Whats the point of the class token here?

[Manojbhat09]

Really appreciate your work.
Question : As the topic.

[lucidrains]

@Manojbhat09 it's a common practice in NLP where you have one token that pools information from the rest of the tokens through rounds of attention, usually to classify the sentence at the end. whether it is completely necessary for ViT to work is up to debate. my take is it isn't that important. you could pool all the embeddings from the last layer and probably still get great results at scale

[Manojbhat09]

Thank you for your clear explanation. Hope others too find it helpful.

[zhongyy]

Thank you for the explaination !