Title：On-chip learning neuromorphic chip and training method

Country：China

Patent No.：202110953844.8

Legal Status：Under review

Inventor：Hong Chen, Jilin Zhang

Assignee：Tsinghua University

Address：Tsinghua University,Haidian District Beijing 100094, China

Filing Date：2021-08-19

Issue Date：

The invention provides an on-chip neural morphology learning chip and a training method. According to the technical scheme, a data buffer outputs a training sample to a neural network; the neural network is used for training an input training sample, and inputting an intermediate pulse generated in the training process into the pulse buffer, so that the pulse buffer stores the intermediate pulse input by the neural network after learning a neuromorphic chip on a training state machine control chip to enter and exit a training window, storage of the intermediate pulse input by the neural network is stopped when leaving the training window, and the stored intermediate pulse is input to a weight updating module; meanwhile, the neural network outputs a result pulse obtained by training to an error calculation module, so that the error calculation module calculates an error according to the result pulse and outputs an error calculation result to a weight updating module; and a weight updating module calculates a weight variation according to the intermediate pulse and the error calculation result, and performs weight updating on the neural network according to the weight variation. The training energy consumption can be reduced.

Patent Certificate： PDF/Jpg