Explore Across-Dimensional Feature Correlations for Few-Shot Learning

Few-shot learning (FSL) aims to learn new concepts with only few examples, which is a challenging problem in deep learning. Attention-oriented methods have shown great potential in addressing the FSL problem. However, many of these methods tend to separately focus on different dimensions of the samples, like channel dimension or spatial dimension only, which may lead to limitations in extracting discriminative features. To address this problem, we propose an across-dimensional attention network (ADANet) to explore the feature correlations across channel and spatial dimensions of input samples. The ADANet can capture across-dimensional feature dependencies and produce reliable representations for the similarity metric. In addition, we also design a three-dimensional offset position encoding (TOPE) method that embeds the 3D position information into the across-dimensional attention, enhancing the robustness and generalization ability of the model. Extensive experiments on standard FSL benchmarks have demonstrated that our method can achieve better performances compared to the state-of-the-art approaches.