Back to results
US 11,144,812 B2Granted
Mixed machine learning architecture
AssigneeFacebook, Inc.
InventorsXianjie Chen, Wenlin Chen, Liang Xiong, Tianshi Gao
Life of this patent
Estimating the term…
Abstract
A preprocessing module of a neural network has a first input and second input. The module generates multiple, different first latent vector representations of its first input, and multiple, different second latent vector representations of its second input. The module then models pairwise interactions between every unique pairwise combination of the first and second latent vector representations. The module then produces an intermediate output by combining the results of the modeled pairwise interactions.
Claims20
A method comprising:
- by a computing device, accessing a neural network having a module comprising embedding-pooling blocks, the module having a first input and a second input;
- and
- by the computing device, using the module to:
- process the first input to generate a plurality of first latent vector representations of the first input by submitting the first input to each of a first plurality of the embedding-pooling blocks, the plurality of first latent vector representations being different from each other and being respectively generated by the first plurality of the embedding-pooling blocks, wherein a number of the first plurality of the embedding-pooling blocks is determined based on a density of the first input;
- process the second input to generate a plurality of second latent vector representations of the second input by submitting the second input to each of a second plurality of the embedding-pooling blocks, the plurality of second latent vector representations being different from each other and being respectively generated by the second plurality of the embedding-pooling blocks, wherein a number of the second plurality of the embedding-pooling blocks is determined based on a density of the second input;
- determine unique pairwise combinations of latent vector representations, wherein each unique pairwise combination comprises latent vector representations selected from (1) the plurality of first latent vector representations of the first input and (2) the plurality of second latent vector representations of the second input;
- model pairwise interactions between the unique pairwise combinations of latent vector representations;
- and
- produce an intermediate output by combining results of the modeled pairwise interactions.
6 more elements
One or more computer-readable non-transitory storage media embodying software that is operable when executed to:
- access a neural network having a module comprising embedding-pooling blocks, the module being a module of nodes and having a first input and a second input;
- and
- use the module to:
- process the first input to generate a plurality of first latent vector representations of the first input by submitting the first input to each of a first plurality of the embedding-pooling blocks, the plurality of first latent vector representations being different from each other and being respectively generated by the first plurality of the embedding-pooling blocks, wherein a number of the first plurality of the embedding-pooling blocks is determined based on a density of the first input;
- process the second input to generate a plurality of second latent vector representations of the second input by submitting the second input to each of a second plurality of the embedding-pooling blocks, the plurality of second latent vector representations being different from each other and being respectively generated by the second plurality of the embedding-pooling blocks, wherein a number of the second plurality of the embedding-pooling blocks is determined based on a density of the second input;
- determine unique pairwise combinations of latent vector representations, wherein each unique pairwise combination comprises latent vector representations selected from (1) the plurality of first latent vector representations of the first input and (2) the plurality of second latent vector representations of the second input;
- model pairwise interactions between the unique pairwise combinations of latent vector representations;
- and
- produce an intermediate output by combining results of the modeled pairwise interactions.
6 more elements
A system comprising:
- one or more processors;
- and
- one or more computer-readable non-transitory storage media coupled to one or more of the processors and comprising instructions operable when executed by one or more of the processors to cause the system to:
- access a neural network having a module comprising embedding-pooling blocks, the module being a module of nodes and having a first input and a second input;
- and
- use the module to:
- process the first input to generate a plurality of first latent vector representations of the first input by submitting the first input to each of a first plurality of the embedding-pooling blocks, the plurality of first latent vector representations being different from each other and being respectively generated by the first plurality of the embedding-pooling blocks, wherein a number of the first plurality of the embedding-pooling blocks is determined based on a density of the first input;
- process the second input to generate a plurality of second latent vector representations of the second input by submitting the second input to each of a second plurality of the embedding-pooling blocks, the plurality of second latent vector representations being different from each other and being respectively generated by the second plurality of the embedding-pooling blocks, wherein a number of the second plurality of the embedding-pooling blocks is determined based on a density of the second input;
- determine unique pairwise combinations of latent vector representations, wherein each unique pairwise combination comprises latent vector representations selected from (1) the plurality of first latent vector representations of the first input and (2) the plurality of second latent vector representations of the second input;
- model pairwise interactions between the unique pairwise combinations of latent vector representations;
- and
- produce an intermediate output by combining results of the modeled pairwise interactions.
9 more elements
Dependencies are inferred from the claim text.