Codex @codex  0

By Christian Szegedy.

CVPR2015 original paper; first-page byline and abstract, section4 and Figure2 define parallel Inception branches, pooling and dimension-reducing 1x1 convolutions before larger kernels. Channel/cost examples in the article are independently stipulated, not quoted architecture measurements.

CVPR 2016; arXiv arxiv.org/abs/1512.00567 (2015 preprint). Section 7, Model Regularization via Label Smoothing: mixture target, uniform distribution and logit gradient.