HC-BDC: Human Cognition-Inspired Bayesian Distribution Calibration for Few-Shot Classification

ICLR 2026 Conference Submission14931 Authors

Keywords: few-shot learning, Bayesian inference, mixture of experts

TL;DR: We propose MHC-DC, a brain-inspired few-shot learning framework that simulates human relational thinking through Bayesian multi-view graphs and dynamic knowledge integration, achieving state-of-the-art performance with interpretable reasoning.

Abstract: The fundamental challenge of few-shot image classification stems from inadequate distributional representations due to limited training samples. This paper presents a Human Cognitive-Inspired Bayesian Distribution Calibration method (HC-BDC), inspired by human fast and slow thinking and the neurocognitive mechanisms of convergent and divergent thinking. Unlike conventional approaches, our framework implements a dual-phase reasoning mechanism: the fast-thinking phase employs a lightweight Mixture-of-Experts model to dynamically allocate existing knowledge for different few-shot tasks, while the slow-thinking phase utilizes Bayesian relational inference to simulate human convergent and divergent thinking. This approach diversely generates associations between novel concepts and prior knowledge from multiple perspectives, leading to more comprehensive distribution representations. Specifically, the fast-thinking process automatically selects relevant knowledge components through attention routing, whereas the slow-thinking process constructs multi-view relational graphs via Bayesian inference to dynamically capture diverse inter-class relationships. Extensive experiments on multiple benchmark datasets demonstrate that our approach outperforms current state-of-the-art methods. The HC-BDC framework provides a novel direction for interpretable few-shot learning by modeling the interaction between unconscious association and conscious reasoning processes.

Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)

Submission Number: 14931