Inception Network
A CNN architecture using parallel convolutions of different sizes within a single layer (inception module).
Overview
The Inception network (GoogLeNet), introduced by Google in 2014, uses a novel inception module that applies multiple convolution filter sizes (1x1, 3x3, 5x5) and pooling operations in parallel, then concatenates their outputs. This allows the network to capture features at multiple scales simultaneously.
Key Details
The architecture is much more parameter-efficient than VGGNet, achieving better accuracy with fewer parameters. Later versions (Inception v2, v3, v4) added batch normalization, factorized convolutions, and residual connections. The inception module concept influenced many subsequent architectures that process information at multiple scales.
Related Concepts
cnn • vgg network • resnet