Gaussian Prototype Rectification For Few-shot Image Recognition

Few-shot Learning aims to recognize novel categories with scarce training data, which is a challenging problem in computer vision. Many single-point prototype-based methods are noise-vulnerable and deviate from the expected prototypes in a low-data regime. To address these issues, we firstly model the prototype as a multidimensional Gaussian distribution. Such a vanilla Gaussian prototype gains no significant performance improvement in complicated datasets (e.g., miniImageNet, tiered-ImageNet) due to weakly representativity of the latent vectors, this paper further proposes to rectify it by mutual information. In a nutshell, we use a multidimensional Gaussian distribution with local mutual information estimation and maximization to represent each prototype. The distribution estimation learns to construct a dimension and class-dependent distance metric while local mutual information maximization rectifies the prototype and guarantees the representation's suitability for few-shot classification. Besides, to solve prototype deviation in the low-data regime, we propose the transductive Kalman (t-Kalman) fusion method. To the best of our knowledge, we are the first to give the theoretical derivation of Kalman data fusion for few-shot scenario. Extensive experiments show our method outperforms previous methods on many benchmark datasets under various settings. Ablation study shows the mutual information and t-Kalman bring in significant improvement on miniImageNet (6.15% and 5.16%) and tieredImageNet (9.25% and 4.76%) under 1-shot and 5-shot task, demonstrating the effectiveness of the proposed Gaussian prototype rectification.