1 paper
James B. Simon, Maksis Knutins, Liu Ziyin +3
We present a simple picture of the training process of joint embedding self-supervised learning methods. We find that these methods learn their high-dimensional embeddings one dime…