Implement cross attention between labels and text embedding.
Examples:
Here, they call it Label Attention (queries are labels and key/value are text embedding, the labels attend to tokens to get the label-attend text repr. )
Here, they compute also the text-attended label representations (query = text embedding, key/value = labels)...
This would enable another explainability pipeline, that would be independent from Captum, above all internal to the model structure itself.
Implement cross attention between labels and text embedding.
Examples:
Here, they call it Label Attention (queries are labels and key/value are text embedding, the labels attend to tokens to get the label-attend text repr. )
Here, they compute also the text-attended label representations (query = text embedding, key/value = labels)...
This would enable another explainability pipeline, that would be independent from Captum, above all internal to the model structure itself.