AI-powered
podcast player
Listen to all your favourite podcasts with AI-powered features
Using BRL and SimCm in Deep Learning?
Swav is sort of sitting in between the contrastive and non-contraceive method. When you train the system to classify into those prototypes, sort of fake categories, if you want, invented categories,. You're implicitly doing a bit of contrastive learning. Because of the normalization of softmax, you can only give a high score to 1k, or automatically the score, the other ones go down. But what I'm actually more excited about is other non-contressive methods, which of course I'm biased because they came from my group at Facebook.