Spiking Neural Network Weight Quantization for Resource-Constrained Devices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Spiking neural networks face accuracy degradation when deployed on computation-resource-lacking devices due to quantization of data with lower bit depth, leading to reduced performance compared to resource-rich devices where they were trained.
Innovation Solution
The method involves training a spiking neural network in a computation-resource-rich device to obtain quantized weight values and deploying it in a resource-lacking device with updated weight values, using equations to calculate and update weight values from M bits to N bits, and adjusting the threshold value accordingly to maintain accuracy and reduce power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by stationary object
If weight quantization is applied to reduce bit depth from M bits to N bits for deployment on resource-lacking devices, then power consumption is reduced and device compatibility is improved, but neural network accuracy degrades
Solution Approach 1:
The patent applies preliminary action by retraining the neural network with quantized weight values before deployment. The method calculates quantized weight values from high-precision training weights, then performs retraining using these quantized values to adjust and optimize the network parameters in advance, ensuring accuracy is maintained before the model is deployed on resource-constrained devices.
Solution Approach 2:
The patent implements parameter changes by systematically adjusting the weight values from high-precision (M bits) to low-precision (N bits) representations. The method calculates quantized weight values using specific equations that map high-precision weights to low-precision equivalents, then uses these quantized parameters throughout the network to achieve both reduced power consumption and maintained accuracy.
2Adaptability or versatility
If weight quantization is applied to deploy neural networks on computation-resource-lacking devices, then device compatibility and ease of deployment are improved, but neural network performance is reduced
Solution Approach 1:
The patent applies preliminary action by retraining the neural network with quantized weight values before deployment. The method calculates quantized weight values from high-precision training weights, then performs retraining using these quantized values to adjust and optimize the network parameters in advance, ensuring accuracy is maintained before the model is deployed on resource-constrained devices.
Solution Approach 2:
The patent implements feedback by using the performance difference between high-precision and quantized weight representations to guide the retraining process. The method calculates quantized weight values, retrains the network with these values, and uses the resulting performance characteristics to optimize the quantization parameters, creating a feedback loop that improves both compatibility and performance.
Data Source
AI summary
A method is provided and includes operations as below: training a spiking neural network (SNN) in a first device to generate multiple first weight values of M bits; calculating multiple second weight values of N bits corresponding to the first weight values according to a threshold value, the number M, and the first weight values, wherein the number N is smaller than the number M; retraining the spiking neural network with the second weight values to update the second weight values; and performing a write operation to save the updated plurality of second weight values in a memory in a second device for performing a spiking neural network operation in the second device.