Regularized neural network architecture search

• Main Authors: AGGARWAL, ALOK, LE QUOC V, REAL ESTEBAN ALBERTO, HUANG YANPING
• Format: Patent
• Language: Chinese; English
• Published: 14.06.2024
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; PHYSICS

The invention relates to a regularized neural network architecture search. A method is described for receiving training data for training a neural network (NN) to perform machine learning tasks and for determining an optimized NN architecture for performing ML tasks using the training data. Determining the optimized NN architecture comprises: maintaining, for each candidate architecture in the population of candidate architectures, overall data comprising (i) data defining the candidate architecture, and (ii) data specifying how to recently train a neural network having the candidate architecture in determining the optimized neural network architecture; and repeatedly performing a plurality of operations using each of the plurality of worker computing units to generate a new candidate architecture based on the selected candidate architecture having the best fit measure, adding the new candidate architecture to the population, and removing the candidate architecture that is least recently trained from the popu