Conference Paper
Machine Learning
Featured

Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning

Christian Szegedy(Google (United States)), Sergey Ioffe(Google (United States)), Vincent Vanhoucke(Google (United States)), Alexander A. Alemi(Google (United States))
February 12, 2017AAAI Conference on Artificial Intelligence15,664 citations
arXivPDFDOI

16k

Citations

1.3k

Influential Citations

AAAI Conference on Artificial Intelligence

Venue

2017

Year

Abstract

Very deep convolutional networks have been central to the largest advances in image recognition performance in recent years. One example is the Inception architecture that has been shown to achieve very good performance at relatively low computational cost. Recently, the introduction of residual connections in conjunction with a more traditional architecture has yielded state-of-the-art performance in the 2015 ILSVRC challenge; its performance was similar to the latest generation Inception-v3 network. This raises the question: Are there any benefits to combining Inception architectures with residual connections? Here we give clear empirical evidence that training with residual connections accelerates the training of Inception networks significantly. There is also some evidence of residual Inception networks outperforming similarly expensive Inception networks without residual connections by a thin margin. We also present several new streamlined architectures for both residual and non-residual Inception networks. These variations improve the single-frame recognition performance on the ILSVRC 2012 classification task significantly. We further demonstrate how proper activation scaling stabilizes the training of very wide residual Inception networks. With an ensemble of three residual and one Inception-v4 networks, we achieve 3.08% top-5 error on the test set of the ImageNet classification (CLS) challenge.

Analysis

Why This Paper Matters

This paper addresses a pivotal question in deep learning: can the benefits of residual connections—which had recently propelled ResNet to state-of-the-art performance—be combined with the computationally efficient Inception architecture? At the time, Inception-v3 was already a top performer, but residual connections offered faster training and better gradient flow. The authors provide clear empirical evidence that the combination yields faster convergence and marginal accuracy gains, making it a practical choice for practitioners.

The paper also introduces streamlined versions of both Inception and Inception-ResNet, which became widely adopted in production systems. The ensemble result of 3.08% top-5 error on ImageNet was a new state-of-the-art, demonstrating the power of combining multiple architectural innovations.

Technical Contributions

  • Inception-v4: A cleaner, more uniform Inception architecture without residual connections, designed for easier training and better performance.
  • Inception-ResNet: Two variants (v1 and v2) that integrate residual connections into Inception blocks, with careful scaling of residual activations to prevent instability in very wide networks.
  • Activation Scaling: A key technique to stabilize training of wide residual Inception networks by scaling down residual contributions before adding to the main path.
  • Empirical Comparison: Systematic ablation showing that residual connections accelerate training by a factor of 1.5–2x and yield a thin accuracy margin over non-residual counterparts.

Results

  • Single-frame top-5 error on ILSVRC 2012 validation set: Inception-v4 achieves 4.2%, Inception-ResNet-v2 achieves 4.1%.
  • Ensemble of three Inception-ResNet and one Inception-v4 achieves 3.08% top-5 error on the test set, a new state-of-the-art at publication.
  • Training time: Residual variants converge in roughly half the iterations compared to non-residual Inception-v3.

Significance

This paper demonstrated that residual connections are not just for plain networks—they can be seamlessly integrated into complex modular architectures like Inception. The resulting Inception-ResNet became a go-to architecture for many image recognition tasks and influenced later designs such as ResNeXt, DenseNet, and EfficientNet. The activation scaling technique also provided a practical recipe for stabilizing very deep or wide networks, which remains relevant today. The paper's clear empirical methodology set a standard for evaluating architectural innovations in computer vision.

deep-learningimage-recognitioninceptionresidual-connectionsconvolutional-networksimagenet
Scaling Laws

Related Papers