Meta-learning of Semantic Attention Prototype Correction Based on Bayesian Estimation

The mean visual prototype is a mainstream meta-learning method for few-shot learning. It is often combined with semantic information to compensate for the lack of training samples. As the number of samples increases, the semantic gain reduces or even becomes noise, leading to a bias in the semantic attention prototype. To address the problem of semantic bias, we propose a prototype correction strategy based on Bayesian estimation. The strategy corrects the original semantic attention prototype defined from the sampling perspective according to the overall information. The probability of the corrected semantic attention prototype is calculated from the probability of the visual prototype and the semantic attention. The semantic attention prototype's posteriori probability center is taken as final prototype. Experiments on Mini-ImageNet and Tiered-ImageNet exhibit our method outperforms existing methods using semantic information by 0.7%~8% in terms of classification accuracy.