Few-Shot Image Classification with Meta Random Fourier Features Based on Bayesian Optimization

Deep learning methods rely on large amounts of labeled data, and their highly nonlinear structure limits their wide applicability in the field of small samples. The kernel method has become an effective solution for nonlinear low-data regions. In this paper, a meta-learning network Meta-RFFNet based on convolutional random Fourier feature operators is proposed to improve the generalization ability of deep learning models in small samples and cross-domain classification tasks. It uses random Fourier features to construct deep kernel network, balances the difference of sample distribution by synthetic gradient network, and realizes parametric inference based on empirical Bayesian model. This approach overcomes the dependence of deep learning models on large amounts of labeled data, and provides mathematically interpretable mappings for each layer of the kernel network. Experimental evaluations on mini-ImageNet and CIFAR-FS datasets show that the proposed algorithm outperforms existing standard few-shot classification methods, with classification accuracy of 70.67% and 79.33%, respectively, at 5-shot settings.