Compact neural networks using condensed filters

Patent metadata

Field	Value
Publication number	US-10872292-B1
Application number	US-201816155656-A
Country	US
Kind code	B1
Filing date	Oct 9, 2018
Priority date	Oct 9, 2017
Publication date	Dec 22, 2020
Grant date	Dec 22, 2020

Abstract

A compact neural network system can generate multiple individual filters from a compound filter. Each convolutional layer of a convolutional neural network can include a compound filters used to generate individual filters for that layer. The individual filters overlap in the compound filter and can be extracted using a sampling operation. The extracted individual filters can share weights with nearby filters thereby reducing the overall size of the convolutional neural network.

First claim

What is claimed is:

1. A method comprising:
generating, using one or more processors of a machine, an image;
generating, from a compound neural network filter, a plurality of additional filters, wherein the plurality of additional filters are kernels of a convolution layer in a convolutional neural network;
generating, using the convolutional neural network, a modified image from the image, the convolutional neural network configured to generate the modified image by applying the plurality of additional filters to the image; and
causing the modified image to be displayed on a network site.

2. The method of claim 1 , wherein the plurality of additional filters are generated by sampling the compound neural network filter.

3. The method of claim 2 , wherein the plurality of additional filters overlap in the compound neural network filter.

4. The method of claim 1 , wherein the plurality of additional filters share weights.

5. The method of claim 1 , wherein the convolutional neural network comprises a plurality of convolution layers, each convolution layer having a corresponding compound neural network filter configured to generate a plurality of individual filters for that convolution layer, the plurality of convolution layers comprising the convolution layer.

6. The method of claim 5 , further comprising:
generating, for each of the convolution layers, a set of additional filters from a corresponding compound filter.

7. The method of claim 1 , wherein at least one of the plurality of additional filters is generated using a rotation operation.

8. The method of claim 1 , wherein at least one of the plurality of additional filters is generated using a reflection operation.

9. The method of claim 1 , wherein the convolutional neural network is configured to perform image segmentation.

10. The method of claim 1 , wherein the convolutional neural network is configured to perform image style transfer.

11. The method of claim 1 , wherein the modified image is published as an ephemeral message on the network site.

12. A system comprising:
one or more processors of a machine; and
a memory storing instructions that, when executed by the one or more processors, cause the machine to perform operations comprising:
generating, using one or more processors of a machine, an image;
generating, from a compound neural network filter, a plurality of additional filters, wherein the plurality of additional filters are kernels of a convolution layer in a convolutional neural network;
generating, using the convolutional neural network, a modified image from the image, the convolutional neural network configured to generate the modified image by applying the plurality of additional filters to the image; and
causing the modified image to be displayed on a network site.

13. The system of claim 12 , wherein the plurality of additional filters are generated by sampling the compound neural network filter.

14. The system of claim 13 , wherein the plurality of additional filters overlap in the compound neural network filter.

15. The system of claim 12 , wherein the plurality of additional filters share weights.

16. The system of claim 12 , wherein the convolutional neural network comprises a plurality of convolution layers, each convolution layer having a corresponding compound neural network filter configured to generate a plurality of individual filters for that convolution layer, the plurality of convolution layers comprising the convolution layer.

17. A machine-readable storage device embodying instructions that, when executed by a machine, cause the machine to perform operations comprising:
generating, using one or more processors of a machine, an image;
generating, from a compound neural network filter, a plurality of additional filters, wherein the plurality of additional filters are kernels of a convolution layer in a convolutional neural network;
generating, using the convolutional neural network, a modified image from the image, the convolutional neural network configured to generate the modified image by applying the plurality of additional filters to the image; and
causing the modified image to be displayed on a network site.

18. The machine-readable storage device of claim 17 , wherein the plurality of additional filters are generated by sampling the compound neural network filter.

Assignees

• Yang Yingzhen
• Yang Jianchao
• Xu Ning
• Snap Inc

Inventors

• Yang Yingzhen
• Yang Jianchao
• Xu Ning

Classifications

• G06N3/045 Primary

  Combinations of networks · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/0495

  Quantised networks; Sparse networks; Compressed networks · CPC title

• G06N3/08

  Learning methods · CPC title

• G06T7/10

  Segmentation; Edge detection (motion-based segmentation G06T7/215) · CPC title