Visualizing and Comparing Convolutional Neural Networks

By Wei Yu, Kuiyuan Yang, Yalong Bai, Hongxun Yao, Yong Rui. 2014

54 citations recorded by the source corpus.

Convolutional Neural Networks (CNNs) have achieved comparable error rates to well-trained human on ILSVRC2014 image classification task. To achieve better performance, the complexity of CNNs is continually increasing with deeper and bigger architectures. Though CNNs achieved promising external classification behavior, understanding of their internal work mechanism is still limited. In this work, we attempt to understand the internal work mechanism of CNNs by probing the internal representations in two comprehensive aspects, i.e., visualizing patches in the representation spaces constructed by different layers, and visualizing visual information kept in each layer. We further compare CNNs with different depths and show the advantages brought by deeper architecture.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

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

Citations in the corpus

Explaining nonlinear classification decisions with deep Taylor decomposition (2015, neutral): “The observation that more accurate predictions are supported by better resolved input patterns is also in line with other studies [51, 52].”
How Deep is the Feature Analysis underlying Rapid Visual Categorization? (2016, background): “[21], who have shown that both sparsity of the representation and distinctiveness of the matched features increase monotonously with the depth of the network.”
Deep learning for visual understanding: A review (2016, methodology): “Similar visualization was also studied by Yu et al. [141].”
Visual Analytics in Deep Learning: An Interrogative Survey for the Next Frontiers (2018, unknown): “Although the image size must be greatly reduced to prevent excessive overlap, viewing all the images at once can provide insight into what a deep learning model has learned, as seen in this example where the authors visualize ImageNet test data [76], or in this example where the authors create many ”
Understanding convolutional neural networks with a mathematical model (2016, background): “Keywords: Convolutional Neural Network (CNN), Nonlinear Activation, RECOS Model, Rectified Linear Unit (ReLU), MNIST Dataset.”
Identifying Individual Facial Expressions by Deconstructing a Neural Network (2016, methodology): “Explanation methods such as [3,26,22,12,13,7,27,25,19] have been able to successfully solve this important open problem.”
Deep Convolutional Neural Networks for Image Classification: A Comprehensive Review (2017, background): “…5.4.2), Yu, Wang, Chen, and Wei (2014) introduced a novel mixed pooling technique to further boost the regularization abilities of DCNNs and address the known issues associated with average and max pooling (Zeiler & Fergus, 2013; D. Yu et al., 2014; Sainath, Kingsbury, Mohamed et al., 2013).”