Here is what I meant: Standard SVM classifier: Maps an input sequence to a 0-1 label. Example: Take a paragraph and return its sentiment. During training, labe

Here is what I meant: Standard SVM classifier: Maps an input sequence to a 0-1 label. Example: Take a paragraph and return its sentiment. During training, label is specified. Transformer's SVM: Takes input sequence, suppresses bad tokens and passes good tokens to the next layer. This is a token-selector rather than classifier. Example: Take a paragraph and output the salient words in the paragraph. We don't know which words are salient during training, the model has to figure them out during training.