Spiking Neural Network Training via Membrane Potential and Lateral Inhibition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
There is no effective method for training multi-layer deep spiking neural networks, which are expected to be more efficient due to their operation scheme similar to the biological brain, as they lack a systematic training approach.
Innovation Solution
A recognition apparatus and training method are introduced, featuring a neural network with neurons activated by first and second synaptic signals, where an activation neuron generates signals to excite or inhibit next-layer neurons and adjusts membrane potentials based on lateral inhibition coefficients, and a processor determines parameters like membrane potential thresholds and synaptic weights using a gradient descent scheme.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a multi-layer deep spiking neural network is used to improve recognition efficiency similar to biological brain operation, then recognition efficiency is improved, but training capability deteriorates because no training method exists
Solution Approach 1:
The patent introduces a membrane potential as an intermediary variable that mediates between synaptic inputs and neuron activation. This membrane potential mechanism enables the spiking neural network to accumulate and integrate signals over time, providing a trainable parameter that bridges the gap between biological-inspired operation and systematic training capability
Solution Approach 2:
The patent changes the operational parameters of the spiking neural network by introducing membrane potential thresholds and time constants as trainable parameters. This allows the network to adapt its activation characteristics through training, transforming the fixed biological-inspired model into a flexible, trainable system
2Measurement precision
If lateral inhibition signals are introduced to enable winner-take-all competition among neurons, then recognition precision is improved, but network complexity increases
Solution Approach 1:
The patent merges the lateral inhibition mechanism with the standard feedforward propagation by having neurons generate both excitatory outputs to the next layer and inhibitory outputs to同级 neurons. This unified approach enables competition without requiring separate inhibition networks, reducing overall system complexity
Solution Approach 2:
The patent creates a competitive equilibrium among neurons through lateral inhibition, where neurons with stronger inputs drive others toward their inhibition thresholds. This equipotential mechanism ensures that only the most activated neurons (winners) remain active, achieving precise recognition through balanced competition
Data Source
AI summary
A recognition apparatus and a training method are provided. The recognition apparatus includes a memory configured to store a neural network including a previous layer of neurons, and a current layer of neurons that are activated based on first synaptic signals and second synaptic signals, the first synaptic signals being input from the previous layer, and the second synaptic signals being input from the current layer. The recognition apparatus further includes a processor configured to generate a recognition result based on the neural network. An activation neuron among the neurons of the current layer generates a first synaptic signal to excite or inhibit neurons of a next layer, and generates a second synaptic signal to inhibit neurons other than the activation neuron in the current layer.