Confidential tuning of pre-trained machine learning models

Patent metadata

Field	Value
Publication number	US-2024177049-A1
Application number	US-202218058840-A
Country	US
Kind code	A1
Filing date	Nov 25, 2022
Priority date	Nov 25, 2022
Publication date	May 30, 2024
Grant date	—

Abstract

Confidential tuning of pre-trained machine learning models may be provided. A request associated with a model user account to fine-tune a pre-trained machine learning model with model access restrictions may be received. The pre-trained machine learning model may be one of many pre-trained machine learning models uploaded for selection and fine-tuning. The pre-trained machine learning model may be further trained using a request specified data set, with the model access restrictions and access restrictions for the data set being enforced as part of the training. Then, the fine-tuned machine learning model may be made available for invocation by an application associated with the model user account without violating the model access restrictions and data access restrictions.

First claim

What is claimed is:

1 . A system, comprising:
at least one processor; and a memory, storing program instructions that when executed by the at least one processor, cause the at least one processor to implement a machine learning service as part of a provider network, wherein the machine learning service is configured to:
receive a request associated with a model user account of the provider network to fine-tune a pre-trained machine learning model with model access restrictions uploaded in association with a model provider account of the provider network, wherein the request specifies a data set with data access restrictions for fine-tuning the pre-trained machine learning model;
load the pre-trained machine learning model, one or more tuning instructions specified by the provider of the pre-trained machine learning model, and the data set onto one or more computing resources provisioned by the machine learning service for performing the fine-tuning;
initiate training of the pre-trained machine learning model at the one or more computing resources according to the one or more tuning instructions and the data set to generate a fine-tuned machine learning model, wherein the model access restrictions and the data access restrictions are enforced at the one or more computing resources as part of the training;
store the fine-tuned machine learning model in a data store with the model access restrictions enforced;
provide an indication via an interface of the machine learning system that the fine-tune machine learning model is generated; and
responsive to a request to deploy the fine-tuned machine learning model,
obtain the fine-tuned machine learning model from the data store; and
deploy the fine-tuned machine learning model for invocation by an application associated with the model user account without violating the model access restrictions and the data access restrictions, wherein the invocation by the application is received via a network endpoint provisioned for the fine-tuned machine learning model by the machine learning service.

2 . The system of claim 1 , wherein the machine learning service is further configured to:
identify one or more tuning analyses to include as part of performing the request to fine tune the pre-trained machine learning model; include the one or more tuning analyses in a training job that is used to execute the training of the pre-trained machine learning model; and provide one or more analysis reports generated based on respective results of the one or more tuning analyses.

3 . The system of claim 2 , wherein one of the one or more analysis reports compares performance of the tuned machine learning model with a different pre-trained machined learning model.

4 . The system of claim 1 , wherein the request further specifies one or more tuning parameters that are used as part of training the pre-trained machine learning model to generate the tuned machine learning model.

5 . A method, comprising:
receiving, by a machine learning system, a request associated with a model user account to fine-tune a pre-trained machine learning model with model access restrictions uploaded in association with a model provider account, wherein the request specifies a data set with data access restrictions for fine-tuning the pre-trained machine learning model; loading, by the machine learning system, the pre-trained machine learning model, one or more tuning instructions specified by the provider of the pre-trained machine learning model, and the data set onto one or more computing resources for performing the fine-tuning; initiating, by the machine learning system, training of the pre-trained machine learning model at the one or more computing resources according to the one or more tuning instructions and the data set to generate a fine-tuned machine learning model, wherein the model access restrictions and the data access restrictions are enforced at the one or more computing resources as part of the training; and making, by the machine learning system, the fine-tuned machine learning model available for invocation by an application associated with the model user account without violating the model access restrictions and the data access restrictions.

6 . The method of claim 5 , further comprising:
identifying, by the machine learning system, one or more tuning analyses to include as part of performing the request to fine tune the pre-trained machine learning model; including, by the machine learning system, the one or more tuning analyses in a training job that is used to execute the training of the pre-trained machine learning model; and providing, via an interface of the machine learning system, one or more analysis reports generated based on respective results of the one or more tuning analyses.

7 . The method of claim 6 , wherein one of the one or more analysis reports compares performance of the tuned machine learning model with a different pre-trained machined learning model.

8 . The method of claim 6 , wherein one of the one or more analysis reports compares changes to the pre-trained model with respect to different epochs when training the pre-trained machine learning model to generate the tuned machine learning model.

9 . The method of claim 6 , wherein one of the one or more analysis reports recommends a change in makeup of the data set for a generate a different tuned version of the pre-trained machine learning model.

10 . The method of claim 6 , wherein identifying the one or more tuning analyses is based on one or selected tuning analyses specified in the request.

11 . The method of claim 5 , wherein the request enables a debug mode for performance of the request that allows sharing of the data set in a debug data store accessible to the model provider account.

12 . The method of claim 5 , wherein the request further specifies one or more tuning parameters that are used as part of training the pre-trained machine learning model to generate the tuned machine learning model.

13 . The method of claim 5 , wherein making the fine-tuned machine learning model available for invocation by an application associated with the model user account without violating the model access restrictions comprises provisioning a network endpoint that is accessible to the model user account and deploying the fine-tuned machine learning model to a model host that is accessible via the network endpoint.

14 . One or more non-transitory, computer-readable storage media, storing program instructions that when executed on or across one or more computing devices cause the one or more computing devices to implement:
receiving, by a machine learning system, a request associated with a model user account to fine-tune a pre-trained machine learning model with model access restrictions uploaded in association with a model provider account, wherein the request specifies a data set with data access restrictions for fine-tuning the pre-trained machine learning model; loading, by the machine learning system, the pre-trained machine learning model, one or more tuning instructions specified by the provider of the pre-trained machine learning model, and the data set onto one or more computing resources for performing the fine-tuning; initiating, by the machine learning system, training of the pre-trained machine learning model at the one or more computing resources according to the one or more tuning instructions and the data set to generate a fine-tuned machine learning model, wherein the model access restrictions and the data access restrictions are enforced at the one or more computing resources as part of the t

Assignees

• Amazon Tech Inc

Inventors

• Ramakrishnan Lakshmi Naarayanan
• Olgiati Andrea
• Mehrotra Ankur
• Bharathy Karthik Gurumoorthy Subramanya
• Ramakrishnan Rakesh

Classifications

• G06N20/00 Primary

  Machine learning · CPC title