System and method for generating training sets for neural networks
A system and method for generating training sets for training neural networks. The method includes receiving a plurality of query pairs, wherein each of the plurality of query pairs includes a query and a real result previously determined for the query; determining at least one variable element of each query in the plurality of received query pairs; determining a variance for the at least determined variable element of each query in the plurality of received query pairs; and generating a training set based on the determined variable element, the determined variance, and the previously determined real result.
1 . A method for developing a trained computer neural network capable of providing responses to queries to a table-based database, comprising:
receiving a plurality of query pairs, wherein each of the plurality of query pairs includes a query for the table-based database and a real result previously determined for the query that is a response to the query;
determining, for each query in the received plurality of query pairs, a variable and a first subset of potential values for the variable;
determining at least one second subset of potential values for the variable of each query in the plurality of received query pairs, wherein each of the at least one second subset of potential values for the variable is different from the first subset of potential values for the variable; and
generating a training set based on the determined variable for each query, the at least one first subset of potential values for each variable, the at least one second subset of potential values for each variable, and the previously determined real result for each query, wherein generating the training set includes generating at least one training query, wherein each training query is generated for the variable of one of the queries in the received plurality of query pairs based on the at least one second subset of potential values determined for the variable; and
causing a recurrent neural network to become a trained recurrent neural network through training of the recurrent neural network using the generated training set.
2 . The method of claim 1 , further comprising:
determining at least one predicate for each of the plurality of query pairs, wherein a predicate is an expression used to determine if the query will return any one of: a true result and a false result.
3 . The method of claim 1 , wherein the previously determined real result is determined by querying at least one data set using the respective query.
4 . The method of claim 3 , further comprising:
determining if the generated training set is representative of the at least one data set.
5 . The method of claim 4 , wherein determining if the generated training set is representative of the data set further comprises at least one of: determining if the query pairs are directed to all portions of the data set and determining if the query pairs are directed to a number of portions of the data set above a predetermined threshold.
6 . The method of claim 1 , wherein a variance of the query is configured to accommodate for potential bias within a determined real result.
7 . The method of claim 1 , wherein the causing further comprises:
providing the generated training set to the neural network which is one of a plurality of neural networks.
8 . The method of claim 7 , wherein the provided training set is vectorized to a matrix representation configured to be fed to input neurons of the neural network.
9 . The method of claim 1 , further comprising:
continuously generating a plurality of test queries; and
ceasing to generate test queries when the plurality of generated test queries is equal to a representative sample size of the data set.
10 . The method of claim 1 , wherein a first subset of query pairs corresponds to a first column of the data set.
11 . The method of claim 10 , wherein a second subset of query pairs corresponds to the first column and a second column of the data set, and a first plurality of test queries is generated based on the first subset of query pairs, and a second plurality of test queries is generated based on the second subset.
12 . The method of claim 11 , wherein the second plurality of test queries is generated first, and the first plurality of test queries is generated in response to the size of the second plurality of test queries being less than the sample size for a variance of the first column.
13 . A non-transitory computer readable medium having stored thereon instructions for causing a processing circuitry to perform a process, the process being to develop a trained computer neural network capable of providing responses to queries to a table-based database and comprising:
receiving a plurality of query pairs, wherein each of the plurality of query pairs includes a query for the table-based database and a real result previously determined for the query that is a response to the query;
determining, for each query in the received plurality of query pairs, a variable and a first subset of potential values for the variable;
determining at least one second subset of potential values for the variable of each query in the plurality of received query pairs, wherein each of the at least one second subset of potential values for the variable is different from the first subset of potential values for the variable; and
generating a training set based on the determined variable for each query, the at least one first subset of potential values for each variable, the at least one second subset of potential values for each variable, and the previously determined real result for each query, wherein generating the training set includes generating at least one training query, wherein each training query is generated for the variable of one of the queries in the received plurality of query pairs based on the at least one second subset of potential values determined for the variable; and
causing a recurrent neural network to become a trained recurrent neural network through training of the recurrent neural network using the generated training set.
14 . A system for developing a trained computer neural network capable of providing responses to queries to a table-based database, comprising:
a processing circuitry; and
a memory, the memory containing instructions that, when executed by the processing circuitry, configure the system to:
receive a plurality of query pairs, wherein each of the plurality of query pairs includes a query for the table-based database and a real result previously determined for the query that is a response to the query;
determine, for each query in the received plurality of query pairs, a variable and a first subset of potential values for the variable;
determine at least one second subset of potential values for the variable of each query in the plurality of received query pairs, wherein each of the at least one second subset of potential values for the variable is different from the first subset of potential values for the variable; and
generate a training set based on the determined variable for each query, the at least one first subset of potential values for each variable, the at least one second subset of potential values for each variable, and the previously determined real result for each query, wherein generating the training set includes generating at least one training query, wherein each training query is generated for the variable of one of the queries in the received plurality of query pairs based on the at least one second subset of potential values determined for the variable; and
cause a recurrent neural network to become a trained recurrent neural network through training of the recurrent neural network using the generated training set.
15 . The system of claim 14 , wherein the system is further configured to:
determine at least one predicate for each of the plurality of query pairs, wherein a predicate is an expression used to determine if the query will return any one of: a true result and a false result.
16 . The system of claim 14 , wherein the previously determined real result is determined by querying at least one data set using the respective query.
17 . The system of claim 16 , wherein the system is further configured to:
determine if the generated training set is representative of the at least one data set.
18 . The system of claim 17 , wherein determining if the generated training set is representative of the data set further comprises at least one of: determining if the query pairs are directed to all portions of the data set and determining if the query pairs are directed to a number of portions of the data set above a predetermined threshold.
19 . The system of claim 14 , wherein a variance of the query is configured to accommodate for potential bias within a determined real result.
20 . The system of claim 14 , wherein the system is further configured when performing the causing to:
provide the generated training set to the neural network which is one of a plurality of neural networks.
21 . The system of claim 20 , wherein the provided training set is vectorized to a matrix representation configured to be fed to input neurons of the neural network.
22 . The system of claim 14 , wherein the system is further configured to:
continuously generate a plurality of test queries; and
cease to generate test queries when the plurality of generated test queries is equal to a representative sample size of the data set.
23 . The system of claim 14 , wherein a first subset of query pairs corresponds to a first column of the data set.
24 . The system of claim 23 , wherein a second subset of query pairs corresponds to the first column and a second column of the data set, and a first plurality of test queries is generated based on the first subset of query pairs, and a second plurality of test queries is generated based on the second subset.
25 . The system of claim 24 , wherein the second plurality of test queries is generated first, and the first plurality of test queries is generated in response to the size of the second plurality of test queries being less than the sample size for a variance of the first column.
26 . The method of claim 1 , further comprising:
receiving, by the trained neural network, a query for execution on a data set; and
generating, by the trained neural network, a predicted result for the received query.
27 . The method of claim 1 , further comprising:
determining a confidence level of the predicted result based on a loss function of the trained neural network;
executing the query on the data set in response to determining that the confidence level is below a first predetermined threshold; and
executing the query on another neural network in response to determining that the confidence level is below a second predetermined threshold and above the first predetermined threshold.
28 . The system of claim 14 , wherein the memory contains further instructions which when executed by the processing circuitry further configures the system to:
receive a query for execution on a data set; and
generate a predicted result for the received query by providing the received query to the trained neural network.
29 . The system of claim 28 , wherein the memory contains further instructions which when executed by the processing circuitry further configures the system to:
determine a confidence level of the predicted result based on a loss function of the trained neural network;
execute the query on the data set in response to determining that the confidence level is below a first predetermined threshold; and
execute the query on another neural network in response to determining that the confidence level is below a second predetermined threshold and above the first predetermined threshold.