Spiking Neural Network Inference with Negative Membrane Potential

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Current spiking neural networks (SNNs) do not match the performance of analog-valued neural networks (ANNs) due to limitations in information representation and processing, particularly in handling negative membrane potentials and internal delays, which result in information loss and reduced inference accuracy.

Innovation Solution

A spiking neural network is generated based on a pre-learned ANN, with artificial neurons configured to have negative membrane potentials and pre-charged membrane potentials, and output spikes are filtered after a predetermined time to minimize errors and enhance inference accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If conventional SNNs are used, then the network structure resembles biological nervous systems, but the inference accuracy is reduced due to information loss from negative membrane potentials and internal delays

Engineering Contradiction:

Improveinference accuracyVSAvoidinformation loss from negative membrane potentials

Core Design Contradiction:

ReliabilityVSLoss of information

Solution Approach 1:

The patent changes the membrane potential parameter to allow negative values, enabling the SNN to accurately represent information without loss. By permitting negative membrane potentials, the network can capture the full range of signal intensities and temporal patterns, resolving the information loss issue that limited inference accuracy in conventional SNNs.

Inventive Principle:

Principle #35Parameter changes

2Productivity

If SNNs operate with real-time spikes, then the processing speed is fast, but internal delays cause errors in inference results

Engineering Contradiction:

Improveprocessing speedVSAvoidinference result accuracy

Core Design Contradiction:

ProductivityVSMeasurement precision

Solution Approach 1:

The patent applies preliminary action by pre-charging the membrane potential before the input signal is received. This preliminary charging compensates for internal delays and ensures that the neuron is in the optimal state to process the incoming signal accurately, maintaining both fast processing speed and high inference accuracy.

Inventive Principle:

Principle #10Preliminary action

3Adaptability or versatility

If SNNs use spike-based information representation, then the network is more biologically realistic, but the performance lags behind ANNs

Engineering Contradiction:

Improvebiological realismVSAvoidinference performance

Core Design Contradiction:

Adaptability or versatilityVSReliability

Solution Approach 1:

The patent enhances the dynamics of the SNN by allowing membrane potentials to vary continuously including negative values, and by implementing pre-charging mechanisms. These dynamic adjustments enable the SNN to adapt its behavior to match ANN performance while preserving biological realism through spike-based representation.

Inventive Principle:

Principle #15Dynamics

Data Source

PatentUS20240419969A1Inference method and device using spiking neural network

Publication Date: 2024.12.19 SAMSUNG ELECTRONICS CO LTD

AI summary

Embodiments relate to an inference method and device using a spiking neural network including parameters determined using an analog-valued neural network (ANN). The spiking neural network used in the inference method and device includes an artificial neuron that may have a negative membrane potential or have a pre-charged membrane potential. Additionally, an inference operation by the inference method and device is performed after a predetermined time from an operating time point of the spiking neural network.