Statistical max pooling with deep learning

Patent metadata

Field	Value
Publication number	US-10013644-B2
Application number	US-201615346662-A
Country	US
Kind code	B2
Filing date	Nov 8, 2016
Priority date	Nov 8, 2016
Publication date	Jul 3, 2018
Grant date	Jul 3, 2018

Abstract

A method is provided for image processing. The method includes receiving an image dataset having image data for a plurality of images and a group dataset having group information for each image. The method further includes transforming the image dataset into a sparse vector by applying sparse coding and using the group information. The method also includes dividing each of dimensions of the sparse vector into multiple groups with a matrix. The matrix is generated such that the matrix takes on a higher value when samples in a same group from among the multiple groups have same non-zero dimensions above a threshold amount of time and takes on a lower value compared to the higher value when the samples in the same group lack the same non-zero coefficients above the threshold amount of time. The method additionally includes transforming the sparse vector into a summation vector using the multiple groups.

First claim

The invention claimed is:

1. A computer-implemented method for image processing, comprising: receiving an image dataset having image data for a plurality of images and a group dataset having group information for each of the plurality of images;
transforming the image dataset into a sparse vector by applying sparse coding and using the group Information;
dividing each of dimensions of the sparse vector into multiple groups with a matrix, the matrix being generated such that the matrix takes on a higher value when samples in a same group from among the multiple groups have same non-zero dimensions above a threshold amount of time and takes on a lower value compared to the higher value when the samples in the same group lack the same non-zero dimensions above the threshold amount of time; and
transforming the sparse vector into a summation vector using the multiple groups.

2. The computer-implemented method of claim 1 , wherein the matrix is generated by identifying one or more non-zero dimensions from a pair of sparse vectors in the same group.

3. The computer-implemented method of claim 2 , further comprising:
creating a pair of non-zero dimension index lists from the pair of sparse vectors in the same group;
randomly reordering at least one of the non-zero dimension index lists; and
computing a pairing probability between the non-zero dimension index lists subsequent to the reordering.

4. The computer-implemented method of claim 3 , wherein the image data comprises data samples, and the method further comprises:
selecting particular ones of the data samples using the pairing probability; and
updating the matrix with the particular ones of the data samples.

5. The computer-implemented method of claim 1 , further comprising using the summation vector as an input vector for a next layer of a neural network.

6. The computer-implemented method of claim 1 , wherein the group information comprises label information.

7. The computer-implemented method of claim 1 , further comprising sampling, wherein said sampling step comprises computing a first probability and a second probability, the first probability being a total sum of within-group transition probabilities and the second probability being a between-group transition probability.

8. The computer-implemented method of claim 7 , wherein said sampling step obtains samples based on {the first probability/(the first probability+the second probability)} p , wherein p is a hand-tuned parameter.

9. The computer-implemented method of claim 7 , wherein said sampling step is performed using a Markov Chain Monto Carlo technique.

10. The computer-implemented method of claim 1 , wherein said transforming step comprises extracting and outputting a respective maximum value, from the sparse vector, which belongs to the same group, for at least one of the multiple groups.

11. The computer-implemented method of claim 1 , further comprising performing image recognition, based on the summation vector or on information obtained from the summation vector.

12. A computer program product for image processing, the computer program product comprising a non-transitory computer readable storage medium having program instructions embodied therewith, the program instructions executable by a computer to cause the computer to perform a method comprising:
receiving an image dataset having image data for a plurality of images and a group dataset having group information for each of the plurality of images;
transforming the image dataset into a sparse vector by applying sparse coding and using the group information;
dividing each of dimensions of the sparse vector into multiple groups with a matrix, the matrix being generated such that the matrix takes on a higher value when samples in a same group from among the multiple groups have same non-zero dimensions above a threshold amount of time and takes on a lower value compared to the higher value when the samples in the same group lack the same non-zero dimensions above the threshold amount of time; and
transforming the sparse vector into a summation vector using the multiple groups.

13. The computer program product of claim 12 , wherein the matrix is generated by identifying one or more non-zero dimensions from a pair of sparse vectors in the same group.

14. The computer program product of claim 13 , wherein the method further comprises:
creating a pair of non-zero dimension index lists from the pair of sparse vectors in the same group;
randomly reordering at least one of the non-zero dimension index lists; and
computing a pairing probability between the non-zero dimension index lists subsequent to the reordering.

15. The computer program product of claim 14 , wherein the image data comprises data samples, and the method further comprises:
selecting particular ones of the data samples using the pairing probability; and
updating the matrix with the particular ones of the data samples.

16. The computer program product of claim 12 , further comprising sampling, wherein said sampling step comprises computing a first probability and a second probability, the first probability being a total sum of within-group transition probabilities and the second probability being a between-group transition probability.

17. The computer program product of claim 16 , wherein said sampling step obtains samples based on {the first probability/(the first probability+the second probability)} p , wherein p is a hand-tuned parameter.

18. The computer program product of claim 12 , further comprising performing image recognition, based on the summation vector or on information obtained from the summation vector.

19. A computing device, the computing device comprising:
a processor and a memory operably coupled to the processor, configured to:
receive an image dataset having image data for a plurality of images and a group dataset having group information for each of the plurality of images; transform the image dataset into a sparse vector by applying sparse coding and using the group information; divide each of dimensions of the sparse vector into multiple groups with a matrix, the matrix being generated such that the matrix takes on a higher value when samples in a same group from among the multiple groups have same nonzero dimensions above a threshold amount of time and takes on a lower value compared to the higher value when the samples in the same group lack the same non-zero dimensions above the threshold amount of time; and transform the sparse vector into a summation vector using the multiple groups.

20. The computing device of claim 19 , wherein said processor and memory are further configured to perform image recognition, based on the summation vector or on information obtained from the summation vector.

Assignees

• IBM

Inventors

• Takahashi Toshihiro

Classifications

• G06V30/19127

  Extracting features by transforming the feature space, e.g. multidimensional scaling; Mappings, e.g. subspace methods · CPC title

• G06V30/19173

  Classification techniques · CPC title

• G06V10/82 Primary

  using neural networks · CPC title

• G06V10/7715

  Feature extraction, e.g. by transforming the feature space, e.g. multi-dimensional scaling [MDS]; Mappings, e.g. subspace methods · CPC title

• G06F18/24

  Classification techniques · CPC title