Torrent Systems, Inc.

Anthony Passera

Michael James Beckerle

Edward S Zyszkowski

John R Thorp

A computer system splits a data space to partition data between processors or processes. The data space may be split into sub-regions which need not be orthogonal to the axes defined the data space's parameters, using a decision tree. The decision tree can have neural networks in each of its non-terminal nodes that are trained on, and are used to partition, training data. Each terminal, or leaf, node can have a hidden layer neural network trained on the training data that reaches the terminal node. The training of the non-terminal nodes' neural networks can be performed on one processor and the training of the leaf nodes' neural networks can be run on separate processors. Different target values can be used for the training of the networks of different non-terminal nodes. The non-terminal node networks may be hidden layer neural networks. Each non-terminal node automatically may send a desired ratio of the training records it receives to each of its child nodes, so the leaf node networks each receives approximately the same number of training records. The system may automatically configures the tree to have a number of leaf nodes equal to the number of separate processors available to train leaf node networks. After the non-terminal and leaf node networks have been trained, the records of a large data base can be passed through the tree for classification or for estimation of certain parameter values.

Computer system and computerized method for partitioning data for parallel processing

A database often contains sparse, i.e., under-represented, conditions which might be not represented in a training data set for training an analytical model if the training data set is created by stratified sampling. Sparse conditions may be represented in a training set by using a data set which includes essentially all of the data in a database, without stratified sampling. A series of samples, or &ldquo;windows,&rdquo; are used to select portions of the large data set for phases of training. In general, the first window of data should be a reasonably broad sample of the data. After the model is initially trained using a first window of data, subsequent windows are used to retrain the model. For some model types, the model is modified in order to provide it with some retention of training obtained using previous windows of data. Neural networks and Kohonen networks may be used without modification. Models such as probabilistic neural networks, generalized regression neural networks, Gaussian radial basis functions, decision trees, including K-D trees and neural trees, are modified to provide them with properties of memory to retain the effects of training with previous training data sets. Such a modification may be provided using clustering. is Parallel training models which partition the training data set into disjoint subsets are modified so that the partitioner is trained only on the first window of data, whereas subsequent windows are used to train the models to which the partitioner applies the data in parallel.

Computer system and process for training of analytical models using large data sets

The present invention provides a description of the behavior of a model that indicates the sensitivity of the model in subspaces of the input space and which indicates which dimensions of the input data are salient in subspaces of the input space. By implementing this description using a decision tree, the subspaces and their salient dimensions are both described and determined hierarchically. A sensitivity analysis is performed on the model to provide a sensitivity profile of the input space of the model according to sensitivity of outputs of the model to variations in data input to the model. The input space is divided into at least two subspaces according to the sensitivity profile. A sensitivity analysis is performed on the model to provide a sensitivity profile of each of the subspaces according to sensitivity of outputs of the model to variations in data input to the model.

Computer system and process for explaining behavior of a model that maps input data to output data

A computer system splits a data space to partition data between processors or processes. The data space may be split into sub-regions which need not be orthogonal to the axes defined by the data space's parameters, using a decision tree. The decision tree can have neural networks in each of its non-terminal nodes that are trained on, and are used to partition, training data. Each terminal, or leaf, node can have a hidden layer neural network trained on the training data that reaches the terminal node. The training of the non-terminal nodes' neural networks can be performed on one processor and the training of the leaf nodes' neural networks can be run on separate processors. Different target values can be used for the training of the networks of different non-terminal nodes. The non-terminal node networks may be hidden layer neural networks. Each non-terminal node automatically may send a desired ratio of the training records it receives to each of its child nodes, so the leaf node networks each receives approximately the same number of training records. The system may automatically configures the tree to have a number of leaf nodes equal to the number of separate processors available to train leaf node networks. After the non-terminal and leaf node networks have been trained, the records of a large data base can be passed through the tree for classification or for estimation of certain parameter values.

Growing community of inventors

Watertown, MA, United States of America

Anthony Passera