A storage-efficient ensemble classification using filter sharing on binarized convolutional neural networks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Binary Neural Networks (BNNs) offer computational efficiency but suffer from low inference accuracy.
• Ensemble methods can improve accuracy by aggregating multiple classifiers, but increase memory demands.
• Lightweight systems face storage limitations for complex models.

Purpose of the Study:

• To propose a storage-efficient ensemble classification method for Binary Neural Networks (BNNs).
• To overcome the low inference accuracy of individual BNNs while managing memory constraints.
• To enhance classification performance in dynamic powered, lightweight systems.

Main Methods:

• Developed a novel ensemble classification scheme for BNNs.
• Implemented filter sharing from a trained Convolutional Neural Network (CNN) model to reduce storage.
• Retrained the model by training only unfrozen learnable parameters.
• Evaluated performance across various ensemble types and BNN structures on CIFAR datasets.

Main Results:

• The proposed filter-sharing method is scalable with the number of classifiers.
• Significant enhancement in classification accuracy was observed.
• Achieved 56.74% (ResNet-20) and 70.29% (ReActNet-10) Top-1 accuracies on CIFAR-100 with 10 BNN classifiers.
• Performance improved by 7.6% and 3.6% compared to single BNN classifiers.

Conclusions:

• The filter-sharing ensemble classification is effective for improving BNN accuracy.
• The method successfully addresses the storage burden of ensemble methods in BNNs.
• Demonstrated scalability and significant accuracy gains on benchmark datasets.