Gated Convolutional Networks with Hybrid Connectivity for Image Classification

Chuanguang Yang, Zhulin An, Hui Zhu, Xiaolong Hu, Kun Zhang, Kaiqiang Xu, Chao Li, Yongjun Xu

Abstract

HCGNet (Hybrid Convolutional Gated Network) addresses the architectural redundancy found in modern deep learning models by integrating the global feature exploration of DenseNet with the local efficiency of ResNet. Central to this architecture is the Squeeze Multi-scale Gated (SMG) module, which utilizes "Forget" and "Update" gates to actively manage feature flow and prevent representation collapse. Evaluations on CIFAR and ImageNet benchmarks show that HCGNet achieves superior performance compared to models like NASNet and ResNeXt while utilizing significantly fewer parameters. Furthermore, HCGNet demonstrates enhanced robustness to adversarial attacks and improved feature interpretability.

Introduction

Deep learning has been dominated by two primary architectural paradigms: ResNet, which employs additive residual connections for efficient feature reuse, and DenseNet, which uses concatenations to explore new feature combinations. While DenseNet is powerful, its quadratic growth in connections leads to significant redundancy and "superfluous features." Conversely, ResNet can suffer from "representation collapse" when excessive feature addition dilutes information. HCGNet seeks a "Goldilocks" zone, combining the strengths of both while mitigating their respective weaknesses through active feature management.

Architectural Design: Hybrid Connectivity and the SMG Module

The HCGNet architecture is built upon two foundational concepts: Hybrid Connectivity and the Squeeze Multi-scale Gated (SMG) module.

Hybrid Connectivity

HCGNet employs a dual-connectivity strategy. It utilizes global dense connectivity (concatenation) to link different modules, ensuring that early-stage features remain accessible throughout the network. Within each module, however, it applies local residual connectivity (addition) to refine features efficiently without the memory overhead of continuous concatenation.

Hybrid Block architecture

The SMG Module

The SMG module is the primary processing unit of HCGNet, designed to reduce redundancy and enhance feature diversity. It consists of two main cells and two gates:

  • Squeeze Cell: Compresses dense features using 1x1 and 3x3 group convolutions.
  • Multi-scale Excitation Cell: Processes features through parallel 3x3 and 5x5 depthwise convolutions to capture multi-scale information simultaneously.
  • Update Gate: An attention mechanism using tanh activations to dynamically fuse multi-scale features based on spatial and channel importance.
  • Forget Gate: A gating mechanism on the residual shortcut that learns to suppress redundant or noisy historical features before they are combined with new information, acting similarly to a biological synapse.
SMG Module and Gating mechanisms

Experimental Results

HCGNet was evaluated across several standard benchmarks, including CIFAR-10/100, ImageNet (ILSVRC 2012), and MS-COCO for transfer learning.

  • Efficiency and Performance: On CIFAR, HCGNet-A3 outperformed the auto-searched NASNet-A with 4.5x fewer parameters. On ImageNet, HCGNet-C achieved a Top-1 error of 19.5%, surpassing larger models like ResNeXt-101 and DPN-98.
  • Redundancy Reduction: HCGNet-B matched or exceeded the accuracy of DenseNet-169 while reducing the number of stacked modules by 88%.
  • Robustness: The gated architecture provides inherent resistance to adversarial attacks (e.g., FGSM) compared to standard ResNet and DenseNet models, likely due to its active feature filtering.
  • Transferability: As a backbone for Mask R-CNN on MS-COCO, HCGNet-B demonstrated superior performance in object detection and instance segmentation tasks over ResNet-50 and AOGNet.

Key Insights: Active Feature Management and Interpretability

HCGNet shifts the architectural paradigm from passive feature stacking to active feature management. Network Dissection analysis reveals that HCGNet naturally learns a higher number of unique, high-level semantic detectors (e.g., specific object parts and textures) than its predecessors. This suggests that its internal representations are cleaner and more diverse, avoiding the redundancy common in purely dense architectures.

Conclusion

HCGNet demonstrates that architectural redundancy is not a prerequisite for high performance. By intelligently combining dense and residual connectivity with active gating mechanisms, HCGNet provides a highly efficient, robust, and interpretable framework for image classification and beyond. Future work may involve integrating these manual design principles into Neural Architecture Search (NAS) to discover even more optimized configurations.

Summary updated

Abstract

We propose a simple yet effective method to reduce the redundancy of DenseNet by substantially decreasing the number of stacked modules by replacing the original bottleneck by our SMG module, which is augmented by local residual. Furthermore, SMG module is equipped with an efficient two-stage pipeline, which aims to DenseNet-like architectures that need to integrate all previous outputs, i.e., squeezing the incoming informative but redundant features gradually by hierarchical convolutions as a hourglass shape and then exciting it by multi-kernel depthwise convolutions, the output of which would be compact and hold more informative multi-scale features. We further develop a forget and an update gate by introducing the popular attention modules to implement the effective fusion instead of a simple addition between reused and new features. Due to the Hybrid Connectivity (nested combination of global dense and local residual) and Gated mechanisms, we called our network as the HCGNet. Experimental results on CIFAR and ImageNet datasets show that HCGNet is more prominently efficient than DenseNet, and can also significantly outperform state-of-the-art networks with less complexity. Moreover, HCGNet also shows the remarkable interpretability and robustness by network dissection and adversarial defense, respectively. On MS-COCO, HCGNet can consistently learn better features than popular backbones.
BibTeX
@inproceedings{yang2020gated,
  author = {Chuanguang Yang and Zhulin An and Hui Zhu and Xiaolong Hu and Kun Zhang and Kaiqiang Xu and Chao Li and Yongjun Xu},
  title = {{Gated Convolutional Networks with Hybrid Connectivity for Image Classification}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i07.6948},
}
Figure