Open OUTSIDER-CODE opened 8 months ago
in the formula 16
Thank you for your interest in our work.
Note that only the unseen category prototypes are utilized in our approach, aligning with the definition of zero-shot learning as outlined in reference [1]. For instance, generative-based GZSL methods require these category prototypes to synthesize visual features for unseen classes. Furthermore, in embedding-based approaches, many works, such as DVBE, TCN, and DPPN, rely on these class prototypes for score computation. [1] Triple Verification Network for Generalized Zero-Shot Learning, TIP19
Why does
unsenn
appear in the loss function even though it is not used during the training phase to mitigate bias?