Learning Adaptive Classifiers Synthesis for Generalized Few-Shot Learning

Object recognition in the real-world requires handling long-tailed or even open-ended data. An ideal visual system needs to recognize populated head concepts reliably and meanwhile efficiently learn about emerging new tail categories with a few training instances. Class-balanced many-shot learning few-shot tackle one side of this problem, by either strong classifiers for tail. In paper, we investigate problem generalized (GFSL)—a model during deployment is required shots simultaneously classify classes. We propose ClAssifier SynThesis LEarning (Castle), framework that learns how synthesize calibrated addition multi-class classes shared neural dictionary, shedding light upon inductive GFSL. Furthermore, an adaptive version Castle (a Castle) adapts conditioned on incoming examples, yielding allows effective backward knowledge transfer. As consequence, can handle GFSL from heterogeneous domains effectively. demonstrate superior performances than existing algorithms baselines MiniImageNet as well TieredImageNet datasets. More interestingly, they outperform previous state-of-the-art methods when evaluated standard criteria.