Inception

Inception v3 is the third generation of the Inception family and crystallises a set of design principles that the authors extracted from earlier experiments — most notably that one should avoid representational bottlenecks (do not aggressively shrink spatial resolution in early layers) and should factorise large convolutions to gain both parameter efficiency and additional nonlinearity.

The headline factorisation idea is to replace a single $n\times n$ convolution with an $n\times1$ followed by a $1\times n$ convolution. For $n=7$ this trades $49 C^2$ parameters for $14 C^2$ while adding an intermediate ReLU and matching the original receptive field exactly along a separable manifold. Inception v3 uses three different block topologies (Inception-A, B, C) at three spatial resolutions (35×35, 17×17, 8×8), each tailored to the receptive- field budget at that stage of the network.

Two further innovations debuted in this paper are label smoothing, which replaces the one-hot target with a softer distribution $(1 - \epsilon)\,\mathbf{1}_y + \epsilon/K$ to discourage overconfident predictions, and RMSProp with a carefully tuned learning-rate schedule for ImageNet-scale training. Inception v3 reaches a top-5 error of 3.5% on ImageNet validation with only 24 M parameters, illustrating that thoughtful architectural surgery can outperform brute scaling.