- I am still considering the zero-shot problem here. What I mentioned is that ‘the mapping from image embedding to class embedding’ could be implemented by the approach proposed in your post or alternatively by multi-output regresssion. After the mapping function is found, we still do the same thing: feed in a zero-shot image, get the predicted class embedding, and do the prediction.
- I just come across this post https://medium.com/@alitech_2017/from-zero-to-hero-shaking-up-the-field-of-zero-shot-learning-c43208f71332 There is a figure there showing ‘bias towards source classes’. Now I feel that we need more info than class embedding, e.g., adding descriptions or attributes on classes.
- I had tried the regression approach on a fruit dataset (http://www.vicos.si/Downloads/FIDS30), but the result is not good.
