Interactive visualizations of a convolutional neural network

Patent metadata

Field	Value
Publication number	US-10048826-B2
Application number	US-201715724029-A
Country	US
Kind code	B2
Filing date	Oct 3, 2017
Priority date	Oct 4, 2016
Publication date	Aug 14, 2018
Grant date	Aug 14, 2018

Abstract

Interactive visualizations of a convolutional neural network are provided. For example, a graphical user interface (GUI) can include a matrix having symbols indicating feature-map values that represent likelihoods of particular features being present or absent at various locations in an input to a convolutional neural network. Each column in the matrix can have feature-map values generated by convolving the input to the convolutional neural network with a respective filter for identifying a particular feature in the input. The GUI can detect, via an input device, an interaction indicating that that the columns in the matrix are to be combined into a particular number of groups. Based on the interaction, the columns can be clustered into the particular number of groups using a clustering method. The matrix in the GUI can then be updated to visually represent each respective group of columns as a single column of symbols within the matrix.

First claim

The invention claimed is:

1. A system for providing an interactive visualization of a convolutional neural network, 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:
display, via a display device, a graphical user interface comprising a matrix having rows and columns of symbols indicating feature-map values that represent likelihoods of particular features being present or absent at various locations 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 particular feature in the input;
detect, via an input device, an interaction with the graphical user interface indicating that the columns in the matrix are to be combined into a particular number of groups; and
in response to detecting the interaction:
cluster the columns into the particular number of groups using a clustering method;
for each group of columns, determine a plurality of average feature-map values, each average feature-map value being determined by averaging the feature-map values represented in a respective row of the columns; and
display an updated version of the matrix within the graphical user interface by visually representing each respective group of columns as a single column of symbols within the matrix, each symbol in the single column of symbols having visual characteristics representing an average feature-map value, from among the plurality of average feature-map values, that corresponds to a row in which the symbol is positioned.

2. The system of claim 1 , wherein the interaction is a first interaction, and wherein the memory device further includes instructions executable by the processing device for causing the processing device to:
detect a second interaction with the graphical user interface, the second interaction indicating that a group of columns represented by a single column in the updated version of the matrix is to be expanded; and
based on detecting the second interaction, display an expanded version of the matrix that visually includes all of the columns in the group of columns.

3. The system of claim 2 , wherein the memory device further includes instructions executable by the processing device for causing the processing device to:
detect a third interaction indicating that the expanded version of the matrix is to be contracted; and
based on detecting the third interaction, display a contracted version of the matrix in which the group of columns are merged into the single column.

4. The system of claim 1 , wherein the interaction is a first interaction, and wherein the memory device further includes instructions executable by the processing device for causing the processing device to:
detect a second interaction with the graphical user interface, the second interaction comprising hovering a cursor over a symbol in the matrix; and
based on detecting the second interaction, modify the graphical user interface by:
displaying a first matrix of blocks within the graphical user interface, the first matrix of blocks being color coded to represent weights of a filter used to generate a feature-map value indicated by the symbol; and
displaying a second matrix of blocks within the graphical user interface, the second matrix of blocks being color coded to represent values in a convolutional matrix formed by multiplying the weights of the filter by embedding weights corresponding to the symbol.

5. The system of claim 4 , wherein:
the clustering method includes a k-means clustering method;
the first matrix of blocks is positioned above and adjacent to the second matrix of blocks;
the first matrix of blocks and the second matrix of blocks are positioned to a right of the matrix that has the rows and columns of symbols; and
the symbols in the matrix are color coded to represent the feature-map values.

6. The system of claim 1 , wherein the interaction is a first interaction, and wherein the memory device further includes instructions executable by the processing device for causing the processing device to:
detect a second interaction with the graphical user interface, the second interaction indicating a threshold value for compressing the rows in the matrix; and
based on detecting the second interaction, display a compressed version of the matrix by:
determining that adjacent rows in the matrix have a Euclidian distance that is below the threshold value; and
merging the adjacent rows into a single row in the matrix, wherein the single row has symbols with visual characteristics that represent averages of the feature-map values represented by the adjacent rows.

7. The system of claim 6 , wherein the memory device further includes instructions executable by the processing device for causing the processing device to:
detect a third interaction with the graphical user interface, the third interaction indicating that a group of rows represented by a single row in the compressed version of the matrix is to be expanded; and
based on detecting the third interaction, display an expanded version of the matrix by visually displaying all of the rows in the group of rows.

8. The system of claim 1 , wherein the interaction is a first interaction, and wherein the memory device further includes instructions executable by the processing device for causing the processing device to include a node-link diagram within the graphical user interface, wherein the node-link diagram includes:
a first row of symbols representing an input layer to a feed forward neural network that is part of the convolutional neural network, the first row of symbols having visual characteristics representative of values at the input layer;
one or more rows of symbols representing one or more hidden layers of the feed forward neural network, the one or more rows of symbols having visual characteristics representative of values at the one or more hidden layers;
a final row of symbols representing an output layer of the feed forward neural network, the final row of symbols having visual characteristics representative of values at the output layer; and
lines between (i) the first row of symbols representing the input layer, (ii) the one or more rows of symbols representing the one or more hidden layers, and (iii) the final row of symbols representing the output layer, wherein the lines represent connections between the input layer, the one or more hidden layers, and the output layer.

9. The system of claim 8 , wherein the interaction is a first interaction, and wherein the memory device further includes instructions executable by the processing device for causing the processing device to:
detect a second interaction indicating that a layer of the feed forward neural network is to be deactivated in the graphical user interface; and
based on detecting the second interaction:
visually represent the layer of the feed forward neural network as a line of blocks, each block representing a node in the layer of the feed forward neural network and being color coded to represent an activation value of the node; and
visually hide at least a portion of the lines that represent connections between the layer of the feed forward neural network and an adjacent layer of the feed forward neural network.

10. The system of claim 9 , wherein the memory device further includes instructions executable by the processing device for causing the processing device to:
determine that the adjacent layer is also to be deactivated in the graphical user interface; and
based on determining that the

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

• G06N3/045

  Combinations of networks · CPC title

• G06F8/34

  Graphical or visual programming · CPC title

• G06F3/04812 Primary

  Interaction techniques based on cursor appearance or behaviour, e.g. being affected by the presence of displayed objects · CPC title

• G06F9/451

  Execution arrangements for user interfaces · CPC title