Systems and methods for a neural network database framework for answering database query types
A neural network database is disclosed. A learning task to teach a single model to answer any query is formulated. The example neural network database learns existing patterns between query input and output and by exploits the query and data distributions through a decision tree having multiple neural network leaf nodes representing partitions of the queries from the database. The neural network architecture is used to answer different query types efficiently. A generic neural network database framework can learn to answer different query types such as distance to nearest neighbor queries and range aggregate queries. The example neural database answers these two query types with orders of magnitude improvement in query time over the state-of-the-art competitions, and by constructing a model that takes only a fraction of data size.
1 . A system for answering database queries, comprising:
a query input for accepting a query of a first type of query to a database;
a decision tree including a plurality of leaf nodes, each leaf node corresponding to a partition of queries from the database;
a plurality of neural networks executed by a processor coupled to the query input, each of the neural networks trained to output a response to the first type of query corresponding to one of the partitions, wherein the training is based on a training set of queries in the partition and correct answers to the queries from the database, wherein a corresponding partition is determined by the decision tree for the query, wherein each neural network includes a plurality of layers, including a first layer and a last layer including a number of units corresponding to the output dimensionality of the first query type, and wherein batch normalization is used for all the layers, sine activation is used for the first layer and swish activation is used for the first layer and a set of intervening layers between the first layer and last layer; and
an output that outputs the output of a corresponding neural network to the determined partition in response to the query.
2 . The system of claim 1 , wherein each of the plurality of neural networks is trained by minimizing the average approximation error.
3 . The system of claim 1 , wherein each of the plurality of neural networks are retrained to respond to queries corresponding to a second query type different from the first query type.
4 . The system of claim 1 , wherein the first query type includes one of a nearest neighbor, a distance to nearest neighbor query and a range aggregate query.
5 . The system of claim 1 , wherein the partitions include a merged partition that combines two initial partitions, wherein the merged partition corresponds to one of the plurality of neural networks.
6 . The system of claim 1 , further comprising an evaluation module executed by the processor operable to evaluate the first type of query and selecting either the database or the corresponding neural network of the plurality of neural networks to answer the query.
7 . The system of claim 1 , wherein the data in the database is multi-dimensional data.
8 . A method of querying a database, comprising:
receiving a query of a first query type;
matching the query to a partition of a plurality of partitions of queries of a database, the plurality of partitions organized in a decision tree;
selecting a neural network corresponding to the matched partition from a plurality of neural networks each having an identical structure, wherein each neural network includes a plurality of layers, including a first layer and a last layer including a number of units corresponding to the output dimensionality of the first query type, and wherein batch normalization is used for all the layers, sine activation is used for the first layer and swish activation is used for the first layer and a set of intervening layers between the first layer and last layer, wherein the selected neural network is trained from a dataset of queries and answers from the database; and
activating the selected neural network to output an answer to the query.
9 . The method of claim 8 , wherein each of the plurality of neural networks is trained by minimizing the average approximation error.
10 . The method of claim 8 , further comprising retraining the plurality of neural networks to respond to queries corresponding to a second query type different from the first query type.
11 . The method of claim 8 , wherein the first query type includes one of a nearest neighbor, a distance to nearest neighbor query and a range aggregate query.
12 . The method of claim 8 , wherein the partitions include a merged partition that combines two initial partitions, wherein the merged partition corresponds to one of the plurality of neural networks.
13 . The method of claim 8 , further comprising evaluating the query and selecting either the database or the corresponding neural network of the plurality of neural networks to answer the query.
14 . A non-transitory computer readable medium having stored thereon software instructions that, when executed by a processor, cause the processor to:
receive a query of a first query type;
match the query to a partition of a plurality of partitions of queries of the database, the plurality of partitions organized in a decision tree;
select a neural network corresponding to the matched partition from a plurality of neural networks each having an identical structure, wherein each neural network includes a plurality of layers, including a first layer and a last layer including a number of units corresponding to the output dimensionality of the first query type, and wherein batch normalization is used for all the layers, sine activation is used for the first layer and swish activation is used for the first layer and a set of intervening layers between the first layer and last layer, wherein the selected neural network is trained from a dataset of queries and answers from the database; and
activate the selected neural network to output an answer to the query.