Inception-v4

Inception-v4 is the purely non-residual member of the 2016 Inception generation. Where Inception v3 had grown piecemeal — each stage partitioned differently so the model could be trained across many machines under memory limits — Inception-v4 re-draws the whole network as a uniform design: one fixed stem, three kinds of Inception module repeated in blocks, and two dedicated reduction modules between them. Every module is a set of parallel branches whose outputs are concatenated along the channel axis,

where each branch $\mathcal{B}_i$ is a short chain of convolutions (or a pooling followed by a $1\times1$ projection), each followed by batch normalisation and a ReLU.

The layout is Stem → 4× Inception-A ($35\times35\times384$) → Reduction-A → 7× Inception-B ($17\times17\times1024$) → Reduction-B → 3× Inception-C ($8\times8\times1536$) → global average pool → dropout (keep 0.8) → linear classifier. The stem itself is multi-branch: three filter-concatenation junctions take a $299\times299$ image to a $35\times35\times384$ grid, and every spatial reduction in it runs a stride-2 convolution and a stride-2 max pool side by side instead of choosing one. Large kernels are factorised throughout — an $n\times n$ convolution becomes a $1\times n$ followed by an $n\times1$, which covers the same receptive field for $2n$ rather than $n^2$ weights per input–output channel pair.

The paper's point in building Inception-v4 alongside Inception-ResNet was a controlled comparison: with a comparable compute budget, residual connections speed training up markedly but are not what makes a very deep Inception network accurate — Inception-v4 reaches roughly the accuracy of Inception-ResNet-v2 (single-crop top-5 error 5.0% on the ImageNet validation set, against 4.9%) with no shortcuts at all.