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.
核心贡献 · Key contributions
提出了高效扩展 CNN 的设计原则，包括避免表征瓶颈和平衡宽度与深度。 Proposes design principles for scaling CNNs efficiently, including avoiding representational bottlenecks and balancing width/depth.
引入将大卷积分解为小卷积（如 5x5 分解为两个 3x3）的方法，在保持表达能力的同时降低计算量。 Introduces factorization of large convolutions into smaller ones (e.g., 5x5 into two 3x3) to reduce computation without loss of expressiveness.
开发了非对称卷积分解（如 1xn 后接 nx1），在中尺寸特征图上进一步节省计算。 Develops asymmetric convolution factorization (e.g., 1xn followed by nx1) for further computational savings on medium grid sizes.
证明辅助分类器主要起正则化作用而非改善梯度流动，尤其在批归一化后。 Demonstrates that auxiliary classifiers act as regularizers rather than improving gradient flow, especially with batch normalization.
提出标签平滑正则化，通过软化真实标签防止过拟合并提升泛化能力。 Proposes label-smoothing regularization to prevent overfitting and improve generalization by softening ground-truth labels.
在 ILSVRC 2012 上以 Inception-v3 达到 21.2% top-1 和 5.6% top-5 错误率，同时保持计算高效。 Achieves state-of-the-art results on ILSVRC 2012 with 21.2% top-1 and 5.6% top-5 error using Inception-v3, while being computationally efficient.
局限 · Limitations
设计原则是推测性的，需要在不同任务上进一步实验验证。 Design principles are speculative and require further experimental validation across diverse tasks.
分解技术在前几层效果不佳，仅在中尺寸特征图上有效。 Factorization techniques may not work well on early layers; effectiveness is limited to medium grid sizes.
标签平滑引入超参数 epsilon，可能需要针对不同数据集调整。 Label-smoothing introduces a hyperparameter epsilon that may need tuning for different datasets.
研究聚焦于图像分类，对其他视觉任务（如检测或分割）的适用性未充分探索。 The study focuses on image classification; applicability to other vision tasks like detection or segmentation is not fully explored.
计算成本节省依赖于特定架构选择；简单缩放仍可能导致效率低下。 Computational cost savings assume specific architectural choices; naive scaling may still lead to inefficiencies.
论文章节 · Sections（共 14）
摘要Abstract
1 引言1 Introduction
2 通用设计原则2 General Design Principles
3 大卷积核的分解3 Factorizing Convolutions with Large Filter Size
3.1 分解为小卷积3.1 Factorization into smaller convolutions
3.2 空间分解为非对称卷积3.2 Spatial Factorization into Asymmetric Convolutions
4 辅助分类器的效用4 Utility of Auxiliary Classifiers
5 高效的网格尺寸缩减5 Efficient Grid Size Reduction
6 Inception-v26 Inception-v2
7 通过标签平滑进行模型正则化7 Model Regularization via Label Smoothing
8 训练方法8 Training Methodology
9 低分辨率输入的性能9 Performance on Lower Resolution Input