Time-to-First-Spike Learning With Neuron-Indexed Reference Times
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing spiking neural networks face challenges in preventing neuron instability due to small membrane potential thresholds, leading to inefficiencies in information transmission and reduced estimation accuracy.
Innovation Solution
A learning device and method that utilize time-to-first-spike coding with a reference time and neuron index to adjust firing times and membrane potentials, optimizing the learning process for spiking neural networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the membrane potential threshold is set to a small value to enable precise neural coding, then the coding precision is improved, but neuron instability occurs and information transmission efficiency deteriorates
Solution Approach 1:
The patent changes the parameter of membrane potential threshold from a fixed small value to a dynamically adjusted value based on the firing time of neurons. By adjusting the threshold parameter according to the timing information, the system maintains both coding precision and neuron stability, resolving the contradiction between precise coding and neural reliability.
Solution Approach 2:
The patent introduces a feedback mechanism where the evaluation function uses the actual firing time of neurons to adjust the membrane potential threshold. This feedback loop ensures that neurons fire at appropriate times while maintaining stability, thereby improving both coding precision and neural reliability simultaneously.
2Productivity
If the membrane potential threshold is lowered to increase neuron firing probability, then information transmission is improved, but estimation accuracy deteriorates due to excessive firing
Solution Approach 1:
The patent makes the membrane potential threshold dynamic rather than static. The threshold adapts based on the firing time of neurons, allowing the system to optimize both information transmission efficiency and estimation accuracy. This dynamic adjustment prevents excessive firing while maintaining appropriate neural activity levels.
Solution Approach 2:
The patent changes the threshold parameter dynamically based on firing time information. By adjusting the threshold parameter according to when neurons fire, the system achieves both efficient information transmission and accurate estimation, resolving the contradiction between productivity and measurement precision.
3Reliability
If synaptic weight is increased to compensate for small membrane potential, then neuron stability is improved, but learning efficiency deteriorates
Solution Approach 1:
The patent introduces a feedback mechanism where the evaluation function uses actual firing times to adjust synaptic weights dynamically. This feedback loop allows the system to maintain neuron stability while improving learning efficiency, as the synaptic weights are adjusted based on actual neural behavior rather than being fixed.
Solution Approach 2:
The patent makes synaptic weights dynamic by adjusting them based on firing time information. This dynamic adjustment allows the system to maintain both neuron stability and learning efficiency, resolving the contradiction between reliability and productivity.
Data Source
AI summary
A learning device performs learning a spiking neural network using time-to-first-spike coding, by using an evaluation function including a predetermined reference time given to each layer of the spiking neural network and an index related to neurons.