Abstract
The versatility to learn from a handful of samples is the hallmark of human intelligence. Few-shot learning is an endeavour to transcend this capability down to machines. Inspired by the promise and power of probabilistic deep learning, we propose a novel variational inference network for few-shot classification (coined as TRIDENT) to decouple the representation of an image into semantic and label latent variables, and simultaneously infer them in an intertwined fashion. To induce task-awareness, as part of the inference mechanics of TRIDENT, we exploit information across both query and support images of a few-shot task using a novel built-in attention-based transductive feature extraction module (we call AttFEX). Our extensive experimental results corroborate the efficacy of TRIDENT and demonstrate that, using the simplest of backbones, it sets a new state-of-the-art in the most commonly adopted datasets miniImageNet and tieredImageNet (offering up to 4% and 5% improvements, respectively), as well as for the recent challenging cross-domain miniImagenet → CUB scenario offering a significant margin (up to 20% improvement) beyond the best existing cross-domain baselines.
Introduction
Few-shot learning (FSL) represents a critical frontier in artificial intelligence, aiming to bridge the gap between machine learning and human-like sample efficiency. Traditional FSL methods, predominantly based on metric learning, encounter significant obstacles: sample bias and feature entanglement. Sample bias occurs when the limited number of available samples fails to represent the true class distribution, while feature entanglement refers to the difficulty in separating label-relevant features from task-irrelevant semantic noise (e.g., background, lighting).
Methodology: The TRIDENT Framework
We propose TRIDENT (Transductive Relational Inference for DEcoupled Network Training) to address these limitations through a principled probabilistic approach. TRIDENT is built on three innovative pillars:
1. Decoupled Latent Variables
Using Variational Inference (VI), TRIDENT explicitly decouples an image's representation into two latent variables: $z_l$ (label-specific features) and $z_s$ (semantic/contextual features). The classification head operates solely on $z_l$, ensuring that the model learns to ignore distracting semantic information. Simultaneously, both variables are used to reconstruct the input image, ensuring that the decoupling process does not lose essential information.
2. Attention-based Feature Extraction (AttFEX)
To mitigate sample bias, we introduce AttFEX, a transductive module that processes the entire task (all support and query images) as a single unit. By employing a self-attention mechanism, AttFEX allows the model to identify which features are most discriminative for the specific set of classes in the current task, effectively making the feature extraction process task-aware.
3. Meta-Learning Optimization
The framework is optimized using a meta-learning strategy, enabling the network to adapt its parameters rapidly to novel, unseen tasks during inference.
Experimental Results
TRIDENT was rigorously evaluated on the standard benchmarks miniImageNet and tieredImageNet, as well as the challenging miniImageNet → CUB cross-domain scenario.
State-of-the-Art Performance
As shown in the results table below, TRIDENT consistently outperforms existing methods, often by significant margins. Remarkably, it achieves these results using a simple Conv4 backbone, surpassing competitors that utilize much deeper architectures like ResNet-12 or Wide ResNet (WRN).
Key Findings
- Cross-Domain Excellence: The most striking result is in the cross-domain setting, where TRIDENT achieves an accuracy of 84.61% (1-shot), representing a ~20% improvement over the best previous baselines. This highlights the robustness of the decoupled representation.
- Latent Space Separability: Qualitative analysis via UMAP visualizations demonstrates that the label space ($z_l$) becomes highly organized and separable after meta-adaptation, whereas the semantic space ($z_s$) remains unclustered, confirming the efficacy of the decoupling mechanism.
- Model Calibration: Beyond raw accuracy, TRIDENT exhibits superior calibration, meaning its predicted probabilities more accurately reflect the true likelihood of correctness, a vital trait for reliable real-world deployment.
Conclusion and Future Work
TRIDENT demonstrates that a transductive, decoupled approach to variational inference can significantly enhance few-shot classification performance and generalization. By focusing on the architecture of the inference process rather than just model depth, we achieve state-of-the-art results even with minimal backbones. Future research will explore the application of TRIDENT to semi-supervised and unsupervised few-shot learning tasks.
Summary updated