Canonical mean filter for almost zero-shot multi-task classification

The support set plays a key role in providing conditional prior for fast adapting the feature extractors in few-shot tasks. The representative few-shot method CNAPs used a simple conditional feature encoder to extract the prior, which was a conditional feature of the task-specific support set. However, the strict form required by the support set makes its construction difficult in practical applications. Motivated by ANIL, we first investigate the role of the adaption in CNAPs by designing Almost Zero-Shot (AZS) tasks in which a fixed support set is used across all tasks instead of the task-specific support set. The AZS experimental results indicate that the adaptation contributes little to the feature extractor. Nevertheless, simply removing the adaptation from CNAPs will cause severe performance drop on some sub-datasets of the Meta-Dataset. To alleviate this problem, we further proposed a Canonical Mean Filter (CMF) module that maps any support set to a canonical form. The CMF yields a stable conditional feature, and hence allows for CNAPs to perform well without the conditional feature encoder and the parameter adaptation network at the test stage. CNAPs embedded with CMF significantly outperform the vanilla CNAPs for the AZS tasks, and even achieve comparable performance to one-shot tasks but with 40.48% parameter reduction and without task-specific support sets at the test stage.