Visualizing convolutional neural networks

Patent metadata

Field	Value
Publication number	US-10192001-B2
Application number	US-201715725026-A
Country	US
Kind code	B2
Filing date	Oct 4, 2017
Priority date	Oct 4, 2016
Publication date	Jan 29, 2019
Grant date	Jan 29, 2019

Abstract

Convolutional neural networks can be visualized. For example, a graphical user interface (GUI) can include a matrix of symbols indicating feature-map values that represent a likelihood of a particular feature being present or absent in an input to a convolutional neural network. The GUI can also include a node-link diagram representing a feed forward neural network that forms part of the convolutional neural network. The node-link diagram can include a first row of symbols representing an input layer to the feed forward neural network, a second row of symbols representing a hidden layer of the feed forward neural network, and a third row of symbols representing an output layer of the feed forward neural network. Lines between the rows of symbols can represent connections between nodes in the input layer, the hidden layer, and the output layer of the feed forward neural network.

First claim

The invention claimed is:

1. A system for visualizing convolutional neural networks, the system comprising:
a processing device; and
a memory device on which instructions executable by the processing device are stored for causing the processing device to:
generate a matrix of symbols to be positioned in a graphical user interface, each symbol in the matrix indicating a feature-map value that represents a likelihood of a particular feature being present or absent at a location in an input to a convolutional neural network, each column in the matrix having feature-map values generated by convolving the input to the convolutional neural network with a respective filter for identifying a specific feature in the input;
generate a node-link diagram to be positioned in the graphical user interface, wherein the node-link diagram represents a feed forward neural network that forms part of the convolutional neural network and comprises:
a first row of symbols representing an input layer to the feed forward neural network, wherein the input layer is also a maxpooling layer of the convolutional neural network, the first row of symbols is color coded to represent activation values for nodes in the input layer, and each symbol in the first row of symbols is vertically aligned with a respective column in the matrix of symbols and indicates a maximum value in the respective column;
a second row of symbols representing a hidden layer of the feed forward neural network, the second row of symbols being color coded to represent activation values of nodes in the hidden layer;
a third row of symbols representing an output layer of the feed forward neural network, the third row of symbols being color coded to represent activation values of nodes in the output layer; and
lines between the first row of symbols, the second row of symbols, and the third row of symbols, the lines representing connections between nodes in the input layer, the hidden layer, and the output layer of the feed forward neural network;

generate the graphical user interface at least in part by positioning the matrix of symbols above and adjacent to the node-link diagram in the graphical user interface; and
transmit a display communication to a display device for causing the display device to output the graphical user interface.

2. The system of claim 1 , wherein the memory device further comprises instructions that are executable by the processing device for causing the processing device to determine each respective feature-map value represented in the matrix by:
convolving filter values with input values to generate a convolutional matrix;
summing values in the convolutional matrix to determine a sum;
applying a weight to the sum to determine a weighted sum; and
using the weighted sum as the respective feature-map value.

3. The system of claim 2 , wherein the input values include embedding values determined at least in part by training the convolutional neural network.

4. The system of claim 1 , wherein the matrix of symbols is a first matrix of symbols, and the graphical user interface further comprises a second matrix of symbols positioned adjacent to the first matrix of symbols, each symbol in the second matrix of symbols indicating a respective weight for a node in the input layer of the convolutional neural network, and each row of symbols in the second matrix of symbols being horizontally aligned with a corresponding row in the first matrix of symbols.

5. The system of claim 4 , wherein the graphical user interface further comprises a textual representation of the input positioned adjacent to the second matrix of symbols such that each character in the textual representation of the input is horizontally aligned with a row of symbols in the second matrix of symbols.

6. The system of claim 4 , wherein each symbol in the first matrix of symbols is color coded to represent a respective feature-map value, and each symbol in the second matrix of symbols is color coded to represent the respective weight.

7. The system of claim 1 , wherein the graphical user interface further comprises a row of letters positioned below the matrix of symbols, each letter in the row of letters being vertically aligned with a column in the matrix of symbols and corresponding to a maximum value in the column.

8. The system of claim 1 , wherein the memory device further comprises instructions that are executable by the processing device for causing the processing device to incorporate at least one legend into the graphical user interface, wherein the at least one legend indicates a meaning of the color coding of the first row of symbols, second row of symbols, and third row of symbols.

9. The system of claim 1 , wherein the memory device further comprises instructions that are executable by the processing device for causing the processing device to execute the convolutional neural network to determine a respective feature-map value for each symbol in the matrix of symbols.

10. A method for visualizing convolutional neural networks, the method comprising:
generating, by a processing device, a matrix of symbols to be positioned in a graphical user interface, each symbol in the matrix indicating a feature-map value that represents a likelihood of a particular feature being present or absent at a location in an input to a convolutional neural network, each column in the matrix having feature-map values generated by convolving the input to the convolutional neural network with a respective filter for identifying a specific feature in the input;
generating, by the processing device, a node-link diagram to be positioned in the graphical user interface, wherein the node-link diagram represents a feed forward neural network that forms part of the convolutional neural network and comprises:
a first row of symbols representing an input layer to the feed forward neural network, wherein the input layer is also a maxpooling layer of the convolutional neural network, the first row of symbols is color coded to represent activation values for nodes in the input layer, and each symbol in the first row of symbols is vertically aligned with a respective column in the matrix of symbols and indicates a maximum value in the respective column;
a second row of symbols representing a hidden layer of the feed forward neural network, the second row of symbols being color coded to represent activation values of nodes in the hidden layer;
a third row of symbols representing an output layer of the feed forward neural network, the third row of symbols being color coded to represent activation values of nodes in the output layer; and
lines between the first row of symbols, the second row of symbols, and the third row of symbols, the lines representing connections between nodes in the input layer, the hidden layer, and the output layer of the feed forward neural network;

generating, by the processing device, the graphical user interface at least in part by positioning the matrix of symbols above and adjacent to the node-link diagram in the graphical user interface; and
transmitting, by the processing device, a display communication to a display device for causing the display device to output the graphical user interface.

11. The method of claim 10 , further comprising determining each respective feature-map value represented in the matrix by:
convolving filter values with input values to generate a convolutional matrix;
summing values in the convolutional matrix to determine a sum;
applying a weight to the sum to determine a weighted sum; and
using the weighted sum as the respective feature-map value.

12. The method of claim 11 , wherein the input values inclu

Assignees

• Sas Inst Inc
• Univ North Carolina State

Inventors

• Leeman-Munk Samuel Paul
• Sethi Saratendu
• Healey Christopher Graham
• Nie Shaoliang
• Padia Kalpesh
• Devarajan Ravinder
• Caira David James
• Benson Jordan Riley
• Cox James Allen
• Lewis Lawrence E
• Kabul Mustafa Onur

Classifications

• G06T11/26

  Drawing of charts or graphs · CPC title

• G06N3/048

  Activation functions · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/0499

  Feedforward networks · CPC title