PolyNet: A Pursuit of Structural Diversity in Very Deep Networks

By Xingcheng Zhang, Zhizhong Li, Chen Change Loy, Dahua Lin. 2016

279 citations recorded by the source corpus.

A number of studies have shown that increasing the depth or width of convolutional networks is a rewarding approach to improve the performance of image recognition. In our study, however, we observed difficulties along both directions. On one hand, the pursuit for very deep networks is met with a diminishing return and increased training difficulty, on the other hand, widening a network would result in a quadratic growth in both computational cost and memory demand. These difficulties motivate us to explore structural diversity in designing deep networks, a new dimension beyond just depth and width. Specifically, we present a new family of modules, namely the PolyInception, which can be flexibly inserted in isolation or in a composition as replacements of different parts of a network. Choosing PolyInception modules with the guidance of architectural efficiency can improve the expressive power while preserving comparable computational cost. The Very Deep PolyNet, designed following this direction, demonstrates substantial improvements over the state-of-the-art on the ILSVRC 2012 benchmark. Compared to Inception-ResNet-v2, it reduces the top-5 validation error on single crops from 4.9% to 4.25%, and that on multi-crops from 3.7% to 3.45%.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Industrial and Manufacturing Engineering, Model Architectures

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Industrial and Manufacturing Engineering

In Model Architectures

Citations in the corpus

EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks (2019, neutral): “0x 32B 18x PolyNet (Zhang et al., 2017) 81.”
Measuring Robustness to Natural Distribution Shifts in Image Classification (2020, neutral): “polynet [116].”
Do ImageNet Classifiers Generalize to ImageNet? (2019, background): “[62] Xingcheng Zhang, Zhizhong Li, Chen Change Loy, and Dahua Lin.”
Progressive Neural Architecture Search (2017, neutral): “ F= 54) 5.1M 588M 74.2 91.9 Table 4. ImageNet classication results in the Mobile setting. Model Image Size Params Mult-Adds Top-1 Top-5 ResNeXt-101 (64x4d) [36] 320 320 83.6M 31.5B 80.9 95.6 PolyNet [38] 331 331 92M 34.7B 81.3 95.8 Dual-Path-Net-131 [6] 320 320 79.5M 32.0B 81.5 95.8 Squeeze-Excite-N”
Regularized Evolution for Image Classifier Architecture Search (2018, neutral): “arked on the same ImageNet dataset version. Model # Parameters # Multiply-Adds Top-1 / Top-5 Accuracy (%) Incep-ResNet V2 [44] 55.8M 13.2B 80.4 / 95.3 ResNeXt-101 [48] 83.6M 31.5B 80.9 / 95.6 PolyNet [51] 92.0M 34.7B 81.3 / 95.8 Dual-Path-Net-131 [7] 79.5M 32.0B 81.5 / 95.8 GeNet-2 [47] 156M – 72.1 ”
Dual Path Networks (2017, background): “Very Deep PolyNet [23], with a much smaller model size (304 MB v.s. 365 MB).”
A Comprehensive Survey of Neural Architecture Search (2020, neutral): “9]. Search mothed Reference Venue Top 1/Top 5 Acc(%) Params (Millions) Image Size (squarred) GPU Days Human Mobilenets [6] CoRR17 70.6/89.5 4.2 224 - ResNeXt [138] CVPR17 80.9/95.6 83.6 320 - Polynet [139] CVPR17 81.3/95.8 92.0 331 - DPN [140] NIPS17 81.5/95.8 79.5 320 - Shufflenet [137] CVPR18 70.9”
A Survey on Neural Architecture Search (2019, neutral): “l. (2018) 61.4/83.7 16.2 32 3 Zhang et al. (2019) 73.0/91.3 6.1 224 0.84 Human Howard et al. (2017) 70.6/89.5 4.2 224 - Zhang et al. (2018b) 70.9/89.8 5.0 224 - Xie et al. (2017) 80.9/95.6 83.6 320 - Zhang et al. (2017) 81.3/95.8 92.0 331 - Chen et al. (2017) 81.5 95.8 79.5 320 - 28 A Survey on Neur”
ResNeSt: Split-Attention Networks (2020, background): “e of 224 when comparing our ResNeSt with ResNet variants, and a crop size of 256 when comparing with other approaches. Regularization. Very deep neural networks tend to overt even for large datasets [68]. To prevent this, dropout regularization randomly masks out some neurons during training (but no”
Beyond Finite Layer Neural Networks: Bridging Deep Architectures and Numerical Differential Equations (2017, methodology): “This explains why PolyNet is able to reduce depth by increasing width of each residual block to achieve state-of-the-art classiﬁcation accuracy.”
Between-Class Learning for Image Classification (2017, background): “Many types of image recognition networks have been proposed mainly in ILSVRC competition [17, 23, 25, 11, 28, 29, 12].”
Image Aesthetic Assessment: An experimental survey (2016, methodology): “ields the best performance in both the balanced accuracy and overall accuracy. It is worth pointing out that other more recent and deeper models such as ResNet [86], Inception-ResNet [87], or PolyNet [88], could serve as a pre-trained model. Nevertheless, owing to the typically small size of aesthet”
Non-Parametric Calibration for Classification (2019, unknown): “The following CNNs were used: AlexNet (Krizhevsky et al., 2012), VGG19 (Simonyan and Zisserman, 2014, Liu and Deng, 2015), ResNet-50, ResNet-152 (He et al., 2016), WideResNet (Zagoruyko and Komodakis, 2016), DenseNet-121, DenseNet-BC-190, DenseNet-201 (Huang et al., 2017), InceptionV4 (Szegedy et al”
AOGNets: Deep AND-OR Grammar Networks for Visual Recognition (2017, background): “More than 100 layers are popular design in the recent literature [14, 36], and even more than 1000 layers trained on large scale datasets such as ImageNet are not rare any more [19, 44].”