US 10,474,950 B2Granted

Training and operation of computational models

AssigneeMicrosoft Technology Licensing, LLC

InventorsXiaodong He, Jianshu Chen, Brendan W L Clement, Li Deng, Jianfeng Gao, Bochen Jin, Prabhdeep Singh, Sandeep P. Solanki, LuMing Wang, Hanjun Xian, Yilei Zhang, Mingyang Zhao, Zijian Zheng

Life of this patent

Estimating the term…

patentgrep

Loading patent…

Abstract

A processing unit can acquire datasets from respective data sources, each having a respective unique data domain. The processing unit can determine values of a plurality of features based on the plurality of datasets. The processing unit can modify input-specific parameters or history parameters of a computational model based on the values of the features. In some examples, the processing unit can determine an estimated value of a target feature based at least in part on the modified computational model and values of one or more reference features. In some examples, the computational model can include neural networks for several input sets. An output layer of at least one of the neural networks can be connected to the respective hidden layer(s) of one or more other(s) of the neural networks. In some examples, the neural networks can be operated to provide transformed feature value(s) for respective times.

Claims18

Dependent claims

A system comprising:

one or more computer-readable media having thereon:
a plurality of modules;
and
a computational model comprising respective neural networks and respective input-specific parameters for each of a plurality of input sets of the computational model and one or more history parameters;
and
one or more processing unit(s) operably coupled to one(s) of the computer-readable media and communicatively connectable with a plurality of data sources, each of the data sources having a respective unique data domain and the processing unit(s) adapted to execute modules of the plurality of modules comprising:
an accessing module configured to acquire a plurality of datasets from respective ones of the data sources, wherein the data in each dataset is organized according to the respective data domain;
an extraction module configured to determine values of a plurality of features based at least in part on the plurality of datasets, the plurality of features including a target feature and one or more reference features;
an estimating module configured to determine an estimated value of the target feature based at least in part on the computational model and the values of the one or more reference features;
and
an updating module configured to modify at least some of the input-specific parameters or the history parameters of the computational model based at least in part on the value of the target feature and the estimated value of the target feature.

8 more elements

A method, comprising:

receiving, from a plurality of data sources, respective datasets organized according to respective unique data domains, wherein each dataset includes data value(s) associated with individual ones of a plurality of times;
determining, using a processing unit, feature value(s) associated with each of the plurality of times based at least in part on the data value(s) associated with the corresponding ones of the plurality of times;
and
operating respective neural networks for each of the plurality of times to provide transformed feature value(s) based at least in part on ones of the feature value(s) associated with the corresponding ones of the plurality of times, wherein:
a first one of the neural networks includes a hidden layer configured to provide feedforward value(s);
a second, different one of the neural networks is configured to provide at least some of the respective transformed feature value(s) based at least in part on the feedforward value(s);
and
the time of the first one of the neural networks is earlier than the time of the second one of the neural networks.

5 more elements

A tangible computer storage medium having thereon computer-executable instructions, the computer-executable instructions upon execution configuring a computer to perform operations comprising:

extracting feature values from a plurality of datasets organized according to respective, different data domains, each feature value corresponding to a time;
operating a plurality of neural networks to provide an estimated value of a target feature based at least in part on the feature values, wherein each of the neural networks corresponds to a respective relative time period and includes a respective hidden layer communicatively connected with the hidden layer of another of the neural networks having a later relative time period;
determining an error value of the estimated value of the target feature based at least in part on a corresponding training value;
and
training the plurality of neural networks based at least in part on the error value and the feature values having times in the corresponding relative time period, wherein the training includes adjusting parameters of the respective hidden layers of at least two of the neural networks.

2 more elements

Dependencies are inferred from the claim text.

Facts

Application: 14/754,474Filed 29 Jun 2015
Published: US 2016/0379112 A129 Dec 2016
Granted: 12 Nov 2019
Term ends: Not estimatedThe source records don’t support an estimate.

Classification7 codes

G06N 3/092
Learning methods: reinforcement learning
G06N 3/0442
Recurrent networks, e.g. Hopfield networks: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
G06N 3/045
Architecture, e.g. interconnection topology: combinations of networks
G06N 3/049
Architecture, e.g. interconnection topology: temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs