The central building block of convolutional neural networks (CNNs) is the convolution operator, which enables networks to construct informative features by fusing both spatial and channel-wise information within local receptive fields at each layer. We propose the Squeeze-and-Excitation (SE) block, an architectural unit designed to improve the representational power of a network by enabling it to perform dynamic channel-wise feature recalibration. SE networks formed the foundation of our ILSVRC 2017 classification submission which won first place and reduced the top-5 error to 2.251%.

CNNs have proven to be effective models for image understanding. At each convolutional layer, a set of filters express local spatial connectivity patterns along input channels. However, the channel relationships are implicitly modeled by the filters and are entangled with spatial information. We argue that explicitly modeling the interdependencies between channels of convolutional features can significantly improve feature quality. To achieve this, we propose an SE block that operates in two phases: squeeze (aggregating spatial information) and excitation (learning channel-wise dependencies).

Prior work has explored deeper architectures (VGGNet, ResNet, Inception) and architectural search methods (NAS) to improve representations. Attention mechanisms have been applied in various forms: spatial attention focuses on informative spatial regions, while channel attention has been less explored. The SE block can be viewed as a lightweight gating mechanism applied to channels, related to but distinct from highway networks and gated recurrent units. The key difference is that SE blocks operate as self-attention on channel features, conditioned on the input.

The squeeze operation aggregates feature maps across spatial dimensions to produce a channel descriptor. This is achieved through global average pooling, generating a vector z of dimension C where each element z_c is the spatial average of the c-th channel. The rationale is that each channel of the output after convolution is unable to exploit contextual information outside of its local receptive field. The squeeze operation addresses this by embedding global spatial information into the channel descriptor.

The excitation operation uses the channel descriptor to learn channel-wise dependencies. It employs a bottleneck architecture with two fully-connected layers: first a dimensionality reduction layer with ratio r (default r=16) followed by ReLU, then a dimensionality increasing layer followed by sigmoid activation. The output is a set of per-channel modulation weights that are applied to the original feature maps via channel-wise multiplication. This mechanism introduces dynamics conditioned on the input, which can be regarded as a self-attention function on channels.

ImageNet classification: SE-ResNet-50 achieves 6.62% top-5 error versus ResNet-50's 7.48% — a substantial improvement. More remarkably, SE-ResNet-101 (6.07% error) outperforms the deeper ResNet-152 (6.34% error), demonstrating that SE blocks provide more effective improvement than simply adding more layers. The computational overhead is minimal: only a 0.26% increase in GFLOPs for SE-ResNet-50. SE blocks generalize across architectures: they improve VGGNet, Inception, ResNeXt, and MobileNet consistently.

Beyond classification: On COCO object detection with Faster R-CNN, SE blocks provide a 2.4% AP improvement over the ResNet-50 baseline. On Places365 scene classification, SE-ResNet-152 surpasses the previous state-of-the-art. On CIFAR-10 and CIFAR-100, SE blocks improve ResNet-110, ResNet-164, and WideResNet across the board. These results confirm that the benefits of channel recalibration extend well beyond ImageNet classification to diverse tasks and datasets.

Comprehensive ablation studies reveal several design insights. Reduction ratio r=16 provides the optimal accuracy-complexity tradeoff. Global average pooling outperforms max pooling for the squeeze operation. For the excitation nonlinearity, sigmoid is essential — replacing it with ReLU causes significant performance degradation, since ReLU's zero-clipping prevents the suppression of uninformative channels. A role analysis of SE activations shows that early layers exhibit class-agnostic behavior (similar weights across categories), while deeper layers become increasingly class-specific, dynamically selecting the most relevant channels for each input.

We have proposed the Squeeze-and-Excitation block, a lightweight architectural unit that enhances the representational capacity of CNNs by explicitly modeling channel interdependencies. Through the squeeze operation (global information embedding) and the excitation operation (adaptive recalibration), SE blocks learn to dynamically emphasize informative features and suppress less useful ones. With minimal computational overhead (~0.26% GFLOPs increase), SE blocks deliver consistent improvements across multiple architectures, datasets, and tasks, including the winning entry of ILSVRC 2017 with 2.251% top-5 error.