AresB-Net: accurate residual binarized neural networks using shortcut concatenation and shuffled grouped convolution

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Binarized neural networks (BNNs) offer computational efficiency but suffer from degraded accuracy compared to real-valued networks.
• Residual convolutional neural networks (CNNs) improve accuracy but challenges remain in BNNs.
• Existing BNNs struggle with maintaining high classification accuracy, especially with increasing feature resolution.

Purpose of the Study:

• To propose a novel network model, AresB-Net, for achieving higher accuracy in residual binarized CNNs.
• To address the accuracy degradation issue in BNNs while maintaining computational efficiency.
• To develop a network structure that mitigates binarization errors.

Main Methods:

• Introduced novel basic blocks within the AresB-Net architecture to amortize binarization errors.
• Implemented a well-balanced pyramid structure without downsampling convolution.
• Utilized max-pooling for low-cost shortcuts during downsampling (stride > 1).
• Incorporated shortcut connections added to convolution output, followed by concatenation and channel shuffling for grouped convolutions.

Main Results:

• AresB-Net demonstrates improved classification accuracies on benchmark datasets like CIFAR and ImageNet.
• The proposed structure effectively facilitates feature reuse, alleviating binarization errors.
• Achieved reduced computational costs and small weight storage requirements.
• The model shows remarkable classification performance despite low hardware costs associated with binarized computations.

Conclusions:

• AresB-Net offers a promising approach to enhance the accuracy of residual binarized CNNs.
• The novel architecture effectively balances accuracy and computational efficiency.
• The findings suggest AresB-Net is suitable for applications requiring high performance with limited hardware resources.