# US 2021/0241172 A1 · Machine learning model compression system, pruning method, and computer program product

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

- Number: US 2021/0241172 A1
- Document ID: US20210241172A1
- Status: Published application
- Kind: A1
- Application number: 17/002,820
- Filed: 2020-08-26
- Published: 2021-08-05
- Earliest claimed: 2020-02-05
- Assignees: Kabushiki Kaisha Toshiba
- Applicants: Kabushiki Kaisha Toshiba
- Inventors: Takahiro TANAKA; Kosuke HARUKI; Ryuji SAKAI; Akiyuki TANIZAWA; Atsushi YAGUCHI; Shuhei NITTA; Yukinobu SAKATA
- Web page: https://patentgrep.com/patent/US20210241172A1

## CPC

- G06N 3/045: Combinations of networks
- G06N 3/082: modifying the architecture, e.g. adding, deleting or silencing nodes or connections
- G06F 18/211: Selection of the most significant subset of features
- G06F 18/214: Generating training patterns; Bootstrap methods, e.g. bagging or boosting
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/0495: Quantised networks; Sparse networks; Compressed networks
- G06N 3/09: Supervised learning
- G06N 3/0985: Hyperparameter optimisation; Meta-learning; Learning-to-learn
- G06N 20/00: Machine learning
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

A machine learning model compression system according to an embodiment includes one or more hardware processors configured to: select a layer of a trained machine learning model in order from an output side to an input side of the trained machine learning model; calculate, in units of an input channel, a first evaluation value evaluating a plurality of weights included in the selected layer; sort, in ascending order or descending order, the first evaluation values each calculated in units of the input channel; select a given number of the first evaluation values in ascending order of the first evaluation values; and delete the input channels used for calculation of the selected first evaluation values.

## Claims

### Claim 1

A machine learning model compression system comprising
one or more hardware processors configured to:
select a layer of a trained machine learning model in order from an output side to an input side of the trained machine learning model;
calculate, in units of an input channel, a first evaluation value evaluating a plurality of weights included in the selected layer;
sort, in ascending order or descending order, the first evaluation values each calculated in units of the input channel;
select a given number of the first evaluation values in ascending order of the first evaluation values; and
delete the input channels used for calculation of the selected first evaluation values.

### Claim 2 (depends on 1)

The system according to claim 1, wherein the first evaluation value is an L1 norm of the plurality of weights.

### Claim 3 (depends on 1)

The system according to claim 1, wherein the one or more processors are further configured to:
execute parameter selection processing to select a parameter for determining a structure of a compressed model included in a given search space;
execute weight extraction processing to extract weights of the compression model from the trained machine learning model by deleting weights corresponding to the deleted input channels;
execute compressed model generation processing to generate the compressed model by using the parameter and to set the extracted weights as initial values of weights of at least one layer of the compressed model;
execute performance evaluation processing to train the compressed model for a given period and to calculate a second evaluation value representing recognition performance of the compressed model; and
determine, based on a given end condition, whether to repeat the parameter selection processing, the weight extraction processing, the compressed model generation processing, and the performance evaluation processing.

### Claim 4 (depends on 3)

The system according to claim 3, wherein, in the compressed model generation processing, the one or more processors are configured to
receive an input of designating one or more layers for which the extracted weights are set as initial values of the weights of the compressed model, and
set the extracted weights as initial values of weights of the designated layers.

### Claim 5 (depends on 3)

The system according to claim 3, wherein the given end condition is a case in which the second evaluation value exceeds an evaluation threshold, a case in which the number of times of evaluation of the second evaluation value exceeds a number-of-times threshold, or a case in which a search time of the compressed model exceeds a time threshold.

### Claim 6

A pruning method implemented by a computer, the method comprising:
selecting a layer of a trained machine learning model in order from an output side to an input side of the trained machine learning model;
calculating, in units of an input channel, a first evaluation value evaluating a plurality of weights included in the selected layer;
sorting, in ascending order or descending order, the first evaluation values each calculated in units of the input channel;
selecting a given number of the first evaluation values in ascending order of the first evaluation value; and
deleting the input channels used for calculation of the selected first evaluation

### Claim 7

A computer program product comprising a non-transitory computer-readable recording medium on which an executable program is recorded, the program instructing the computer to:
select a layer of a trained machine learning model in order from an output side to an input side of the trained machine learning model;
calculate, in units of an input channel, a first evaluation value evaluating a plurality of weights included in the selected layer;
sort, in ascending order or descending order, the first evaluation values each calculated in units of the input channel;
select a given number of the first evaluation values in ascending order of the first evaluation values; and
delete the input channels used for calculation of the selected first evaluation values.
