# US 2022/0036167 A1 · Sorting method, operation method and operation apparatus for convolutional neural network

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2022/0036167 A1
- Document ID: US20220036167A1
- Status: Published application
- Kind: A1
- Application number: 17/335,569
- Filed: 2021-06-01
- Published: 2022-02-03
- Earliest claimed: 2020-07-31
- Assignees: Xiamen Sigmastar Technology Ltd
- Applicants: Xiamen SigmaStar Technology Ltd.
- Inventors: Chao LI; Wei ZHU; Bo LIN
- Web page: https://patentgrep.com/patent/US20220036167A1

## CPC

- G06F 7/08: Sorting, i.e. grouping record carriers in numerical or other ordered sequence according to the classification of at least some of the information they carry (by merging two or more sets of carriers in ordered sequence G06F7/16)
- G06N 3/063: using electronic means
- G06F 7/24: Sorting, i.e. extracting data from one or more carriers, rearranging the data in numerical or other ordered sequence, and rerecording the sorted data on the original carrier or on a different carrier or set of carriers {sorting methods in general}(G06F7/36 takes precedence)
- G06F 7/50: Adding; Subtracting (G06F7/483 - G06F7/491, G06F7/544 - G06F7/556 take precedence)
- G06F 7/523: Multiplying only
- G06F 7/5443: {Sum of products (for applications thereof, see the relevant places, e.g. G06F17/10, H03H17/00)}
- G06N 3/045: Combinations of networks
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/0495: Quantised networks; Sparse networks; Compressed networks
- Y02D 10/00: Energy efficient computing, e.g. low power processors, power management or thermal management
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- Y02D: CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE

## Abstract

An operation method in a convolutional neural network applied to an electronic apparatus having a memory storing convolutional kernel data having undergone a sorting process. The operation method includes: performing the sorting process on a first feature vector of feature map data under process according to a marking sequence corresponding to a first weighting vector of the convolutional kernel data having undergone the sorting process; removing a part of feature values in the first feature vector having undergone the sorting process to generate a second feature vector; and performing a multiply accumulation operation on the basis of the first weighting vector and the second feature vector. The convolutional kernel data having undergone the sorting process is obtained by means of performing sorting and zero-weighting removal processes, and the marking sequence is generated according to the sorting and zero-weighting removal processes corresponding to the first weighting vector.

## Claims

### Claim 1

An operation method for a convolutional neural network, applied to an electronic apparatus, a memory in the electronic apparatus storing convolutional kernel data having undergone a sorting process, the operation method comprising:
performing the sorting process on a first feature vector of feature map data under process according to a marking sequence corresponding to a first weighting vector of the convolutional kernel data having undergone the sorting process;
removing a part of feature values in the first feature vector having undergone the sorting process to generate a second feature vector; and
performing a multiply accumulation operation on the basis of the first weighting vector and the second feature vector;
wherein, the convolutional kernel data having undergone the sorting process is obtained by means of sorting and zero-weighting removal processes, and the marking sequence is generated according to the sorting and zero-weighting removal processes corresponding to the first weighting vector.

### Claim 2 (depends on 1)

The operation method for a convolutional neural network according to claim 1, wherein the marking sequence is generated by replacing, in a second weighting vector of the convolutional kernel data having undergone the sorting process, a zero weighting by a first value and a non-zero weighting by a second value, and performing sorting.

### Claim 3 (depends on 1)

The operation method for a convolutional neural network according to claim 1, wherein the first weighting vector is obtained after correspondingly adjusting an order of the second weighting vector while obtaining the marking sequence.

### Claim 4 (depends on 2)

The operation method for a convolutional neural network according to claim 2, wherein in the step of removing the part of the feature values in the first feature vector having undergone the sorting process, the removed feature values correspond to the first values in the marking sequence.

### Claim 5 (depends on 1)

The operation method for a convolutional neural network according to claim 1, wherein the convolutional kernel data having undergone the sorting process comprises a plurality of first weighting vectors, and the plurality of first weighting vectors have a same length.

### Claim 6 (depends on 6)

A data sorting method for a convolutional neural network, comprising:
acquiring first convolutional kernel data;
splitting the first convolutional kernel data into a plurality of second weighting vectors in a channel direction;
generating a plurality of marking sequences corresponding to the second weighting vectors according to positions of zero weightings in the second weighting vectors;
performing a sorting process on weighting values in the second weighting vectors according to the marking sequences so that the zero weightings are arranged on one ends of the second weighting vectors; and
removing at least one zero weighting arranged on the one ends of the second weighting vectors to obtain a plurality of corresponding first weighting vectors;
wherein, the first weighting vectors form the convolutional kernel data having undergone the sorting process. The data sorting method for a convolutional neural network according to claim 6, wherein the first weighting vectors have a same length.

### Claim 8 (depends on 6)

The data sorting method for a convolutional neural network according to claim 6, wherein in the step of generating the marking sequences corresponding to the second weighting vectors, the marking sequence is generated by replacing, in the second weighting vectors, the zero weighting by a first value and non-zero weightings by a second value.

### Claim 9 (depends on 8)

The data sorting method for a convolutional neural network according to claim 8, wherein the step of generating the marking sequences corresponding to the second weighting vectors comprises performing the sorting process on the first value and the second value in the marking sequences so as to arrange the first value on one ends of the marking sequences.

### Claim 10 (depends on 8)

The data sorting method for a convolutional neural network according to claim 8, wherein the step of performing the sorting process on the weightings in the second weighting vectors according to the marking sequences performs the sorting process on the weightings in the second weighting vectors according to the sorting process performed on the first values and the second values in the marking sequences.

### Claim 11

An operation apparatus for a convolutional neural network, applied to an electronic apparatus, a memory in the electronic apparatus storing convolutional kernel data having undergone a sorting process, the operation apparatus comprising:
a sorting circuit, performing the sorting process on a first feature vector of feature map data under process according to a marking sequence corresponding to a first weighting vector of the convolutional kernel data having undergone the sorting process, and removing a part of feature values in the first feature vector having undergone the sorting process to generate a second feature vector; and
a multiply accumulation operation circuit, performing a multiply accumulation operation on the basis of the first weighting vector and the second feature vector;
wherein, the convolutional kernel data having undergone the sorting process is obtained by means of sorting and zero-weighting removal processes, and the marking sequence is generated according to the sorting and zero-weighting removal processes corresponding to the first weighting vector.
