Task-specific method-agnostic metric for few-shot learning

Heng Wang, Yong Li

Abstract: Metric-based few-shot learning (FSL) methods have been attracting more and more research attention since they reflect a simpler and more effective inductive bias in the limited-data regime. The episodic evaluating method is widely used in the metric-based FSL methods, and the task-wise relative metric is critical to improving the performance of the episodic method. However, the commonly used metrics in existing metric-based FSL methods typically measure the absolute distance in a smooth and uniform feature space. Observing this, this paper proposed mapping the features into the task-specific sub-space by designing the correlation matrix of task-specific prototypical vectors, which induces a task-specific method-agnostic (TSMA) metric. The TSMA can be viewed as an adaptive linear classifier and hence is method-agnostic. In addition, the TSMA is manually designed and thus is parameter-free. The extensive experiments evaluated on various datasets show that TSMA outperformed the SOTA methods by 1.5–4.4%. And the ablation study shows that TSMA could adaptively adjust the scale of the similarity items and the scaling items, allowing for the models to easily optimized.