Rethinking the Inception Architecture for Computer Vision

TLDR

Convolutional networks underpin most state‑of‑the‑art computer vision, and since 2014 very deep models have dominated benchmarks, yet computational efficiency and low parameter counts remain critical for mobile and large‑scale applications. This work investigates scaling up networks by employing factorized convolutions and aggressive regularization to maximize the utility of added computation. The authors design factorized convolutional architectures with strong regularization to build deeper, more efficient models. On ILSVRC 2012, a single‑frame model achieves 21.2 % top‑1 and 5.6 % top‑5 error with 5 B MACs and <25 M parameters, while a 4‑model ensemble with multi‑crop reduces top‑5 error to 3.5 % and top‑1 to 17.3 % on validation and 3.6 % top‑5 on the test set.

Abstract

Convolutional networks are at the core of most state of-the-art computer vision solutions for a wide variety of tasks. Since 2014 very deep convolutional networks started to become mainstream, yielding substantial gains in various benchmarks. Although increased model size and computational cost tend to translate to immediate quality gains for most tasks (as long as enough labeled data is provided for training), computational efficiency and low parameter count are still enabling factors for various use cases such as mobile vision and big-data scenarios. Here we are exploring ways to scale up networks in ways that aim at utilizing the added computation as efficiently as possible by suitably factorized convolutions and aggressive regularization. We benchmark our methods on the ILSVRC 2012 classification challenge validation set demonstrate substantial gains over the state of the art: 21:2% top-1 and 5:6% top-5 error for single frame evaluation using a network with a computational cost of 5 billion multiply-adds per inference and with using less than 25 million parameters. With an ensemble of 4 models and multi-crop evaluation, we report 3:5% top-5 error and 17:3% top-1 error on the validation set and 3:6% top-5 error on the official test set.