InfoNCE loss at 11:14 looks odd as Dp is the distance notation at 9:00, but you say its related to probabilities. It would break the flow to introduce new notation though. But as it stands it was a little confusing to me to see that the loss would be minimized by maximizing Dp. I checked the paper and it seems the term is an approximator for "mutual information" which we want bigger for positive samples. At least thats my rough understanding... Thanks for the video its a fantastic explanation!
Thanks, it pushes negative pairs apart until their distance reaches the margin, by minimizing the difference between the margin and the distance between the points. This is the quantity in red at 06:40 :)
How exactly in the original contrast loss, does y = 0 in the positive case and the y=1 in the negative case? In addition what is y representing here? 6:27
Nice explanation! It still isn't clear to me how to choose the metric to determine how similar or dissimilar two samples are, is it also learned by the network?
At 12:04 you say that SimCLR select multiple negative pairs and then you show a picture of a cat, and a dog. I am confused, the second dog picture is also considered as a negative pair even though it's the same animal? If yes, does this mean the model train to lower the distance ONLY with the original image even though other could be dogs?
@@Deepia-ls2fo That is very interesting, thank you for your answer, I have another question if you do not mind At the end when comparing classification accuracy you compare supervised, SimCLR+finetune and SimCLR, the last one have me confused, how can the model without any finetuning even work for classification? Or do they not count a trained dense layer that learn to use the latent space of SimCLR for classification, and SimCLR+finetune mean finetuning the latent space instead? My question is that does fine-tune mean finetuning a dense layer or the latent space? Your videos are high quality and I really love them, sometimes I just wish they would be longer and slightly more into the implementation details, thank you! Edit: Regarding my first question, since the negative pair can be the same class (if we imagine the ultimate goal is classification), would a low amount of class (let's say only 2) lower the quality of the latent space due to a high amount of class "collision" ? And in the opposite if there is hundreds of class it will rarely select the same class as a negative pair and improve latent space representation?
@@itz_lucky6472 I strongly advise you to read the SimCLR paper as it is a very easy read and they detail everything. About the classification task: for SimCLR they use what we call "linear eval", meaning they plug a fully connected head on the model and train only this part. The difference between "SimCLR" and "SimCLR fine-tune" is that the weights of the backbone are modified in a supervised fashion with a small portion of the data for "SimCLR fine-tune". For your second question I did not read a lot about this, and I'm myself new to self-supervised learning in general, so I can't answer for sure. I guess you could easily do the experiment with 2 MNIST classes though. Intuitively I think taking many semantically similar objects and treating them as negatives is bad for the representation space.