Chip-in-loop proxy training method and device, chip and electronic device

Abstract

The invention discloses a chip-in-loop proxy training method and device, a chip and an electronic device. In order to solve various error problems existing when the SNN network is deployed to the chip, the chip is incorporated into the SNN network training process, and quantization errors are eliminated. According to the method and the device, chip in-loop training is adopted, and the ANN network in the training equipment and the SNN network in the chip share the network configuration data, so that the technical problems of conversion errors, quantization errors and the like when the SNN network is deployed to the chip are solved, and the technical effect that the SNN network deployed to the chip can also obtain high reasoning precision is achieved. The method is suitable for the fields of brain-like chip, AIoT and spiking neural network training.

Description

technical field [0001] The invention relates to a chip-in-loop agent training method and device, a chip and an electronic device, in particular to an artificial neural network (ANN) as a proxy (proxy) and a chip-in-loop (chip-in-loop) spiking neural network (SNN) ) training methods and equipment, chips and electronic equipment. Background technique [0002] In the post-artificial intelligence era, how to realize edge intelligence with low power consumption is an important topic of academic and commercial attention, and the bionic Spiking Neural Networks (SNN) is one of the most promising research directions. How to efficiently train (also known as learning) a high-precision spiking neural network is an extremely important research content in this field. [0003] At present, the methods of training SNN in the existing technology mainly include the following categories: [0004] 1) ANN to SNN: that is, train an ANN (Artificial Neural Networks) and then convert it to SNN. Th...

AI Technical Summary

Problems solved by technology

[0011] The applicant found that although the accuracy of the SNN network trained by prior art 2 is excellent, it is still difficult to obtain equivalent precision performance on the chip due to the existence of the aforementioned quantization error when the trained SNN is actually deployed to the chip.

Images