Rethinking the Inception Architecture for Computer Vision

This paper revisits the Inception architecture for convolutional networks in computer vision, motivated by the observation that while increasing model size and compute tends to yield immediate quality gains given enough labeled data, computational efficiency and low parameter counts remain important for use cases like mobile vision and big-data scenarios. The core method explores ways to scale up networks that use the added computation as efficiently as possible, achieved through suitably factorized convolutions combined with aggressive regularization techniques.

Benchmarked on the ILSVRC 2012 classification validation set, the approach demonstrates substantial gains over the prior state of the art, reaching 21.2% top-1 and 5.6% top-5 error for single-frame evaluation using a network with 5 billion multiply-adds per inference and fewer than 25 million parameters; an ensemble of 4 models with multi-crop evaluation further reduces error to 3.5% top-5 and 17.3% top-1 on validation and 3.6% top-5 on the official test set. This mattered because it showed strong accuracy could be achieved with efficient, low-parameter architectures suitable for resource-constrained deployment, not just by brute-force scaling.