hi, very nice job and very well explained. The subject of self-supervised learning always inspires me a feeling that I'm watching a sci-fi movie. It's so cool.
I do not understand. If this loss is calculated per image and simply averaged/summed over the batch how does this not just have the encoder collapse every single image to the same set of orthogonal vectors which would minimize the loss?
Great work and nice illustration. However, I recommend that you change the naming of what you call "cross-correlation matrix" (in particular wrt the discussion starting at 20:00). What you optimize is not a correlation matrix in the proper sense (but just a matrix multiplication between ZA and ZB). The values in a real correlation matrix (between two matrices ZA and ZB being both hypothetically composed of 1's) would not be 1's but NaN because the covariance matrix is zero everywhere
Can you elaborate that how they are not 1's? because according to the code he showed the cross-correlation matrix values will be divided by "N" which is the batch size, means normalization. Please correct me if I understood something wrong here. Thanks
Really appreciated this video, really interesting work. The frequent referral to other methods and how they differ was very clear and useful. Thanks!
very nice explanation. I learnt a lot, thanks.
hi,
very nice job and very well explained.
The subject of self-supervised learning always inspires me a feeling that I'm watching a sci-fi movie.
It's so cool.
Excellent talk
I do not understand. If this loss is calculated per image and simply averaged/summed over the batch how does this not just have the encoder collapse every single image to the same set of orthogonal vectors which would minimize the loss?
very well explained
Great work and nice illustration. However, I recommend that you change the naming of what you call "cross-correlation matrix" (in particular wrt the discussion starting at 20:00). What you optimize is not a correlation matrix in the proper sense (but just a matrix multiplication between ZA and ZB). The values in a real correlation matrix (between two matrices ZA and ZB being both hypothetically composed of 1's) would not be 1's but NaN because the covariance matrix is zero everywhere
Can you elaborate that how they are not 1's? because according to the code he showed the cross-correlation matrix values will be divided by "N" which is the batch size, means normalization. Please correct me if I understood something wrong here. Thanks
It is a matrix of correlations between specific features of the latent, averaged over the batch.