MetaKernel: Learning Variational Random Features With Limited Labels.

Area of Science:

• Machine Learning
• Artificial Intelligence

Background:

• Few-shot learning addresses the challenge of generalizing from limited data.
• Extracting prior knowledge from related tasks is key for rapid adaptation.

Purpose of the Study:

• Propose MetaKernel, a novel meta-learning approach using random Fourier features for few-shot learning.
• Develop data-driven, task-specific kernels by leveraging shared knowledge.

Main Methods:

• Learn variational random features via variational inference, treating the basis as a latent variable.
• Incorporate shared knowledge using a long-short term memory module for posterior inference.
• Employ conditional normalizing flows for richer posterior distributions over random Fourier bases.

Main Results:

• MetaKernel generates informative and discriminative kernels, enhancing few-shot learning performance.
• Extensive experiments on image classification and regression tasks show consistent improvements.
• Ablation studies confirm the effectiveness of individual components.

Conclusions:

• MetaKernel achieves comparable or superior performance to state-of-the-art methods across fourteen datasets.
• The proposed method effectively improves few-shot learning through learned, expressive kernels.