Semi-supervised active learning using convolutional auto- encoder and contrastive learning.

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Obtaining labeled data for machine learning is often expensive and time-consuming.
• Active learning aims to optimize data annotation within budget constraints.
• Existing active learning methods often overlook the potential of unlabeled data.

Purpose of the Study:

• To propose a novel pool-based semi-supervised active learning method for image classification.
• To enhance the utilization of both labeled and unlabeled data in the active learning process.
• To improve the efficiency and accuracy of image classification with limited labeled samples.

Main Methods:

• Clustering the latent space of a pre-trained convolutional autoencoder.
• Employing a clustering contrastive loss to refine latent space clustering with minimal labeled data.
• Querying samples with high uncertainty for annotation by an oracle in an iterative process.

Main Results:

• The proposed method demonstrates effectiveness, particularly when the number of annotated samples is small.
• Experiments on benchmark datasets validate the method's efficacy in image classification.
• Empirical results show significant improvements in accuracy terms.

Conclusions:

• The developed semi-supervised active learning approach effectively utilizes unlabeled data to boost image classification performance.
• This method offers a powerful solution for scenarios with limited labeled data budgets.
• The research highlights the value of integrating unlabeled data into active learning strategies for computer vision tasks.