Inception-ResNet

Inception-ResNet hybridises the two dominant architectural ideas of the early ResNet era: the multi-branch Inception module and the residual shortcut. Each Inception sub-network is wrapped in an identity skip connection, so the block computes

where $\alpha$ is a small fixed scale factor (typically 0.10–0.30) applied to the Inception branch before addition. This residual scaling trick was introduced specifically for Inception-ResNet because the unscaled residual branches occasionally caused training to diverge on networks with many filters per Inception module.

Three block families — Block35, Block17, Block8 — mirror Inception v3's A/B/C topologies at the 35×35, 17×17, and 8×8 resolutions respectively, but each is now a residual unit. The network's stem is shared with Inception v4 (an aggressive multi-branch downsampler). Reduction-A and Reduction-B blocks perform spatial reduction between groups of residual Inception blocks while doubling the channel budget.

The empirical message of the paper is precise and influential: residual connections do not improve final accuracy beyond what a comparably-sized non-residual Inception v4 achieves, but they dramatically accelerate convergence — Inception-ResNet v2 reaches the same accuracy roughly twice as fast. Final top-5 error on ImageNet is 3.08%, with the residual variants holding a convergence-speed edge over their non-residual siblings throughout training.