Residual Feature-Reutilization Inception Network

2024-08-01

Pattern Recognition

AI Summary

To address feature redundancy and gradient degradation in deep convolutional networks, this paper proposes an enhanced Inception architecture integrating residual connections with multi-scale feature reuse. The core innovation is a dynamic cross-layer feature reutilization mechanism: within each Inception branch, historical features are iteratively activated and reweighted, guided by channel-wise attention and a gradient-gated unit to improve representational continuity and parameter efficiency. Experiments on ImageNet demonstrate that the proposed method achieves a 1.3% absolute improvement in Top-1 accuracy, reduces model parameters by 18%, accelerates inference speed by 22%, and significantly enhances generalization under few-shot learning settings.

Technology Category

• Machine Learning: Deep Neural Architectures and Foundation Models
• Computer Vision: Learning & Optimization for CV
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Responsible Web: Human-perceived consequences of algorithmic deployment on the web

Problem

• Deep Learning
• Image Recognition
• Complex Feature Handling

Innovation

• ResFRI
• Split-ResFRI
• Multi-perspective Feature Extraction