The NT-Xent loss · MAAI7602 Applied Computer Vision

NT-Xent is the SimCLR loss. For each view it is a cross-entropy over the other views in the batch, with the other view of the same image as the correct answer and a temperature that sharpens the softmax. How does the NT-Xent loss of SimCLR turn the similarities within a batch into a training signal?

This page needs JavaScript for its explanations, activities and practice questions.