7  Modern Convnets

This chapter studies the architectures and training practices that established convolutional networks in computer vision. Since roughly 2021, vision transformers have led many large-scale image classification benchmarks, beginning with Dosovitskiy et al. (2021) and the Swin Transformer (Liu et al. 2021); we cover them in Chapter 11. Convnets remain important where latency is constrained, training data is limited, or the prediction is dense, as in segmentation (Long et al. 2015) and object detection (Redmon and Farhadi 2018). With modern training procedures and matched compute, they can also match transformers on image classification (Smith et al. 2023).

The first group of sections covers the architectural developments of 2012–2015. AlexNet (Krizhevsky et al. 2012) established deep convolutional networks on ImageNet (Section 7.1). VGG, NiN, and GoogLeNet then introduced repeated blocks, \(1 \times 1\) channel mixing, global pooling, and multi-branch design (Section 7.2). Batch normalization (Ioffe and Szegedy 2015) improved optimization (Section 7.3), while residual connections (He et al. 2016) made substantially deeper networks practical (Section 7.4). Normalization and residual connections subsequently became standard components outside computer vision, including in transformers.

The later sections separate architectural choices from training and deployment constraints. Efficient networks use depthwise convolution and structural re-parameterization (Section 7.5). Modern training procedures raise the accuracy of an unchanged ResNet-50 by more than four percentage points (Wightman et al. 2021) (Section 7.6). ConvNeXt applies these practices in a controlled modernization of ResNet (Liu et al. 2022) (Section 7.7). Finally, RegNet studies distributions over network designs rather than selecting a single architecture (Radosavovic et al. 2020) (Section 7.8).

The sections are arranged roughly chronologically because each design responds to a limitation of its predecessors. This also makes it possible to distinguish improvements due to architecture from those due to optimization, data augmentation, and computational budget.

Resources and Further Reading

The references below trace the architectural developments covered in this chapter and current practice: landmark networks from AlexNet to ConvNeXt, the training recipes that confound naive comparisons between them, and efficient networks for deployment. All are freely accessible online except where noted.

Books

Courses and video lectures

Tutorials, notes, and surveys

Foundational papers

All free on arXiv or the proceedings site; these are the primary sources this chapter retells, worth reading in the original: