System and method for training neural networks with sparse data

• Main Authors: HASSELGREN JOHN NIELS THEODOOR, LEHTINEN JAAKKO T, AILA TIMO OSKARI, MUNKBERG CARL J
• Format: Patent
• Language: Chinese; English
• Published: 02.05.2023
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; IMAGE DATA PROCESSING OR GENERATION, IN GENERAL; PHYSICS

The invention relates to a system and method for training a neural network with sparse data. Methods, computer readable media, and systems for training neural network models are disclosed. The method comprises the step of selecting an input vector from a set of training data comprising the input vector and sparse target vectors, wherein each sparse target vector comprises target data corresponding to a subset of samples within an output vector of a neural network model. The method further includes the steps of processing the input vector by the neural network model to generate output data for samples within the output vector, and adjusting parameter values of the neural network model to reduce a difference between the output vector and the sparse target vector for the subset of samples. 本申请涉及用稀疏数据训练神经网络的系统和方法。公开了用于训练神经网络模型的方法、计算机可读介质和系统。所述方法包括步骤：从包括输入向量和稀疏目标向量的一组训练数据中选择输入向量，其中每个稀疏目标向量包括对应于神经网络模型的输出向量内的样本子集的目标数据。所述方法还包括步骤：通过神经网络模型处理输入向量以针对输出向量内的样本产生输出数据，以及调整神经网络模型的参数值以针对样本子集减小输出向量与稀疏目标向量之间的差。