[Paper Review] Learning from Web Data with Memory Module

Yi Tu, Li Niu+3 more|arXiv (Cornell University)|Jun 28, 2019

Image Retrieval and Classification Techniques2 citations

TL;DR

This paper proposes a memory-augmented multi-instance learning framework that jointly addresses label and background noise in web-crawled images without requiring clean supervision. By grouping region proposals into bags and using a learnable memory module to assign dynamic weights based on cluster discriminativeness, the method achieves end-to-end training and outperforms existing approaches on four benchmark datasets.

ABSTRACT

Learning from web data has attracted lots of research interest in recent years. However, crawled web images usually have two types of noises, label noise and background noise, which induce extra difficulties in utilizing them effectively. Most existing methods either rely on human supervision or ignore the background noise. In this paper, we propose a novel method, which is capable of handling these two types of noises together, without the supervision of clean images in the training stage. Particularly, we formulate our method under the framework of multi-instance learning by grouping ROIs (i.e., images and their region proposals) from the same category into bags. ROIs in each bag are assigned with different weights based on the representative/discriminative scores of their nearest clusters, in which the clusters and their scores are obtained via our designed memory module. Our memory module could be naturally integrated with the classification module, leading to an end-to-end trainable system. Extensive experiments on four benchmark datasets demonstrate the effectiveness of our method.

Motivation & Objective

• To address the dual challenge of label noise and background noise in web-crawled images, which hinder effective self-supervised learning.
• To develop a method that does not require human-annotated clean images during training.
• To enable end-to-end learning by integrating a memory module with a classification head.
• To improve model robustness and accuracy on noisy web data through representative region weighting.

Proposed method

• The method groups images and their region proposals (ROIs) from the same category into bags, following a multi-instance learning paradigm.
• A memory module learns representative clusters of ROIs and assigns each ROI a weight based on the discriminativeness score of its nearest cluster.
• The memory module is differentiable and jointly trained with the classification module, enabling end-to-end optimization.
• Cluster scores are updated dynamically during training using a key-value memory mechanism that stores and retrieves feature representations.
• ROIs are weighted according to their proximity to high-scoring clusters, emphasizing more representative and discriminative regions.
• The framework is trained end-to-end without requiring clean image supervision, relying solely on noisy web data.

Experimental results

Research questions

• RQ1 Can a self-supervised method effectively handle both label and background noise in web-crawled images without clean supervision?
• RQ2 How can representative and discriminative ROIs be automatically identified and weighted in a noisy web image setting?
• RQ3 Can a memory module be effectively integrated into a multi-instance learning framework to improve robustness and performance?
• RQ4 What is the impact of dynamic ROI weighting based on cluster discriminativeness on classification accuracy?

Key findings

• The proposed method achieves state-of-the-art performance on four benchmark datasets despite training on noisy web data without clean supervision.
• The memory module significantly improves model robustness by focusing on more representative and discriminative ROIs during training.
• The end-to-end trainable architecture enables consistent performance gains across diverse web image datasets.
• The ablation study confirms that both label noise and background noise handling contribute to the overall performance improvement.