Modifying artificial intelligence models using model fragments

Patent metadata

Field	Value
Publication number	US-11829888-B2
Application number	US-201916365721-A
Country	US
Kind code	B2
Filing date	Mar 27, 2019
Priority date	Mar 27, 2019
Publication date	Nov 28, 2023
Grant date	Nov 28, 2023

Abstract

An example system includes a processor to monitor system resources and performance preferences. The processor is to select model fragments based on the system resources and the performance preferences. The processor is to also construct a running artificial intelligence (AI) model from the selected model fragments. The processor is to further automatically modify the running AI model using the model fragments in response to detecting a change in the system resources or a change in the performance preferences.

First claim

What is claimed is:

1. A system, comprising a processor to:
monitor system resources and performance preferences, wherein the system resources comprise currently available processing at an edge device running an artificial intelligence (AI) model;
receive, at the edge device, model fragments and model fragments information that describes properties of each model fragment and properties of a plurality of different combinations of the model fragments, wherein the model fragments comprise different additional incremental compressions of a base model that is initially compressed using a lossless compression;
select, at the edge device, model fragments based on the system resources, the performance preferences, and the model fragments information comprising properties of the different additional incremental compressions;
construct a second AI model from the selected model fragments; and
automatically modify the running AI model based on the second AI model using the model fragments in response to detecting a change in the system resources or a change in the performance preferences.

2. The system of claim 1 , wherein the system resources further comprise free memory, free storage space, available power, or any combination thereof.

3. The system of claim 1 , wherein the performance preferences comprise power consumption, model size, inference time, model accuracy, adaptability to new input, or any combination thereof.

4. The system of claim 1 , wherein the processor is to select the model fragments based on the system resources, the performance preferences, current model fragments being used and a cost of migrating to a new set of model fragments.

5. The system of claim 1 , wherein the model fragments comprise different model types, parameter tunings, or compressions of an AI model.

6. The system of claim 1 , wherein the processor is to predict a change in system resources and construct a modified AI model to replace the running AI model.

7. The system of claim 1 , wherein the model fragments comprise differences between compressed models generated using reversable compression techniques.

8. A computer-implemented method, comprising:
monitoring, via a processor of an edge device, system resources and performance preferences, wherein the system resources comprise currently available processing at the edge device running an artificial intelligence (AI) model;
receiving, via the processor, model fragments and model fragments information that describes properties of each model fragment and properties of a plurality of different combinations of the model fragments, wherein the model fragments comprise different additional incremental compressions of a base model that is initially compressed using a lossless compression;
selecting, via the processor, model fragments based on the system resources, the performance preferences, and the model fragments information comprising properties of the different additional incremental compressions;
constructing, via the processor, a second AI model from the selected model fragments; and
automatically modify, via the processor, the running AI model based on the second AI model using the model fragments in response to detecting a change in the system resources or a change in the performance preferences.

9. The computer-implemented method of claim 8 , wherein selecting the model fragments is based on the system resources and the performance preferences, current model fragments being used, and a cost of migrating to a new set of model fragments.

10. The computer-implemented method of claim 8 , wherein automatically modifying the running AI model comprises enhancing the running AI model using a model fragment in response to detecting an increase in a system resource or reducing the running AI model using a model fragment in response to detecting a decrease in a system resource.

11. The computer-implemented method of claim 8 , comprising predicting a change in system resources and constructing a modified AI model to replace the running AI model.

12. The computer-implemented method of claim 8 , comprising generating the model fragments, wherein generating the model fragments comprises pruning a node or an edge of a generated AI model.

13. The computer-implemented method of claim 8 , comprising generating the model fragments, wherein generating the model fragments comprises quantizing a weight of a generated AI model.

14. The computer-implemented method of claim 8 , comprising generating the model fragments, wherein generating the model fragments comprises weight rounding a weight of a generated AI model.

15. The computer-implemented method of claim 8 , comprising retraining the modified running AI model.

16. A computer-readable storage medium for automatically modify running AI models using model fragments, the computer-readable storage medium having program code embodied therewith, the program code executable by a processor to cause the processor to:
monitor system resources and performance preferences, wherein the system resources comprise currently available processing at an edge device running an artificial intelligence (AI) model;
receive, at the edge device, model fragments and model fragments information that describes properties of each model fragment and properties of a plurality of different combinations of the model fragments, wherein the model fragments comprise different additional incremental compressions of a base model that is initially compressed using a lossless compression;
select model fragments based on the system resources, the performance preferences, and the model fragments information comprising properties of the different additional incremental compressions;
construct a second AI model from the selected model fragments; and
automatically modify the running AI model based on the second AI model using the model fragments in response to detecting a change in the system resources or a change in the performance preferences.

17. The computer-readable storage medium of claim 16 , further comprising program code executable by the processor to select the model fragments based on the system resources, the performance preferences, current model fragments being used, and a cost of migrating to a new set of model fragments.

18. The computer-readable storage medium of claim 16 , further comprising program code executable by the processor to enhance the running AI model using a model fragment in response to detecting an increase in a system resource or reduce the running AI model using a model fragment in response to detecting a decrease in a system resource.

19. The computer-readable storage medium of claim 16 , further comprising program code executable by the processor to predict a change in system resources and construct a modified AI model to replace the running AI model.

20. The computer-readable storage medium of claim 16 , further comprising program code executable by the processor to predict the system resources based on historical data or usage profiles, wherein the processor is to store a subset of the model fragments on a local storage based on the predicted system resources.

Assignees

• IBM

Inventors

• Naaman Nir
• Rosen Ira
• Turgeman Lior
• Guy Nili
• Kallner Samuel
• Amir Adar

Classifications

• G06N5/02 Primary

  Knowledge representation; Symbolic representation · CPC title

• G06F9/5011

  the resources being hardware resources other than CPUs, Servers and Terminals · CPC title

• G06F11/3409

  for performance assessment · CPC title

• G06N20/00

  Machine learning · CPC title

• G06F9/5061 Primary

  Partitioning or combining of resources · CPC title