Systems and methods for online distributed embedding services
AssigneeFacebook, Inc.
InventorBradley Ray Green
Life of this patent
Estimating the term…
Abstract
Systems, methods, and non-transitory computer-readable media can receive JOINKEY information identifying a JOINKEY and embedding element information identifying a plurality of embedding elements associated with the JOINKEY. A training instance is created comprising a pre-determined number of embedding elements from the plurality of embedding elements. A plurality of negative samples from a sample cache are added to the training instance. One or more embeddings are updated based on the training instance.
Claims19
A computer-implemented method comprising:
- receiving, by a computing system, JOINKEY information identifying a JOINKEY and embedding element information identifying a plurality of embedding elements associated with the JOINKEY;
- creating, by the computing system, a training instance comprising a pre-determined number of embedding elements from the plurality of embedding elements;
- labeling, by the computing system, a first embedding element of the training instance as an item element, and one or more other embedding elements of the training instance as context elements;
- adding, by the computing system, a plurality of negative samples from a sample cache to the training instance, wherein the plurality of negative samples comprises a pre-determined number of negative samples that are determined based on a number of context elements;
- updating, by the computing system, one or more embeddings based on the training instance;
- and
- writing, by the computing system, the context elements in the training instance to the sample cache for potential selection as negative samples in future training instances.
A system comprising:
- at least one processor;
- and
- a memory storing instructions that, when executed by the at least one processor, cause the system to perform a method comprising:
- receiving JOINKEY information identifying a JOINKEY and embedding element information identifying a plurality of embedding elements associated with the JOINKEY;
- creating a training instance comprising a pre-determined number of embedding elements from the plurality of embedding elements;
- labeling a first embedding element of the training instance as an item element, and one or more other embedding elements of the training instance as context elements;
- adding a plurality of negative samples from a sample cache to the training instance, wherein the plurality of negative samples comprises a pre-determined number of negative samples that are determined based on a number of context elements;
- updating one or more embeddings based on the training instance;
- and
- writing the context elements in the training instance to the sample cache for potential selection as negative samples in future training instances.
A non-transitory computer-readable storage medium including instructions that, when executed by at least one processor of a computing system, cause the computing system to perform a method comprising:
- receiving JOINKEY information identifying a JOINKEY and embedding element information identifying a plurality of embedding elements associated with the JOINKEY;
- creating a training instance comprising a pre-determined number of embedding elements from the plurality of embedding elements;
- labeling a first embedding element of the training instance as an item element, and one or more other embedding elements of the training instance as context elements;
- adding a plurality of negative samples from a sample cache to the training instance, wherein the plurality of negative samples comprises a pre-determined number of negative samples that are determined based on a number of context elements;
- updating one or more embeddings based on the training instance;
- and
- writing the context elements in the training instance to the sample cache for potential selection as negative samples in future training instances.
Dependencies are inferred from the claim text.