Squeeze-and-Excitation (SE) es una técnica arquitectónica avanzada utilizada en redes neuronales convolucionales (CNNs) to improve their ability to represent complex features in data. The primary goal of SE is to adaptively recalibrate the feature maps produced by a CNN enfocándose en características útiles mientras suprime las menos informativas.
The technique consists of two main steps: ‘squeeze’ and ‘excitation’. In the ‘squeeze’ step, global information is captured by applying agrupamiento promedio global to the feature maps, creating a channel descriptor that summarizes the presence of features across the spatial dimensions. This results in a compact representation that reflects the importance of each channel.
In the ‘excitation’ step, the channel descriptor is passed through a series of fully connected layers and funciones de activación, typically using a sigmoid activation to yield a set of weights that indicate the significance of each channel. These weights are then used to scale the original feature maps, effectively enhancing the response of important features while diminishing the influence of less pertinent ones.
The SE block can be integrated into various neural network architectures, including ResNet and Inception, and has been shown to improve performance on numerous tasks, including clasificación de imágenes and object detection. By recalibrating the feature responses, Squeeze-and-Excitation allows the network to focus on the most relevant information, leading to more accurate predictions and improved model robustness.