Execution of trained neural networks using a database system

A network traffic manager apparatus for efficient execution of a trained neural network, the trained neural network comprising a plurality of layers each comprising a plurality of weight values and bias values and programmed at each of the layers to execute an affine transformation of an activation function and an input data, the network traffic manager apparatus, comprising a database and a memory, the memory comprising programmed instructions stored in the memory and one or more processors configured to be capable of executing the programmed instructions stored in the memory to:

• execute, with the one or more processors, the trained neural network based on the input data as follows, the input data being stored in the database in columnar format comprising one or more rows and one or more columns, wherein each of the one or more rows corresponds to a respective input set, and each of the one or more columns corresponds to a respective element of the respective input set,
• at a first layer of the trained neural network:
• divide the affine transformation into a plurality of transformation pieces based on one or more predetermined factors, to generate one or more affine transformation queries for each transformation piece, each of the transformation pieces applying one of the plurality of weight values to a portion of the input data, the one or more predetermined factors comprising a desired query length;
• execute the generated one or more affine transformation queries in parallel by applying an operation to each of the one or more columns in parallel to obtain computed pieces, and write the computed pieces to a first temporary database table of the database;
• generate one or more activation function queries to combine the computed pieces and apply the activation function to the combined computed pieces to generate a set of output data;
• write, in columnar format, the output data to one of a plurality of different second database tables of the database that respectively correspond to the layers, and subsequently delete the first temporary database table;
• at each subsequent layer of the trained neural network, use the output data of one or more previous layers as input data and repeat the divide, execute, combine, apply, and write operations;
• and
• display the output data via a user interface of an application program.

A database system configured to execute a trained neural network, the trained neural network comprising a plurality of layers each comprising a plurality of weight values and bias values and programmed at each of the layers to execute an affine transformation of an activation function and input data, the database system comprising traffic management apparatuses, client devices, or server devices, the database system comprising memory comprising programmed instructions stored thereon and processors configured to be capable of executing the stored programmed instructions to:

• execute, with the processors, the trained neural network based on the input data as follows, the input data being stored in the database system in columnar format comprising one or more rows and one or more columns, wherein each of the one or more rows corresponds to a respective input set, and each of the one or more columns corresponds to a respective element of the respective input set,
• at a first layer of the trained neural network:
• divide the affine transformation into a plurality of transformation pieces based on one or more predetermined factors, to generate one or more affine transformation queries for each transformation piece, each of the transformation pieces applying one of the plurality of weight values to a portion of the input data, the one or more predetermined factors comprising a desired query length;
• execute the generated one or more affine transformation queries in parallel by applying an operation to each of the one or more columns in parallel to obtain computed pieces, and write the computed pieces to a first temporary database table of the database system;
• generate one or more activation function queries to combine the computed pieces and apply the activation function to the combined computed pieces to generate a set of output data;
• write, in columnar format, the output data to one of a plurality of different second database tables of the database system that respectively correspond to the layers, and subsequently delete the first temporary database table;
• at each subsequent layer of the trained neural network, use the output data of one or more previous layers as input data and repeat divide, execute, combine, apply, and write operations;
• and
• display the output data via a user interface of an application program.

A non-transitory computer readable medium having stored thereon instructions for efficient execution of a trained neural network, the trained neural network comprising a plurality of layers each comprising a plurality of weight values and bias values and programmed at each of the layers to execute an affine transformation of an activation function and input data, comprising executable code which when executed by processors, causes the processors to:

• execute, with the processors, the trained neural network based on the input data as follows, the input data being stored in a database system in columnar format comprising one or more rows and one or more columns, wherein each of the one or more rows corresponds to a respective input set, and each of the one or more columns corresponds to a respective element of the respective input set,
• at a first layer of the trained neural network:
• divide the affine transformation into a plurality of transformation pieces based on one or more predetermined factors, to generate one or more affine transformation queries for each transformation piece, each of the transformation pieces applying one of the plurality of weight values to a portion of the input data, the one or more predetermined factors comprising a desired query length;
• execute the generated one or more affine transformation queries in parallel by applying an operation to each of the one or more columns in parallel to obtain computed pieces, and writing the computed pieces to a first temporary database table of the database;
• generate one or more activation function queries to combine the computed pieces and apply the activation function to the combined computed pieces to generate a set of output data;
• write, in columnar format, the output data to one of a plurality of different second database tables of the database that respectively correspond to the layers, and subsequently delete the first temporary database table;
• at each subsequent layer of the trained neural network, use the output data of one or more previous layers as input data and repeat divide, execute, combine, apply, and write operations;
• and
• display the output data via a user interface of an application program.