Spiking Neural Network Neuron Threshold Compensation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Spiking neural networks (SNNs) face inaccuracies in inference due to variations in threshold voltages among neurons when implemented in hardware, leading to inconsistent input and output characteristics.
Innovation Solution
A method is introduced to compensate for threshold variations by adjusting the effective threshold for subsequent firings to a target threshold, ensuring each neuron discharges the same amount of membrane potential after firing, thereby maintaining consistent threshold values across all neurons.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If threshold voltage of each neuron is allowed to vary in hardware implementation, then manufacturing complexity is reduced, but inference accuracy deteriorates due to inconsistent input and output characteristics
Solution Approach 1:
The patent applies parameter changes by dynamically adjusting the threshold voltage parameter of each neuron based on its actual hardware characteristics. Each neuron's threshold voltage is modified to compensate for manufacturing variations, ensuring that all neurons operate with consistent effective thresholds despite hardware differences. This resolves the contradiction by allowing flexible hardware manufacturing while maintaining precise inference through parameter adaptation.
2Ease of manufacture
If threshold voltage is uniformly set for all neurons, then manufacturing is simplified, but neuron performance deteriorates due to inability to compensate for device characteristic distribution
Solution Approach 1:
The patent implements local quality by assigning individualized threshold voltage adjustments to each neuron based on its specific device characteristics. Instead of a uniform threshold for all neurons, each neuron receives a customized threshold compensation that accounts for its unique hardware properties. This ensures consistent firing behavior across all neurons while maintaining manufacturing simplicity through automated characterization and adjustment processes.
3Measurement precision
If membrane potential is discharged by different amounts for each neuron, then threshold variation is compensated, but calculation complexity increases due to continuous variation calculation requirements
Solution Approach 1:
The patent applies preliminary action by pre-calculating and storing the compensation values for each neuron's threshold voltage before the neural network operates. The characterization process determines the optimal discharge amounts in advance, and these pre-computed values are then used during inference. This eliminates the need for continuous complex calculations during operation, reducing real-time computational burden while maintaining accurate threshold compensation.
Solution Approach 2:
The patent implements self-service by enabling each neuron to automatically adjust its own threshold voltage based on its measured device characteristics. Each neuron performs its own threshold compensation using its specific membrane potential discharge behavior, without requiring external control or complex system-wide calculations. This self-adjustment mechanism simplifies the overall system complexity while achieving precise individualized threshold compensation.
Data Source
AI summary
The present inventive concept provides a method for compensating a neuron threshold variation for firing in a neural network apparatus. The method compensates a threshold variation adjusting an effective threshold to a target threshold in each neuron. The effective threshold is for a next firing after the firing of each neuron.