2:40 PM - 3:00 PM
[2Q3-J-2-05] Visualization of learning process for Convolution Neural Network
Keywords:CNN, visualization, learning process, Grad-CAM
Convolutional Neural Network (CNN) is an image classifier using deep neural network. However, it hardly gives the evidence why it classifies an image into a class.
To solve this problem, some methods producing visual explanations has been proposed.
Grad-CAM produces visual information for localized important regions for a class in an input image.
As well as the visual explanations of classification, it is important to visualize of the learning process.
The performance of CNN, such as accurate classification, is highly rely on the parameters. We convince the visualization of the learning process helps the parameter tuning.
We propose a method that visualize the learning process. It generates the visual explanation images of arbitrary classes for each epoch.
We validated the effectiveness of our method using MNIST dataset. The result shows the proposed method can visualize the learning process for every class for every epoch, whereas the usual method cannot.
To solve this problem, some methods producing visual explanations has been proposed.
Grad-CAM produces visual information for localized important regions for a class in an input image.
As well as the visual explanations of classification, it is important to visualize of the learning process.
The performance of CNN, such as accurate classification, is highly rely on the parameters. We convince the visualization of the learning process helps the parameter tuning.
We propose a method that visualize the learning process. It generates the visual explanation images of arbitrary classes for each epoch.
We validated the effectiveness of our method using MNIST dataset. The result shows the proposed method can visualize the learning process for every class for every epoch, whereas the usual method cannot.