[Paper Review] Exploiting Unsupervised Inputs for Accurate Few-Shot Classification

Yuqing Hu, Vincent Gripon+1 more|arXiv (Cornell University)|Feb 21, 2020

Domain Adaptation and Few-Shot Learning22 citations

TL;DR

This paper proposes a graph-based feature interpolation method for few-shot image classification that leverages pre-trained features and transductive information from unlabeled samples without training additional parameters. By constructing a similarity graph from feature embeddings and applying power iterations to diffuse features across neighbors, the method achieves state-of-the-art accuracy on miniImageNet and CUB benchmarks, outperforming GNN-based approaches while maintaining simplicity and efficiency.

ABSTRACT

In few-shot classification, the aim is to learn models able to discriminate classes using only a small number of labeled examples. In this context, works have proposed to introduce Graph Neural Networks (GNNs) aiming at exploiting the information contained in other samples treated concurrently, what is commonly referred to as the transductive setting in the literature. These GNNs are trained all together with a backbone feature extractor. In this paper, we propose a new method that relies on graphs only to interpolate feature vectors instead, resulting in a transductive learning setting with no additional parameters to train. Our proposed method thus exploits two levels of information: a) transfer features obtained on generic datasets, b) transductive information obtained from other samples to be classified. Using standard few-shot vision classification datasets, we demonstrate its ability to bring significant gains compared to other works.

Motivation & Objective

• Address the challenge of few-shot image classification with limited labeled data by exploiting transductive information from unlabeled samples.
• Overcome the limitations of existing GNN-based methods that introduce additional trainable parameters, increasing model complexity.
• Develop a simple yet effective method that combines transfer learning with graph-based feature propagation to improve generalization in low-data regimes.
• Demonstrate that graph-based interpolation of pre-trained features can significantly enhance few-shot classification performance.
• Analyze the impact of key hyperparameters and class imbalance on model robustness and performance.

Proposed method

• Pre-train a deep neural network backbone on large-scale generic datasets to extract transferable features.
• Construct a similarity graph where nodes represent samples and edge weights are based on cosine similarity between feature vectors.
• Apply a diffusion process using the matrix $ (\alpha I + E)^\kappa $, where $ E $ is the graph adjacency matrix, to interpolate features across neighboring samples.
• Use the interpolated features to train a simple logistic regression classifier on labeled data only.
• Perform inference on unlabeled samples using the trained classifier, leveraging the transductive inductive bias from the graph.
• Optimize hyperparameters $ \kappa $, $ \alpha $, and $ k $ (number of neighbors) to balance smoothing and representation fidelity.

Experimental results

Research questions

• RQ1 Can graph-based feature interpolation improve few-shot classification accuracy without introducing additional trainable parameters?
• RQ2 How does the choice of hyperparameters $ \kappa $, $ \alpha $, and $ k $ affect the performance and robustness of the method?
• RQ3 To what extent does class imbalance in the unlabeled set degrade the performance of the proposed transductive method?
• RQ4 How does the proposed method compare to state-of-the-art GNN-based few-shot learning approaches in terms of accuracy and efficiency?
• RQ5 Does the graph structure effectively capture meaningful relationships between samples, even when features are not perfectly aligned with class boundaries?

Key findings

• The proposed method achieves state-of-the-art performance on miniImageNet and CUB benchmarks, surpassing existing GNN-based approaches in few-shot classification accuracy.
• Best performance is achieved with small $ k $ (e.g., $ k=5 $ or $ k=10 $), indicating that noisy or distant similarities harm performance.
• The method is robust to hyperparameter choices when $ k $ is small, with minimal sensitivity to $ \alpha $ and $ \kappa $ for $ \kappa \leq 3 $.
• Over-smoothing occurs when $ \alpha $ is small and $ k $ is large, especially when the graph is dense, degrading performance.
• Class imbalance in the unlabeled set significantly reduces accuracy, with the worst performance observed when only one sample is available per class.
• Visualization shows that some classes are well-separated in the graph, while others are entangled, suggesting that backbone features may contain task-irrelevant variance.