xception

Xception (Chollet, 2017, the author of Keras) is a reinterpretation of the Inception module that pushes its underlying hypothesis to the limit, hence the name 'Extreme Inception'. The Inception module implicitly assumes that cross-channel correlations and spatial correlations in feature maps can be mapped somewhat separately: its 1x1 convolutions first mix channels, then its 3x3 and 5x5 branches handle spatial structure. Xception asks: what if we assume these two kinds of correlation are completely decoupled and can be handled in fully separate steps? Following that assumption to its conclusion turns the Inception module into a depthwise separable convolution.

Concretely, Xception is a stack of depthwise separable convolutions: a 1x1 pointwise convolution that handles all cross-channel mixing, followed by a depthwise convolution that handles spatial filtering independently per channel. This is almost the same primitive MobileNet uses, with two minor ordering and detail differences (Xception applies the 1x1 first and typically places no nonlinearity between the two operations). The entire network is built from such blocks, organised into entry, middle, and exit flows, with residual connections wrapped around most blocks to aid optimisation, giving roughly the same parameter count as Inception-v3 but better accuracy.

The significance of Xception is conceptual clarity: it frames the depthwise separable convolution not as a mere efficiency hack but as a principled architectural hypothesis about the structure of image features, namely that learning channel mixing and spatial mixing separately is not just cheaper but often better. By outperforming Inception-v3 on ImageNet at a similar size, it validated that hypothesis and helped cement depthwise separable convolutions as a default building block. Xception thus sits at the conceptual junction between the Inception line (multi-scale, factorised modules) and the MobileNet/EfficientNet line (efficient separable convolutions).