Granted Patent B2

US 12,572,841 · App. 17/036,218 · Granted Mar 10, 2026

Inventors: Umar Asif (Melbourne, AU); Stefan von Cavallar (Sandringham, AU); Jianbin Tang (Doncaster East, AU); Stefan Harrer (Hampton, AU)

Assignee: International Business Machines Corporation

G06N20/00G06N3/082

Abstract

Claims (102)

1 . A computer-implemented method of selecting a machine learning model, the computer including one or more processors, one or more computer-readable memories and one or more computer-readable storage media, the method comprising:

receiving hardware specifications of a device;

receiving output-based requirements of the device;

determining a performance requirement based on the hardware specifications and the output-based requirements of the device;

acquiring a set of machine learning models from a repository of machine learning models accessible to the device, the set of machine learning models including:

a first machine learning model having a first set of layers, wherein each layer of the first set of layers manipulates input in a different way in order to produce its individual output; and

a teacher model;

determining a threshold number of layers from the first set of layers to form a pruned set of layers;

deleting one or more layers of the first set of layers from the first machine learning model, resulting in a student model with the pruned set of layers having a number of layers at or below the threshold number of layers;

training the student model based on:

training data; and

the teacher model;

evaluating a performance of the student model;

comparing the performance of the student model and the performance requirement;

determining, based at least on the comparing, to select a model having a highest performance; and

deploying the model having the highest performance on the device,

wherein each of the pruned set of layers is developed over time while training the student model.

2 . The method of claim 1 , further comprising receiving an additional performance requirement from a user, wherein the comparing further includes comparing the performance of the student model and the additional performance requirement.

3 . The method of claim 1 , wherein the deleting one or more layers includes:

determining, based on the performance requirement, a number of layers to delete;

selecting the one or more layers based on the number; and

deleting the selected layers.

4 . The method of claim 3 , wherein the one or more layers are selected at random.

5 . The method of claim 1 , wherein the deleting one or more layers includes deleting one or more blocks, each block including a continuous set of layers.

6 . The method of claim 1 , wherein the training includes:

inputting training data into the student model and the teacher model;

receiving a student feature from the student model;

receiving a teacher feature from the teacher model;

performing a first comparison between the student feature and the teacher feature;

receiving a student output from the student model;

performing a second comparison between the student output and the training data; and

adjusting at least one layer of the pruned set of layers of the student model based on the first comparison and the second comparison.

7 . The method of claim 1 , wherein each of the set of machine learning models are selected from a model repository.

8 . A computer system for selecting a machine learning model, the system comprising:

one or more processors, one or more computer-readable memories and one or more computer-readable storage media;

program instructions, stored on at least one of the one or more storage media for execution by at least one of the one or more processors via at least one of the one or more memories, to receive hardware specifications of a device;

program instructions, stored on at least one of the one or more storage media for execution by at least one of the one or more processors via at least one of the one or more memories, to receive output-based requirements of the device;

program instructions, stored on at least one of the one or more storage media for execution by at least one of the one or more processors via at least one of the one or more memories, to determine a performance requirement based on the hardware specifications and the output-based requirements;

program instructions, stored on at least one of the one or more storage media for execution by at least one of the one or more processors via at least one of the one or more memories, to acquire a set of machine learning models from a repository of machine learning models accessible to the device, the set of machine learning models including:

a first machine learning model having a first set of layers, wherein each layer of the first set of layers manipulates input in a different way in order to produce its individual output; and

a teacher model;

program instructions, stored on at least one of the one or more storage media for execution by at least one of the one or more processors via at least one of the one or more memories, to determine a threshold number of layers from the first set of layers to form a pruned set of layers;

program instructions, stored on at least one of the one or more storage media for execution by at least one of the one or more processors via at least one of the one or more memories, to delete one or more layers of the first set of layers from the first machine learning model, resulting in a student model with the pruned set of layers having a number of layers at or below the threshold number of layers;

program instructions, stored on at least one of the one or more storage media for execution by at least one of the one or more processors via at least one of the one or more memories, to train the first student model based on:

training data; and

the teacher model;

program instructions, stored on at least one of the one or more storage media for execution by at least one of the one or more processors via at least one of the one or more memories, to evaluate a performance of the first student model;

program instructions, stored on at least one of the one or more storage media for execution by at least one of the one or more processors via at least one of the one or more memories, to compare the performance of the first student model to the performance requirement;

program instructions, stored on at least one of the one or more storage media for execution by at least one of the one or more processors via at least one of the one or more memories, to determine, based at least on the comparing, to generate a second student model; and

program instructions, stored on at least one of the one or more storage media for execution by at least one of the one or more processors via at least one of the one or more memories, to deploy the model having the highest performance on the device

wherein each of the pruned set of layers is developed over time while training the student model.

9 . The system of claim 8 , further comprising:

program instructions, stored on at least one of the one or more storage media for execution by at least one of the one or more processors via at least one of the one or more memories, to receive an additional performance requirement from a user, wherein the comparing further includes comparing the performance of the student model and the additional performance requirement.

10 . The system of claim 8 , wherein the deleting one or more layers includes:

determining, based on the performance requirement, a number of layers to delete;

selecting the one or more layers based on the number; and

deleting the selected layers.

11 . The system of claim 10 , wherein the one or more layers are selected at random.

12 . The system of claim 8 , wherein the deleting one or more layers includes deleting one or more blocks, each block including a continuous set of layers.

13 . The system of claim 8 , wherein the training includes:

inputting training data into the student model and the teacher model;

receiving a student feature from the student model;

receiving a teacher feature from the teacher model;

performing a first comparison between the student feature and the teacher feature;

receiving a student output from the student model;

performing a second comparison between the student output and the training data; and

adjusting at least one layer of the pruned set of layers of the student model based on the first comparison and the second comparison.

14 . The system of claim 8 , wherein each of the set of machine learning models are selected from a model repository.

15 . A computer program product for selecting a machine learning model, the computer program product comprising:

one or more computer-readable storage media;

program instructions, stored on at least one of the one or more storage media, to receive hardware specifications of a device;

program instructions, stored on at least one of the one or more storage media, to receive output-based requirements of the device;

program instructions, stored on at least one of the one or more storage media, to determine a performance requirement based on the hardware specifications and the output-based requirements of the device;

program instructions, stored on at least one of the one or more storage media, to acquire a set of machine learning models from a repository of machine learning models accessible to the device, the set of machine learning models including:

a first machine learning model having a first set of layers, wherein each layer of the first set of layers manipulates input in a different way in order to produce its individual output; and

a teacher model;

program instructions, stored on at least one of the one or more storage media, to determine a threshold number of layers from the first set of layers to form a pruned set of layers;

program instructions, stored on at least one of the one or more storage media, to delete one or more layers of the first set of layers from the first machine learning model, resulting in a student model with the pruned set of layers having a number of layers at or below the threshold number of layers;

program instructions, stored on at least one of the one or more storage media, to train the student model based on:

training data; and

the teacher model;

program instructions, stored on at least one of the one or more storage media, to evaluate a performance of the student model;

program instructions, stored on at least one of the one or more storage media, to compare the performance of the student model and the performance requirement;

determine, based at least on the comparing, to select a model having a highest performance; and

program instructions, stored on at least one of the one or more storage media, to deploy the model having the highest performance on the device

wherein each of the pruned set of layers is developed over time while training the student model.

16 . The computer program product of claim 15 , further comprising:

program instructions, stored on at least one of the one or more storage media, to receive an additional performance requirement from a user, wherein the comparing further includes comparing the performance of the student model and the additional performance requirement.

17 . The computer program product of claim 15 , wherein the deleting one or more layers includes:

determining, based on the performance requirement, a number of layers to delete;

selecting the one or more layers based on the number; and

deleting the selected layers.

18 . The computer program product of claim 15 , wherein the deleting one or more layers includes deleting one or more blocks, each block including a continuous set of layers.

19 . The computer program product of claim 15 , wherein the training includes:

inputting training data into the student model and the teacher model;

receiving a student feature from the student model;

receiving a teacher feature from the teacher model;

performing a first comparison between the student feature and the teacher feature;

receiving a student output from the student model;

performing a second comparison between the student output and the training data; and

adjusting at least one layer of the pruned set of layers of the student model based on the first comparison and the second comparison.

20 . The computer program product of claim 15 , wherein each of the set of machine learning models are selected from a model repository.

Assignments (1)

Continuity (1)

References Cited (58)