Spiking Neuron Reinforcing Circuit for Single Event Effect Tolerance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Spiking neurons in spiking neural networks are prone to computing errors due to susceptibility to single event effects, leading to incorrect spiking signals and output results in spatial applications.
Innovation Solution
A spiking neuron reinforcing circuit and method that includes coding and decoding configuration information and model parameters using error checking and correcting (ECC) modules, storing backup data in a shadow memory, and selecting error-corrected data to configure the spiking neuron, thereby improving error tolerance and adaptability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If spiking neuron uses standard storage without error correction, then device complexity is low, but reliability deteriorates due to single event effects causing computing errors
Solution Approach 1:
The patent applies preliminary action by performing error correction coding on configuration information and model parameters before they are stored in registers. The coding module encodes data in advance, and the decoding module decodes and corrects errors before the data is used by the spiking neuron, preventing single event effects from causing computing errors.
Solution Approach 2:
The patent implements copying by creating backup copies of configuration information and model parameters through shadow registers. These shadow registers store redundant copies of the data, allowing the system to compare and correct errors by switching between primary and shadow copies when errors are detected.
2Reliability
If spiking neuron implements error correction coding, then reliability improves, but device complexity increases due to additional coding and decoding modules
Solution Approach 1:
The patent applies segmentation by dividing the storage system into multiple independent register groups: first register group for coded configuration information, second register group for coded model parameters, and shadow register groups for backup copies. This segmentation allows error correction to be applied independently to different data types and facilitates selective recovery of corrupted data.
3Adaptability or versatility
If spiking neuron stores backup data in shadow memory, then adaptability improves, but loss of time increases due to error checking and data replacement operations
Solution Approach 1:
The patent implements self-service by enabling the spiking neuron system to automatically detect and correct its own errors without external intervention. The decoding module autonomously checks for errors in configuration information and model parameters, and the logic module automatically replaces corrupted data with correct copies from shadow registers, allowing continuous operation in spatial environments.
Data Source
AI summary
A spiking neuron reinforcing circuit and a spiking neuron reinforcing method are provided.