BiSPL: Bidirectional Self-Paced Learning for Recognition From Web Data.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning (DL) models require extensive labeled data, which is costly and time-consuming to acquire.
• Utilizing readily available web data is a promising approach to mitigate data scarcity in DL.
• Directly training DL models on raw web data is hindered by inherent data noise and mixed quality.

Purpose of the Study:

• To develop a novel framework for effectively utilizing noisy web data in deep learning.
• To address the data scarcity problem, particularly in fine-grained visual classification (FGVC) tasks.
• To improve the stability and performance of deep models trained with web-sourced data.

Main Methods:

• A bidirectional self-paced learning (BiSPL) framework was developed.
• The framework iteratively samples web data based on its distance to labeled source samples.
• It trains deep models by initially including both easy and hard samples, then progressively dropping hard samples to reduce noise.

Main Results:

• The BiSPL framework effectively reduces the impact of noisy data from web sources.
• Experiments on six FGVC tasks demonstrated superior performance compared to state-of-the-art methods.
• The method achieved high, stable performance even with significantly reduced labeled training set sizes.

Conclusions:

• The proposed BiSPL framework offers a robust solution for training deep learning models using noisy web data.
• This approach is particularly beneficial for data-limited domains like fine-grained visual classification.
• BiSPL enables effective model training and improved performance by intelligently selecting and refining training data.