Electronic apparatus and control method thereof

Patent metadata

Field	Value
Publication number	US-2020234131-A1
Application number	US-201916727323-A
Country	US
Kind code	A1
Filing date	Dec 26, 2019
Priority date	Jan 21, 2019
Publication date	Jul 23, 2020
Grant date	—

Abstract

An electronic apparatus is provided. The electronic apparatus includes sample data and memory storing a first matrix included in an artificial intelligence model trained based on sample data, and a processor configured to prunes each of a plurality of first elements included in the first matrix based on a first threshold, and acquire a first pruning index matrix that indicates whether each of the plurality of first elements has been pruned with binary data, factorize the first matrix to a second matrix of which size was determined based on the number of rows and the rank, and a third matrix of which size was determined based on the rank and the number of columns of the first matrix, prunes each of a plurality of second elements included in the second matrix based on a second threshold, and acquire a second pruning index matrix that indicates whether each of the plurality of second elements has been pruned with binary data, prunes each of a plurality of third elements included in the third matrix based on a third threshold, and acquire a third pruning index matrix that indicates whether each of the plurality of third elements has been pruned with binary data, acquire a final index matrix based on the second pruning index matrix and the third pruning index matrix, and update at least one of the second pruning index matrix or the third pruning index matrix by comparing the final index matrix with the first pruning index matrix.

First claim

What is claimed is:

1 . An electronic apparatus comprising:
a memory configured to store a first matrix, wherein the first matrix is included in an artificial intelligence model, wherein the artificial intelligence model is trained based on sample data; and a processor configured to:
prune each of a plurality of first elements included in the first matrix based on a first threshold,
acquire a first pruning index matrix, wherein the first pruning index matrix indicates for each element of the plurality of first elements, whether each element of the plurality of first elements has been pruned,
factorize the first matrix to a second matrix and a third matrix, wherein a size of the second matrix is determined based on a number of rows of the first matrix and a rank, wherein a size of the third matrix is determined based on the rank and a number of columns of the first matrix,
prune, based on a second threshold, each of a plurality of second elements included in the second matrix,
acquire a second pruning index matrix, wherein the second pruning index matrix indicates for each element of the plurality of second elements, whether each element of the plurality of second elements has been pruned,
prune, based on a third threshold, each of a plurality of third elements included in the third matrix,
acquire a third pruning index matrix, wherein the third pruning index matrix indicates for each element of the plurality of third elements, whether each element of the plurality of third elements has been pruned,
acquire a final index matrix based on the second pruning index matrix and the third pruning index matrix, and
update at least one of the second pruning index matrix or the third pruning index matrix by comparing the final index matrix with the first pruning index matrix.

2 . The electronic apparatus of claim 1 ,
wherein the processor is further configured to: compare elements included in the final index matrix with elements in corresponding positions included in the first pruning index matrix, identify positions at which the elements included in the final index matrix do not match the elements in corresponding positions included in the first pruning index matrix, and update at least one of the second pruning index matrix or the third pruning index matrix, wherein the processor is configured to perform the update by changing, based on the sizes of elements of the first matrix corresponding to the identified positions, at least one of the second threshold or the third threshold.

3 . The electronic apparatus of claim 2 ,
wherein the processor is further configured to: based on identifying a plurality of positions at which the elements included in the final index matrix do not match the elements in corresponding positions included in the first pruning index matrix, sum the sizes of a plurality of elements of the first matrix corresponding to the plurality of identified positions, and based on the summed size being equal to or greater than a threshold size, change at least one of the second threshold or the third threshold.

4 . The electronic apparatus of claim 2 ,
wherein the processor is further configured to: based on one of the second threshold or the third threshold being increased, decrease the other one of the second threshold or the third threshold, and based on one of the second threshold or the third threshold being decreased, increase the other one of the second threshold or the third threshold.

5 . The electronic apparatus of claim 1 ,
wherein the processor is further configured to: ternary quantize each of the plurality of first elements and acquire a quantization matrix including a representative value matrix and first binary data, acquire a random matrix, wherein a size of the random matrix is based on a compression subject unit and a compression target unit of the first binary data, acquire a plurality of equations based on the random matrix and the compression subject unit, and remove at least some of the plurality of equations based on binary data corresponding to the pruned first element among a plurality of binary data corresponding to the compression subject unit, and acquire second binary data corresponding to the compression target unit based on remaining equations of the plurality of equations.

6 . The electronic apparatus of claim 5 ,
wherein the processor is further configured to: based on a number of the remaining equations exceeding a number of unknowns included in the compression target unit, identify, among the remaining equations based on dependency among the remaining equations, a plurality of first equations corresponding to the number of unknowns, and acquire third binary data corresponding to the compression target unit based on the plurality of first equations.

7 . The electronic apparatus of claim 6 ,
wherein the processor is further configured to: identify whether at least one second equation excluding the plurality of first equations is established based on the third binary data, and generate patch information corresponding to third equations that are not established, wherein the plurality of equations includes the third equations, wherein the patch information includes information on the number of the third equations and identification information of each of the third equations that are not established.

8 . The electronic apparatus of claim 5 ,
wherein the processor is further configured to: ternary quantize each of a plurality of first elements that were not pruned in the first matrix, and acquire the quantization matrix including the representative value matrix and the first binary data.

9 . The electronic apparatus of claim 8 ,
wherein the processor is further configured to: identify the plurality of first elements that were not pruned in the first matrix based on the final index matrix.

10 . The electronic apparatus of claim 5 ,
wherein the random matrix includes elements of a first type and elements of a second type, and a number of the elements of the first type included in the random matrix and a number of the elements of the second type included in the random matrix are identical to each other.

11 . A control method of an electronic apparatus storing a first matrix included in an artificial intelligence model trained based on sample data, the control method comprising:
pruning each of a plurality of first elements included in the first matrix based on a first threshold, and acquiring a first pruning index matrix, wherein the first pruning index matrix indicates for each element of the plurality of first elements, whether each element of the plurality of first elements; factorizing the first matrix to a second matrix and a third matrix, wherein a size of the second matrix is determined based on a number of rows of the first matrix and a rank, wherein a size of the third matrix is determined based on the rank and a number of columns of the first matrix; pruning, based on a second threshold, each of a plurality of second elements included in the second matrix; acquiring a second pruning index matrix, wherein the second pruning index matrix indicates for each element of the plurality of second elements, whether each element of the plurality of second elements has been pruned; pruning, based on a third threshold, each of a plurality of third elements included in the third matrix; acquiring a third pruning index matrix, wherein the third pruning index matrix indicates for each element of the plurality of third elements, whether each element of the plurality of third elements has been pruned; acquiring a final index matrix based on the second pruning index matrix and the third p

Assignees

• Samsung Electronics Co Ltd

Inventors

• Lee Dongsoo
• Kwon Sejung
• Kapoor Parichay
• Kim Byeoungwook

Classifications

• G06N3/045

  Combinations of networks · CPC title

• G06N3/0495

  Quantised networks; Sparse networks; Compressed networks · CPC title

• G06N3/082 Primary

  modifying the architecture, e.g. adding, deleting or silencing nodes or connections · CPC title

• G06N3/063 Primary

  using electronic means · CPC title

• G06F17/16

  Matrix or vector computation {, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization (matrix transposition G06F7/78)} · CPC title