Approximation-Free Neural Network Mapping for Spiking Conversion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural networks struggle with efficient mapping and conversion from pre-trained analog-signal-based neural networks to temporal-coding-based spiking neural networks, which affects task accuracy and latency.
Innovation Solution
A computer-implemented method that accesses pre-trained analog-signal-based neural network components, scales them, and generates temporal-coding-based neural network components, ensuring that the converted spiking neural network maintains equivalent performance to the original network.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If approximation-based mapping methods are used to convert analog-signal-based neural networks to spiking neural networks, then conversion complexity is reduced, but task accuracy deteriorates due to approximation errors
Solution Approach 1:
The patent creates an exact copy of the analog neural network's computational behavior in the spiking neural network domain through precise mathematical mapping. The mapping functions replicate the analog network's weight matrices, bias vectors, and activation functions without approximation, ensuring identical task accuracy while adapting to spiking network constraints
Solution Approach 2:
The patent transforms parameters from the analog domain to the spiking domain through systematic parameter changes. Weight matrices W and bias vectors b are mapped to spiking equivalents W' and b' using defined mathematical relationships, while activation functions f are transformed to spiking activation functions f', maintaining functional equivalence without approximation
2Productivity
If direct conversion from analog-signal-based neural networks to spiking neural networks is performed without scaling, then computational efficiency is improved, but mapping accuracy deteriorates
Solution Approach 1:
The patent performs preliminary scaling of the analog neural network parameters before conversion to the spiking domain. Weight matrices and bias vectors are scaled by predetermined factors to ensure compatibility with spiking network dynamics, and this preprocessing step is crucial for achieving both high mapping accuracy and computational efficiency in the converted network
3Use of energy by moving object
If temporal coding is introduced to convert analog neural networks to spiking neural networks, then energy efficiency is improved, but conversion complexity increases
Solution Approach 1:
The patent substitutes the continuous analog signal processing mechanism with a discrete temporal coding mechanism. Instead of using continuous voltage signals as in analog networks, the conversion employs discrete spike events with temporal encoding, where information is conveyed through the timing and pattern of spikes rather than continuous signal amplitudes, achieving energy efficiency inherent to spiking networks
Data Source
AI summary
A computer-implemented method includes accessing neural network (NN) components of a pre-trained analog-signal-based NN. The NN components are processed to generate scaled NN components. Based at least in part on one or more of the scaled NN components, temporal-coding-based NN components of a temporal-coding-based NN are generated. The temporal-coding-based NN components are dependent on the one or more of the scaled NN components.