Compatible machine learning model generation within resource constraints
AssigneeAmazon Technologies, Inc.
InventorsShuo Yang, Hao Zhou, Yuanjun Xiong, Wei Xia, Stefano Soatto
Life of this patent
Estimating the term…
Abstract
Machine learning models may be generated that are compatible with another machine learning model and satisfy a resource constraint. Techniques that ensure weight compatibility and architectural compatibility between a machine learning model being created to be compatible with another machine learning model are applied. The resource constraint is enforced so that the generated machine learning model also fits within the resource constraint.
Claims20
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:
- receive a request to create a query machine learning model that generates an embedding in a vector space used to query for objects corresponding to embeddings generated by a gallery machine learning model, the request specifying a resource constraint for the query machine learning model;
- and
- responsive to the request:
- create the query machine learning model according to an architecture identified using a search technique that satisfies:
- weight compatibility between the query machine learning model and the gallery machine learning model;
- architectural compatibility between the query machine learning model and the gallery machine learning model;
- and
- the resource constraint;
- and
- provide the query machine learning model.
A method, comprising:
- receiving, at a model generation system, a request to create a first machine learning model that generates an embedding in a vector space that is compatible with embeddings generated by a second machine learning model, the request specifying a resource constraint for the first machine learning model;
- and
- creating, by the model generation system, the first machine learning model according to an architecture identified using a search technique that satisfies:
- weight compatibility between the first machine learning model and the second machine learning model;
- architectural compatibility between the first machine learning model and the second machine learning model;
- and
- the resource constraint;
- and
- providing, by the model generation system, the first machine learning model.
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 a request to create a first machine learning model that generates an embedding in a vector space that is compatible with embeddings generated by a second machine learning model;
- responsive to the request:
- identifying a resource constraint for executing the first machine learning model based on the request;
- creating the first machine learning model according to an architecture identified using a search technique that satisfies:
- weight compatibility between the first machine learning model and the second machine learning model;
- architectural compatibility between the first machine learning model and the second machine learning model;
- and
- the resource constraint;
- and
- providing the first machine learning model.
Dependencies are inferred from the claim text.