[Paper Review] Rethinking the Inception Architecture for Computer Vision

Christian Szegedy, Vincent Vanhoucke+3 more|arXiv (Cornell University)|Dec 2, 2015

Advanced Neural Network Applications564 citations

TL;DR

This paper introduces Inception-v3, a refined convolutional neural network architecture that improves accuracy through factorized convolutions, aggressive dimensionality reduction, batch normalization, and label smoothing. It achieves state-of-the-art performance with 21.2% top-1 and 5.6% top-5 error on ImageNet ILSVRC 2012 using only 5 billion multiply-adds and under 25 million parameters, significantly outperforming prior models in computational efficiency.

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 explore 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 on the validation set (3.6% error on the test set) and 17.3% top-1 error on the validation set.

Motivation & Objective

• To design a more computationally efficient and accurate convolutional neural network architecture for computer vision tasks.
• To address the inefficiency of naive network scaling, which increases parameters and FLOPs quadratically.
• To improve performance on ImageNet classification while maintaining low computational cost and parameter count.
• To provide a clearer, more systematic design framework for Inception-style networks beyond GoogLeNet.
• To demonstrate that aggressive regularization and architectural innovations can yield high accuracy even with modest compute budgets.

Proposed method

• Factorizing large convolutional filters (e.g., 7×7) into smaller, sequential 3×3 convolutions to reduce FLOPs and parameters.
• Applying batch normalization to all convolutional layers, including auxiliary classifiers, to stabilize training and improve generalization.
• Using label smoothing during training to reduce overconfidence and improve robustness and generalization.
• Implementing aggressive dimensionality reduction via 1×1 convolutions before large filters to limit computational cost.
• Introducing an auxiliary classifier with batch normalization to provide stronger gradients during early training.
• Employing multi-crop evaluation and model ensembling to further boost performance on the ILSVRC 2012 validation set.

Experimental results

Research questions

• RQ1 Can factorized convolutions and dimensionality reduction significantly reduce FLOPs and parameters without sacrificing accuracy?
• RQ2 How do batch normalization and label smoothing affect generalization and training stability in deep Inception networks?
• RQ3 To what extent can architectural innovations improve top-1 and top-5 accuracy while keeping computational cost low?
• RQ4 Can aggressive architectural scaling be achieved without a quadratic increase in FLOPs or parameters?
• RQ5 How does multi-crop evaluation and model ensembling improve performance on the ILSVRC 2012 benchmark?

Key findings

• Inception-v3 achieved 21.2% top-1 and 5.6% top-5 error on the ILSVRC 2012 validation set with single-frame evaluation, setting a new SOTA.
• With a 4-model ensemble and multi-crop evaluation, the model reached 3.5% top-5 error, a 25% relative reduction over the best published ensemble result.
• The model used only 5 billion multiply-adds per inference and fewer than 25 million parameters, significantly outperforming deeper, denser networks like those in He et al. [6].
• The use of factorized 7×7 convolutions reduced FLOPs while maintaining accuracy, demonstrating the effectiveness of architectural decomposition.
• Label smoothing and batch-normalized auxiliary classifiers contributed to improved generalization, reducing overfitting and boosting performance.
• The model achieved high performance even with a receptive field as small as 79×79, enabling better detection of small objects.