I have some variable length input vectors for my own use case of a 'stylistic transfer'-esque process, and I am wondering if anyone knows of a way to engineer an input that maps to a 0 element in embedding space. This would be an element that simply holds space but would be readily overlaid with vector addition of another embedded input.

My rationale is that I could pad the inputs with these zero elements to mask what I don't care about and have a semantically meaningful vector addition in the embedding space.

I wonder if I could permute some training examples with a chosen value which all map to the same output and this would allow a neural net to learn such a feature.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.