Gated Convolutional Networks with Hybrid Connectivity for Image Classification

Chuanguang Yang

Introduction

Deep learning architectures for image classification have long grappled with the balance between feature exploration and feature reuse. ResNet popularized residual connections to facilitate feature reuse and gradient flow, while DenseNet introduced dense connectivity to maximize feature exploration. However, DenseNet's quadratic growth in connections leads to significant parameter redundancy. This paper introduces HCGNet (Hybrid Connectivity Gated Network), a novel architecture that harmonizes these two paradigms through a gated hybrid connectivity pattern.

Hybrid Connectivity Architecture

The fundamental building block of HCGNet is the Hybrid Block. Unlike standard DenseNets that rely solely on concatenation, HCGNet employs a nested aggregation strategy. Globally, modules within a block are connected via dense concatenation (dense connectivity), ensuring robust feature exploration. Locally, within each module, residual additions (residual connectivity) are used to maintain efficient feature flow and reuse.

The Hybrid Block architecture combining global dense connectivity with local residual additions.

This hybrid approach allows the network to capture the representational power of dense connections while mitigating the explosion of redundant parameters. By structuring the network this way, HCGNet achieves superior performance with significantly fewer modules per block compared to traditional DenseNets.

The SMG Module: Squeeze, Multi-scale, and Gated

At the heart of the Hybrid Block is the SMG Module, which consists of three primary components designed for efficient feature extraction and management:

  1. Squeeze Cell: This component utilizes 1x1 and 3x3 group convolutions to condense incoming feature maps, reducing computational overhead.
  2. Multi-scale Excitation Cell: To capture diverse spatial information, this cell employs parallel 3x3 and 5x5 depthwise convolutions. This "wider" local structure allows the model to learn features at multiple scales without the depth-related parameter cost.
  3. Gated Mechanisms: The module incorporates an Update Gate and a Forget Gate. The Update Gate uses spatial and channel attention to intelligently fuse multi-scale features. Conversely, the Forget Gate selectively decays redundant information from the residual path, mimicking biological neural filtering.

Notably, the authors utilize the tanh activation function within these gates. They argue that tanh is better suited for preserving information in low-dimensional spaces, and the hybrid connectivity effectively prevents the vanishing gradient issues typically associated with non-ReLU activations.

Experimental Evaluation

HCGNet was rigorously tested across multiple benchmarks, demonstrating state-of-the-art efficiency:

  • CIFAR-10/100: HCGNet-A3 outperformed NASNet-A while utilizing 4.5x fewer parameters, highlighting its architectural efficiency.
  • ImageNet: HCGNet-C achieved a 19.5% Top-1 error rate with 42.2M parameters. In comparison, ResNet-101 requires 44.6M parameters to reach a significantly higher error rate of 23.4%.
  • Transfer Learning: When integrated into the Mask R-CNN framework for MS-COCO, HCGNet backbones consistently outperformed ResNet-50 in both object detection and instance segmentation tasks.

Robustness and Interpretability

Beyond raw accuracy, HCGNet exhibits enhanced adversarial robustness. In tests against FGSM attacks, it maintained higher accuracy than both ResNet and DenseNet, likely due to the smoother information flow provided by the gated hybrid connections. Furthermore, network dissection analysis revealed that HCGNet learns more distinct and high-level visual detectors (e.g., specific textures and object parts) than traditional backbones, suggesting a more meaningful internal representation of visual data.

Conclusion

HCGNet represents a significant step forward in efficient CNN design. By combining dense and residual connectivity through intelligent gating mechanisms, it effectively addresses the redundancy issues of DenseNet while surpassing the representational limits of ResNet. Its performance across classification, detection, and segmentation tasks makes it a highly versatile backbone for modern computer vision.

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, ie, 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},
  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