D5AI LLC

Computer systems and computer-implemented methods modify a machine learning network, such as a deep neural network, to introduce judgment to the network. A “combining” node is added to the network, to thereby generate a modified network, where activation of the combining node is based, at least in part, on output from a subject node of the network. The computer system then trains the modified network by, for each training data item in a set of training data, performing forward and back propagation computations through the modified network, where the backward propagation computation through the modified network comprises computing estimated partial derivatives of an error function of an objective for the network, except that the combining node selectively blocks back-propagation of estimated partial derivatives to the subject node, even though activation of the combining node is based on the activation of the subject node.

Computer systems and computer-implemented methods modify a machine learning network, such as a deep neural network, to introduce judgment to the network. A “combining” node is added to the network, to thereby generate a modified network, where activation of the combining node is based, at least in part, on output from a subject node of the network. The computer system then trains the modified network by, for each training data item in a set of training data, performing forward and back propagation computations through the modified network, where the backward propagation computation through the modified network comprises computing estimated partial derivatives of an error function of an objective for the network, except that the combining node selectively blocks back-propagation of estimated partial derivatives to the subject node, even though activation of the combining node is based on the activation of the subject node.

Computer systems and computer-implemented methods train a neural network, by: (a) computing for each datum in a set of training data, activation values for nodes in the neural network and estimates of partial derivatives of an objective function for the neural network for the nodes in the neural network; (b) selecting a target node of the neural network and/or a target datum in the set of training data; (c) selecting a target-specific improvement model for the neural network, wherein the target-specific improvement model, when added to the neural network, improves performance of the neural network for the target node and/or the target datum, as the case may be; (d) training the target-specific improvement model; (e) merging the target-specific improvement model with the neural network to form an expanded neural network; and (f) training the expanded neural network.

Computer systems and computer-implemented methods train a neural network, by: (a) computing for each datum in a set of training data, activation values for nodes in the neural network and estimates of partial derivatives of an objective function for the neural network for the nodes in the neural network; (b) selecting a target node of the neural network and/or a target datum in the set of training data; (c) selecting a target-specific improvement model for the neural network, wherein the target-specific improvement model, when added to the neural network, improves performance of the neural network for the target node and/or the target datum, as the case may be; (d) training the target-specific improvement model; (e)merging the target-specific improvement model with the neural network to form an expanded neural network; and (f) training the expanded neural network.

A system and method for controlling a nodal network. The method includes estimating an effect on the objective caused by the existence or non-existence of a direct connection between a pair of nodes and changing a structure of the nodal network based at least in part on the estimate of the effect. A nodal network includes a strict partially ordered set, a weighted directed acyclic graph, an artificial neural network, and/or a layered feed-forward neural network.

Data-dependent node-to-node knowledge sharing to increase the interpretability of the activation pattern of one or more nodes in a neural network, is implemented by a set of knowledge sharing links. Each link may comprise a knowledge providing node or other source P and a knowledge receiving node R. A knowledge sharing link can impose a node-specific regularization on the knowledge receiving node R to help guide the knowledge receiving node R to have an activation pattern that is more easily interpreted. The specification and training of the knowledge sharing links may be controlled by a cooperative human-AI learning supervisor system in which a human and an artificial intelligence system work cooperatively to improve the interpretability and performance of the client system.

Computer systems and computer-implemented methods modify a machine learning network, such as a deep neural network, to introduce judgment to the network. A “combining” node is added to the network, to thereby generate a modified network, where activation of the combining node is based, at least in part, on output from a subject node of the network. The computer system then trains the modified network by, for each training data item in a set of training data, performing forward and back propagation computations through the modified network, where the backward propagation computation through the modified network comprises computing estimated partial derivatives of an error function of an objective for the network, except that the combining node selectively blocks back-propagation of estimated partial derivatives to the subject node, even though activation of the combining node is based on the activation of the subject node.

Computer systems and methods modify a base deep neural network (DNN). The method comprises replacing the target node of the base DNN with a compound node to thereby create a modified base DNN. The compound node comprises at least first and second nodes. The first node is trained to detect target node patterns in inputs to the first node and the second node is trained to detect an absence of the target node patterns in inputs to the second node, and the first and second nodes are trained to be non-complementary. Replacing the target node with the compound node comprises: connecting the first node to the upper sub-network of the base DNN, such that the first node has a weighted connection for each of the one or more base connection that the target node had to the upper sub-network in the base DNN; and connecting the second node to the upper sub-network of the base deep DNN, such that the second node has a weighted connection for each of the one or more base connection that the target node had to the upper sub-network in the base deep DNN. After replacing the target node, the modified base DNN is trained.

Computer systems and computer-implemented methods modify a machine learning network, such as a deep neural network, to introduce judgment to the network. A “combining” node is added to the network, to thereby generate a modified network, where activation of the combining node is based, at least in part, on output from a subject node of the network. The computer system then trains the modified network by, for each training data item in a set of training data, performing forward and back propagation computations through the modified network, where the backward propagation computation through the modified network comprises computing estimated partial derivatives of an error function of an objective for the network, except that the combining node selectively blocks back-propagation of estimated partial derivatives to the subject node, even though activation of the combining node is based on the activation of the subject node.

Systems and methods improve performance of a classifier, which comprises a neural network and is trained through machine learning. First and second scores are computed, by the classifier, for each a multiple data examples from a generator. The first score is indicative of whether the data example belongs to a first data cluster and the second score is indicative of whether the data example belongs to a second data cluster. The generator is trained with an objective such that, for each data example generated by the generator, the first and second scores computed by the classifier are equal. Partial derivatives from the classifier are back-propagated for multiple data examples generated by the generator, to obtain a vector, for each data example, that is orthogonal to a decision surface for the classifier. A problem with the classifier is detected based on changes in directions of the vectors. Upon detecting a problem, the classifier is adjusted to reduce errors by the classifier caused by overfitting training data.