US Patent 10956818 Data splitting by gradient direction for neural networks

Systems and methods improve the performance of a network that has converged such that the gradient of the network and all the partial derivatives are zero (or close to zero) by splitting the training data such that, on each subset of the split training data, some nodes or arcs (i.e., connections between a node and previous or subsequent layers of the network) have individual partial derivative values that are different from zero on the split subsets of the data, although their partial derivatives averaged over the whole set of training data is close to zero. The present system and method can create a new network by splitting the candidate nodes or arcs that diverge from zero and then trains the resulting network with each selected node trained on the corresponding cluster of the data. Because the direction of the gradient is different for each of the nodes or arcs that are split, the nodes and their arcs in the new network will train to be different. Therefore, the new network is not at a stationary point.