Methods for Visualizing Convolutional Neural Networks

Abstract

The method for visualizing the convolutional neural network of the present invention includes: 1) preparing a data set; 2) customizing the convolutional neural network input layer, and setting the convolutional neural network feature extraction function parameters to generate a convolutional neural network feature extraction program; 3. ) Execute the convolutional neural network feature extraction program, extract the response features of all neurons in the specified layer of all pictures in the data set, and save them; 4) Calculate the response field parameters of the neurons in the specified layer; 5) Visualize and update the maximum response of a single neuron in the specified layer Visualization of large response features. The method for visualizing the convolutional neural network of the present invention can well display the working characteristics of neurons, and is helpful for understanding how the convolutional neural network realizes its superiority.

Description

technical field [0001] The invention relates to image processing technology, in particular to a method for visualizing a convolutional neural network. Background technique [0002] Since the convolutional network (convnets) was first introduced by LeCun.Y et al. in "Backpropagation applied to handwritten zip code recognition" (Neural Comput.1989) in 1989, the convolutional network has achieved good results in the field of image processing. In the ILSVRC (ImageNet Large Scale Visual Recognition Challenge) competition in 2012, the 8-layer convolutional neural network (Convolutional Neural Network) designed by Alex Krizhevsky and others with 5 convolutional layers and 3 fully connected layers achieved 16.4% top -5errorrate, almost half of the 26.1% top-5 error rate of the second place. The entire field of image processing research has also begun to migrate from convex optimization to research on neural networks. Now, the convolutional neural network has become one of the rese...

AI Technical Summary

Problems solved by technology

[0004] In the prior art, research on the display of the trained convolutional neural network is generally limited to the display of some convolution parameters (convolution kernel Kernel) and the display of the final recognition results

However, the display of simple parameters does not give a good idea of ​​what kind of image input the convolutional layer will have a greater response to.

Images