SG-CNN

SG-CNN implements a self-grouping convolutional neural network approach that dynamically groups convolutional filters by importance vectors to reduce parameters and accelerate CNN inference while preserving recognition accuracy for bioinformatics tasks such as large-scale genomic data analysis and protein structure prediction.

Key Features:

• Data-Dependent Grouping: Dynamically groups convolutional layer filters based on similarity of their importance vectors by evaluating each filter's input-channel importance values.
• Importance Vector Clustering: Clusters importance vectors to determine group centroids and uses these clusters to form diverse group convolution filters and prune less critical links.
• Pruning: Removes less significant connections identified from importance vectors to minimize accuracy loss during parameter reduction.
• Fine-Tuning Schemes: Restores recognition capacity after pruning using local and global fine-tuning, or global-only fine-tuning, which yield comparable experimental results.
• Adaptability and Performance: Integrates with architectures such as ResNet and DenseNet and demonstrates improved compression ratio, speedup, and recognition accuracy on CIFAR-10/100 and ImageNet.
• Generalization Capabilities: Supports transfer learning applications including domain adaptation and object detection with strong generalization performance.

Scientific Applications:

• Genomic Data Analysis: Enables computationally efficient CNN models for large-scale genomic datasets while maintaining model accuracy.
• Protein Structure Prediction: Provides parameter-reduced, accelerated CNN inference suitable for deep-learning–based protein structure tasks.
• Transfer Learning and Domain Adaptation: Applies to transfer learning scenarios, including domain adaptation and object detection in biological imaging contexts.

Methodology:

Evaluate input-channel importance values per filter, cluster importance vectors to determine group centroids, prune less critical connections to form group convolutions, and restore performance via local and global or global-only fine-tuning; experiments were conducted on ResNet and DenseNet using CIFAR-10/100 and ImageNet.