Wenhao Li

MCrossEntropy​(pi,yi)+12​(ℒD, D}}). (12) We detail the algorithm of the entire training process in the Appendix. During testing, the prototype with textual prompts ctc_{t} is obtained by averaging the enhanced support features. The prototype with visual prompts cvc_{v} is computed by averaging the expanding support set in Eq. (3). The final classification prototype CC is then obtained by integrating cvc_{v} and ctc_{t} in a convex combination manner [22]: C=u​ct+(1−u)​cv,C=uc_{t}+(1-u)c_{v}, (13) where u∈[0,1]u\in[0,1] is a manually controlled fusion factor, determined in the validation set. Table 1: Results (%) on miniImageNet [9] and tieredImageNet [57]. The average accuracy with 95% confidence interval is reported. Bold and blue font indicates the best and suboptimal results. Model Venue Backbone ≈\approx # Params miniImageNet tieredImageNet 1-shot 5-shot 1-shot 5-shot MatchNet [9] NeurIPS’16 ResNet-12 12.