An Intelligent Multi-View Active Learning Method Based on a Double-Branch Network

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) are vital deep learning models requiring extensive labeled data for optimal performance.
• Manual data labeling is labor-intensive and time-consuming, posing a significant bottleneck in CNN development.
• Active learning strategies aim to minimize labeling effort by intelligently selecting data points for annotation.

Purpose of the Study:

• To propose a novel intelligent active learning method, Multi-View Active Learning based on Double-Branch Network (MALDB), for deep learning.
• To enhance the efficiency of training Convolutional Neural Networks (CNNs) by reducing the need for large labeled datasets.
• To improve classifier performance through iterative expansion of the training dataset using strategically selected unlabeled samples.

Main Methods:

• MALDB integrates two Bayesian Convolutional Neural Networks (BCNNs) with distinct structures as dual branches within a classifier.
• The method analyzes unlabeled datasets, querying informative samples based on the divergent characteristics learned by the two BCNN branches.
• It leverages multi-level feature information from various hidden layers of the BCNNs to ensure stable sample selection.

Main Results:

• Experimental validation was performed on five diverse datasets: Fashion-MNIST, Cifar-10, SVHN, Scene-15, and UIUC-Sports.
• The proposed MALDB method demonstrated significant effectiveness in improving classifier performance.
• Results confirmed the validity and efficiency of the MALDB approach in active learning for deep learning.

Conclusions:

• The Multi-View Active Learning based on Double-Branch Network (MALDB) effectively reduces the dependency on large labeled datasets for training deep learning models.
• MALDB enhances classifier performance and stability by intelligently querying and incorporating informative samples.
• This approach offers a promising solution for efficient model training in domains where data labeling is a constraint.