# Transductive Few-Shot Learning: Clustering is All You Need?

**Authors:** Imtiaz Masud Ziko, Malik Boudiaf, Jose Dolz, Eric Granger, Ismail Ben Ayed  
**arXiv:** [2106.09516](https://arxiv.org/abs/2106.09516)  
**Date:** 2021-06-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2106.09516v1/

## Key points

- Investigates a general formulation for clustering and transductive few-shot learning
- Proposes a method that integrates prototype-based objectives, Laplacian regularization, and supervision constraints from a few labeled data points
- Introduces a concave-convex relaxation of the problem and derives a computationally efficient block-coordinate bound optimizer with convergence guarantee
- Allows for easy distribution in large-scale clustering and few-shot tasks by computing independent updates for each point-to-cluster assignment
- Provides a thorough convergence analysis based on point-to-set maps
- Evaluates the method through comprehensive clustering and few-shot learning experiments on various datasets
- Shows competitive performances in terms of accuracy and optimization quality, even when scaling up to large problems
- Finds that even standard clustering procedures like K-means can achieve competitive performances compared to state-of-the-art few-shot learning methods
- Highlights limitations of current few-shot benchmarks and questions the viability of complex meta-learning and episodic training techniques in recent literature
- Expands significantly on preliminary clustering results published at the NeurIPS conference, providing new formulations, convergence analysis, experiments, comparisons, and discussions

## Summary

The paper investigates a general formulation for clustering and transductive few-shot learning. It proposes a method that integrates prototype-based objectives, Laplacian regularization, and supervision constraints from a few labeled data points. The authors introduce a concave-convex relaxation of the problem and derive a computationally efficient block-coordinate bound optimizer with convergence guarantee. The proposed optimizer computes independent updates for each point-to-cluster assignment, allowing for easy distribution in large-scale clustering and few-shot tasks. The authors provide a thorough convergence analysis based on point-to-set maps. The method is evaluated through comprehensive clustering and few-shot learning experiments on various datasets. The results show that the proposed approach achieves competitive performances in terms of accuracy and optimization quality, even when scaling up to large problems. Surprisingly, the authors find that even standard clustering procedures like K-means can achieve competitive performances compared to state-of-the-art few-shot learning methods. The paper highlights the limitations of current few-shot benchmarks and questions the viability of complex meta-learning and episodic training techniques in recent literature. Moreover, it expands significantly on their preliminary clustering results published at the NeurIPS conference, providing new formulations, convergence analysis, experiments, comparisons, and discussions. Overall, the paper presents a novel approach for integrating clustering and transductive few shot learning, demonstrating its effectiveness through extensive experiments.