Learning From Large-Scale Noisy Web Data With Ubiquitous Reweighting for Image Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning advances often stem from image classification on large datasets.
• Web data is frequently noisy with inaccurate tags, making dataset construction difficult and costly.

Purpose of the Study:

• To propose a novel network, the Ubiquitous Reweighting Network (URNet), capable of learning image classification models from noisy web data.
• To address key challenges in web data: imbalanced classes, high intra-class diversity, inter-class similarity, imprecise and insufficient instances, and ambiguous labels.

Main Methods:

• URNet reweights training instances to mitigate data bias and noise.
• Instance influence is adjusted based on class size, cluster size, confidence, bag size, and labels.
• This gradual alleviation of noise and bias leads to performance improvements.

Main Results:

• URNet demonstrated superior performance on the WebVision 2018 challenge.
• The approach achieved first place in the image classification task using 16 million noisy images across 5000 classes.
• Outperformed existing state-of-the-art models in handling large-scale, noisy web data.

Conclusions:

• URNet effectively learns robust image classification models from challenging, noisy web-scale datasets.
• The proposed reweighting strategy successfully tackles common issues in real-world image data.
• This method offers a significant advancement for deep learning on uncurated image collections.