Spiking Neural Network Weight Updates Without Time Measurement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional spiking neural networks face challenges in learning new information with low spike density, leading to deterioration of inference accuracy due to unchanged synaptic weights when no spikes are input, and require large circuit scales for measuring time differences.
Innovation Solution
A neural network device that updates synaptic weights using Spike Driven Synaptic Plasticity (SDSP), where synaptic weights are potentiated or attenuated based on the internal potential of post-neurons, eliminating the need for time difference measurement, allowing for a smaller circuit configuration and enabling learning with high accuracy even with low spike density.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If Spike Timing Dependent Plasticity (STDP) is used to implement spiking neural networks, then learning operation efficiency is improved, but circuit scale increases due to the need for time difference measurement mechanisms
Solution Approach 1:
The patent extracts the time difference measurement mechanism from the STDP implementation and replaces it with a simplified potential-based mechanism. By removing the complex timing measurement circuitry and retaining only the essential synaptic weight update functionality driven by post-neuron potential, the circuit scale is significantly reduced while maintaining learning efficiency.
Solution Approach 2:
The patent changes the parameter used for synaptic weight update from time difference (dt) to post-neuron internal potential (Vmem). This parameter substitution fundamentally simplifies the measurement mechanism, eliminating the need for precise time difference measurement circuits while preserving the core learning functionality of STDP.
2Loss of energy
If conventional spiking neural networks process low spike density inputs, then energy consumption is reduced, but inference accuracy deteriorates due to unchanged synaptic weights
Solution Approach 1:
The patent introduces a feedback mechanism where the post-neuron's internal potential continuously influences synaptic weight updates. Even during low spike density conditions, this feedback ensures that synaptic weights remain dynamic and adaptable, preventing the deterioration of inference accuracy while maintaining low energy consumption characteristic of spiking neural networks.
Solution Approach 2:
The patent implements preliminary synaptic weight updates based on post-neuron potential changes that occur even before explicit spike events. This preliminary action ensures that synaptic weights are continuously optimized during low spike density periods, maintaining inference accuracy without requiring high spike rates or additional energy input.
3Adaptability or versatility
If synaptic weights are updated using time difference measurement, then Spike Timing Dependent Plasticity is achieved, but device complexity increases
Solution Approach 1:
The patent replaces the mechanical time difference measurement system with an electrical potential-based system. Instead of measuring time intervals using complex timing circuits, the system uses the post-neuron's membrane potential as a direct indicator for synaptic weight updates, significantly simplifying the measurement mechanism while preserving synaptic plasticity capability.
Data Source
AI summary
A neural network device according to an embodiment includes a plurality of neuron circuits and a plurality of synapse circuits. In a case where an internal state value of a second neuron circuit is larger than a set determination reference value, a first synapse circuit among the synapse circuits executes potentiation processing to increase the degree of influence of a synaptic weight on a synapse signal in response to acquiring the input spike from the first neuron circuit. In response to outputting an output spike being a spike signal from the second neuron circuit, the first synapse circuit executes attenuation processing to decrease the degree of influence of the synaptic weight on the synapse signal.