Are most zero-shot learning methods generally unable to provide more information than distance in an abstract embedding space, as I claim in the motivation section? For function prediction, there isn't currently any natural language generation methods. DeepGOZero can give zero-shot predictions, and my understanding is that the new functions have to have some ontology relation to known functions. Can it create new terms with new ontology relations, or must the user input the relations with existing terms into the model in order to predict for them? If it's the latter case, it's even more restrictive, since you can't actually predict whether a protein has a new function without prior knowledge.
Are most zero-shot learning methods generally unable to provide more information than distance in an abstract embedding space, as I claim in the motivation section? For function prediction, there isn't currently any natural language generation methods. DeepGOZero can give zero-shot predictions, and my understanding is that the new functions have to have some ontology relation to known functions. Can it create new terms with new ontology relations, or must the user input the relations with existing terms into the model in order to predict for them? If it's the latter case, it's even more restrictive, since you can't actually predict whether a protein has a new function without prior knowledge.