TASML: Two-Stage Adaptive Semi-supervised Meta-learning for Few-Shot Learning

Human vision system(HVM) is a remarkable computational system that excels at analyzing complex visual scenes and recognizing objects with high accuracy and efficiency, and interrelated hierarchies for integrated processing of visual inputs is a signature of human vision intelligence. Similar to the process of human recognition of unknown objects, meta-learning enables models to “learn to learn" by providing a small number of “support" samples to solve the few-shot or one-shot problem in real scenarios. Inspired by this, we propose a two-stage adaptive semi-supervised meta-learning(TASML) framework with a hierarchical structure to solve the few-shot learning task. Specifically, parallels to the transition from primary to advanced visual cortex procedure in HVM, the invariant information of the target are obtained and features fusion is performed through an unsupervised representation sensing phase. Subsequently, a gradient-based meta-learning mechanism is used to simulate the execution of object targeting and recognition processes in HVM. In addition, a global context-aware(GCA) module is proposed in conjunction with our framework to follow the HVM’s access to context and semantics of visual object. Extensive experiments conducted on Mini-ImageNet and CIFAR-100 datasets validate the effectiveness of our framework and module on few-shot classification tasks, which achieves \(76.66\%\) , \(77.8\%\) accuracy on 5way-5shot classification and \(63.96\%\) , \(62.8\%\) accuracy on 5way-1shot classification for two datasets, respectively.