IP Library Granted Patent US 12670397
Granted Patent B2
US 12670397 · App. 17/993,623 · Granted Jun 30, 2026

Machine learning method and pruning method

Inventors: Ryusuke Seki (Kobe, JP); Yasutaka Okada (Kobe, JP); Yuki Katayama (Kobe, JP)
Assignee: DENSO TEN Limited
G06N3/082
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 12670397
App. No.
17/993,623
Granted
Jun 30, 2026
Kind
B2
Abstract

In a method of creating a learning model using a controller configured to perform pruning on a neural network, the pruning includes a first pruning process in which a pruning process is performed in units of channels of convolutional layers and a second pruning process in which a pruning process is performed in units of weight parameters.

Claims (33)

1 . A method of creating a learning model in which pruning is performed on a neural network, the method executed by a controller and comprising:

the controller performing a first pruning process in which pruning is performed on the neural network in units of channels of convolutional layers; and

the controller performing a second pruning process in which pruning is performed on the neural network in units of weight parameters,

the method further comprising

the controller classifying convolutional layers of the neural network that are linked together by residual connections into a same group and calculating an importance level of the group so that all of the convolutional layers of the group have an identical importance level, and

the controller deciding, according to the importance level that has been calculated, whether to perform the first pruning process in units of channels in the convolutional layers classified into the group.

2 . The method according to claim 1 , wherein further comprising:

the controller performing training using a data set after the first pruning process,

the controller performing the second pruning process after the training, and

the controller performing training using a data set after the second pruning process.

3 . The method according to claim 2 , wherein

the controller performs the first pruning process and the training after the first pruning process continuously a plurality of times.

4 . The method according to claim 3 , wherein

the controller performs the second pruning process and the training after the second pruning process continuously a plurality of times.

5 . A method of pruning a neural network, the method executed by a controller and comprising:

the controller performing a first pruning process in which pruning is performed on the neural network in units of channels of convolutional layers; and

the controller performing a second pruning process in which pruning is performed on the neural network in units of weight parameters,

the method further comprising

the controller classifying convolutional layers of the neural network that are linked together by residual connections into a same group and calculating an importance level of the group so that all of the convolutional layers of the group have an identical importance level, and

the controller deciding, according to the importance level that has been calculated, whether to perform the first pruning process in units of channels in the convolutional layers classified into the group.

6 . A method of machine learning in which pruning is performed on a neural network, the method executed by a controller and comprising:

the controller classifying convolutional layers of the neural network that are linked together by residual connections into a same group and calculating an importance level of the group so that all of the convolutional layers of the group have an identical importance level, and

the controller deciding, according to the importance level that has been calculated, whether to perform the pruning in units of channels in the convolutional layers classified into the group.

7 . The method according to claim 6 , further comprising:

before the classifying, the controller calculating the importance level for each of the convolutional layers that are to be classified into the group, wherein

the calculating the importance level of the group includes calculating an average importance level by averaging the calculated importance levels of the convolutional layers classified into the group, and

the deciding whether to perform the pruning in units of channels in the convolutional layers classified into the group is based on the average importance level calculated for the group.

8 . The method according to claim 6 , further comprising:

the controller calculating, for a each individual convolutional layer of the neural network that is not classified into the group, the importance level of the individual convolutional layer individually, and

the controller deciding, according to the calculated importance level, whether to perform the pruning in units of channels in the individual convolutional layer.

9 . The method according to claim 6 , further comprising:

the controller taking as a target of the pruning a predetermined proportion of the convolutional layers in increasing order of the calculated importance levels, and

for the convolutional layers grouped into the group, the controller using a number of convolutional layers grouped into the group to calculate the predetermined proportion.