MICROTRAINING FOR ITERATIVE REFINEMENT OF A NEURAL NETWORK WITH FEW ADAPTATIONS

摘要

The disclosed microtraining methods improve the accuracy of trained neural networks by performing iterative refinement at low learning rates using a relatively short series of microtraining steps. A framework for training neural networks receives the trained neural network together with a second training data set and a set of hyperparameters. The framework for training neural networks generates a microtrained neural network by adapting one or more weights of the trained neural network at a lower learning rate in order to enable incremental improvements in accuracy without significantly changing the computation structure of the trained neural network. The microtrained neural network can be evaluated for changes in accuracy and / or quality. Additional microtraining sessions can be performed with the microtrained neural network to further improve accuracy or quality.