 1 anch. The class samples produce class features, including z 1 cls. A dot product combines the anchor and class features. Softmax produces output logits y hat. The targets y tilde form an identity matrix with 1 along the diagonal and 0 elsewhere. The output logits and targets are passed to cross-entropy loss L C E. Theta is updated using the gradient of the summed cross-entropy losses over B samples.Instance discrimination in contrastive learning is enhanced by using two-view augmented data with weak bias and strong bias relative to as the anchor samples and class samples , respectively