Gated Channel Transformation for Visual Recognition

1. Introduction

• GCT combines normalization methods and gating mechanisms to create both competition among channels, which increases activation variance, and cooperation among channels, which decreases activation variance.Source: Figure 1 · p. 1
• SE’s fully connected layers limit module deployment across CNN layers and make learned channel interactions difficult to interpret [17].SE modules were applied at the block level, with reduced FC dimensions to save computation, but the FC layers still hindered wider deployment.Source: §1 · p. 1
• GCT replaces fully connected channel modeling with simple ℓ2 normalization and lightweight trainable channel-wise parameters that control gated adaptation.The channel-wise gating weights are intended to make channel relationships easier to analyze across a CNN.Source: §1 · p. 2
• GCT is reported to improve deep CNN generalization across visual recognition tasks and datasets.The paper describes the architecture as a simple and effective way to model relationships among channels.Source: §1 · p. 2

Gating and attention mechanisms

• LSTM uses input, output, and forget gates to regulate information flow, while attention methods [27] direct computation toward informative feature components.Attention has also improved sequence learning [4], lip reading [7], image captioning [42], and image localization and understanding [20].Source: §2 · p. 2
• Compared with the SE module, GCT models cross-channel relationships and is described as achieving better performance gains with less computation and fewer parameters.Source: §2 · p. 2

Normalization layers

• GCT uses channel-related information in a normalization structure, alongside normalization methods such as LRN, BN [19], LN [3], and GN.LRN computes channel-neighborhood statistics per pixel; BN uses batch-dimension spatial information, LN computes along channels, and GN normalizes within channel groups.Source: §2 · p. 2

Deep architectures

• VGGNets [31] and Inception networks [33] improved representation through depth, while ResNets [13] used identity skip connections to build deeper networks.Other work explores computational-element representation [34], multi-branch convolutions or pooling, and grouped convolutions to increase transformation cardinality [41].Source: §2 · p. 2
• For video classification, C3D [36] uses 3D convolutions, while NL-Nets [39] capture long-range non-local dependencies and improve accuracy.Source: §2 · p. 2

• GCT-equipped networks are reported to achieve promising performance improvements with negligible growth in computational complexity.Source: §2 · pp. 2–3

3. Gated Channel Transformation

• GCT combines global context embedding, parameter-free channel normalization, and channel-wise gating to model channel relationships with lightweight, interpretable parameters.The trainable gate parameters can be visualized to explain channel behavior.Source: §3 · p. 3

3.2. Channel Normalization

• GCT’s channel-wise parameters give it O(C) complexity, compared with O(C2) for SE’s two fully connected layers [17].The embedding weight α controls each channel’s contribution to normalization; a value near zero excludes that channel.Source: §3 · p. 3, §3.1 · p. 4

3.1. Global Context Embedding

• Using an ℓp-norm instead of SE’s global average pooling avoids constant pooled outputs after Instance Normalization; ℓ2 is the default, while ℓ1 performs very similarly.When features are consistently non-negative, ℓ1 can be replaced by global average pooling and is more computationally efficient.Source: §3.1 · pp. 3–4

3.3. Gating Adaptation

• The gating factor 1 + tanh(γcŝc + βc) adapts each channel and supports competition when γc is positive or cooperation when it is negative.Unlike LRN, which supports competition only, GCT can model both relationship types.Source: §3.3 · p. 4

3.4. Learning

• GCT can be placed before convolution, before normalization, or after normalization; all are effective, but placement before convolution works better.For training, α is initialized to 1 and γ and β to 0; a 0.01 learning-rate warmup for 1 epoch is used on ImageNet, and β receives no weight decay.Source: §3.4 · pp. 4–5

4.1. ImageNet

• GCT improves ImageNet results across ResNet depths, including a 21.4% top-1 error for GCT-ResNet-101 versus 21.6% for the deeper ResNet-152 baseline.The reported top-1 improvements are 1.1% for ResNet-50, 0.8% for ResNet-152, and 1.0% for ResNet-200.Source: §4.1 · p. 5

4.2. COCO

• On COCO, GCT improves Mask R-CNN detection and segmentation, and BN∗+GCT exceeds GN results in ResNet-101.With the short schedule, ResNet-50 gains 2.0 box AP and 1.8 mask AP, versus SE [17] gains of 0.4 box AP and 0.5 mask AP; ResNet-101 gains 1.9 detection AP and 1.6 segmentation AP. Against GN, ResNet-101 BN∗+GCT improves detection AP by 0.8 and segmentation AP by 1.1; its short-schedule segmentation AP of 37.7 exceeds the long-schedule GN result of 37.2.Source: Table 4 · p. 6, §4.2 · pp. 6–7

4.3. Kinetics

• On Kinetics-400, GCT counterparts consistently improve recognition accuracy over ResNet-50 (3D) and ResNet-101 (3D) baselines, despite limited-memory placement.GCT was applied only in the last two convolutional layers of each Res-Block; the authors note that applying it to all convolutional layers might improve performance further.Source: §4.3 · p. 7

4.4. Analysis

• GCT tends to reduce feature variance in layers far from the output and increase it in layers near the output, supporting cooperation and competition among channels, respectively.The reported gating-weight means tend to be below zero in earlier layers and above zero closer to the output; the measured output-to-input variance ratio is below 1 in earlier layers and above 1 near the output.Source: §4.4 · pp. 7–8

4.5. Ablation Studies

• Ablations find ℓ2 normalization slightly better than ℓ1, residual gating adaptations better than Sigmoid, and GCT before all convolutional layers the best tested placement.Mean-and-variance normalization yields 23.7% top-1 error versus the 23.8% ResNet-50 baseline; among tested adaptations, 1 + tanh gives slightly better improvement, while 1 + tanh and 1+ELU [8] outperform Sigmoid.Source: §4.5 · p. 8

Conclusion and Future Work

• GCT combines normalization and gating to facilitate channel competition and cooperation with negligible parameter complexity.The layer leverages relationships among channels to improve deep CNN discriminability.Source: §5 · p. 8
• Experiments show GCT's effectiveness and robustness across a wide range of modern CNNs and datasets.Source: §5 · p. 8
• Future work will study applying GCT to recurrent networks, including LSTM.Source: §5 · p. 8