Hierarchical neural network with efficient selection inference.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) achieve high image classification precision with increasing complexity.
• Uneven visual separability between categories poses challenges for standard CNNs.
• Existing CNNs often overlook the inherent hierarchical structure of data categories.

Purpose of the Study:

• To propose a novel hierarchical network model that integrates category hierarchies with ResNet-style modules.
• To enhance feature extraction and computational efficiency in image classification.
• To address the limitations of fixed-layer feed-forward computation in current CNNs.

Main Methods:

• A top-down hierarchical network is constructed by integrating category hierarchies with ResNet modules.
• Residual block selection based on coarse categories is employed to create adaptive computation paths.
• Each residual block acts as a switch, enabling 'JUMP' or 'JOIN' modes for specific categories.

Main Results:

• The proposed hierarchical network achieves higher prediction accuracy compared to original residual networks.
• The model demonstrates similar or improved computational efficiency (FLOPs) due to layer skipping for certain categories.
• Experiments on CIFAR-10, CIFAR-100, SVHM, and Tiny-ImageNet datasets validate the effectiveness of the approach.

Conclusions:

• Hierarchical network models offer advantages over standard CNNs by exploiting data structure.
• Selective residual block usage can significantly improve both accuracy and inference speed.
• This approach provides a promising direction for more efficient and accurate image classification systems.