Tong Wu

This table is the same to Table 1 in the paper. However, the performance of FADI is slightly less-than-satisfactory on higher shot and novel split2. Distribution Alignment (Section 3.2 in paper) is not the real distribution of novel classes. As mentioned in Section 4.4 (Paragraph: Superiority of Semantic Similarity over Visual Similarity), the visual representation is not reliable under data-scarce scenarios due to the existence of co-occurrence, thus we adopt semantic as similarity measurement, However, it can not capture some other cues that matter to the performance, e.g., shape similarity, which has been proved to be beneficial to the knowledge generalization [ Empirically, 'bird' is more similar to'aeroplane' from the We study the hyper-parameters, i.e., We found the optimality of hyper-parameters is related to the confusion level with models, hence we adopt'association + disentangling' We first investigate the importance of each component in Table 3. On the contrary, it also has a suppression of novel classes.