We propose a deep convolutional neural network architecture codenamed "Inception", which was responsible for setting the new state of the art for classification and detection in the ImageNet Large-Scale Visual Recognition Challenge 2014 (ILSVRC 2014). The main hallmark of this architecture is the improved utilization of the computing resources inside the network. This was achieved by a carefully crafted design that allows for increasing the depth and width of the network while keeping the computational budget constant. To optimize quality, the architectural decisions were based on the Hebbian principle and the intuition of multi-scale processing. One particular incarnation used in our submission for ILSVRC 2014 is called GoogLeNet, a 22 layers deep network, the quality of which is assessed in the context of classification and detection.
核心贡献 · Key contributions
提出 Inception 架构，通过增加深度和宽度同时保持计算预算不变，提高了计算资源利用率。 Proposed the Inception architecture that improves computing resource utilization by increasing depth and width while keeping computational budget constant.
引入使用 1x1 卷积进行降维，消除计算瓶颈。 Introduced dimension reduction using 1x1 convolutions to remove computational bottlenecks.
使用 GoogLeNet 在 ILSVRC 2014 分类和检测任务上取得最先进结果。 Achieved state-of-the-art results on ILSVRC 2014 classification and detection with GoogLeNet.
证明用密集组件近似最优稀疏结构是改进神经网络的有效方法。 Demonstrated that approximating optimal sparse structure with dense components is viable for improving neural networks.
展示了通过高效设计增加网络深度和宽度可以在不按比例增加计算成本的情况下提高准确率。 Showed that increasing network depth and width with efficient design can improve accuracy without proportional computational cost.
使用辅助分类器促进较低阶段的判别性并提供正则化。 Used auxiliary classifiers to encourage discrimination in lower stages and provide regularization.
局限 · Limitations
Inception 架构专为计算机视觉任务设计，可能无法泛化到其他领域。 The Inception architecture was designed specifically for computer vision tasks and may not generalize to other domains.
该架构依赖手动设计和调优；自动构建类似拓扑仍是未来工作。 The architecture relies on manual design and tuning; automated construction of similar topologies remains future work.
计算效率的提升是以增加架构复杂性为代价的。 The computational efficiency gains are achieved at the cost of increased architectural complexity.
结果基于特定竞赛数据集（ILSVRC 2014），可能不完全代表真实世界性能。 The results are based on a specific competition dataset (ILSVRC 2014) and may not fully represent real-world performance.
检测流程未使用边界框回归或上下文，这些可能进一步提高准确率。 The detection pipeline did not use bounding box regression or context, which could further improve accuracy.
论文章节 · Sections（共 11）
摘要Abstract
1 引言1 Introduction
2 相关工作2 Related Work
3 动机与高层考量3 Motivation and High Level Considerations