Neuron Circuit Stochastic Synapse Weight Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for implementing spiking neural networks (SNNs) require excessive calculations and large memory for storing synaptic weight values, limiting learning speed and increasing costs.
Innovation Solution
A neuron circuit with sub-circuits for determining active signals, comparing cumulative reception counters with threshold values, and performing potentiating and depressing learning processes based on probability values to adjust synaptic weights, reducing memory requirements and improving learning efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing deterministic methods are used for SNN implementation, then learning accuracy can be maintained, but memory requirements increase significantly
Solution Approach 1:
The patent changes the fundamental parameter of synaptic weight representation from deterministic continuous values to stochastic binary values (0 or 1). This parameter change reduces memory requirements while maintaining learning accuracy through probabilistic computation that mimics biological neural behavior.
Solution Approach 2:
The patent employs simple binary synapse states (active/inactive) that can be easily flipped and reset, replacing complex continuous weight storage. This allows for more efficient memory usage and faster learning dynamics, as the system can quickly adapt synapse states without requiring large memory capacity.
2Reliability
If existing methods are used for SNN implementation, then comprehensive learning can be achieved, but calculation requirements become excessive
Solution Approach 1:
The patent replaces complex deterministic calculation mechanisms with stochastic probabilistic mechanisms. Instead of performing extensive mathematical computations to determine weight updates, the system uses random number generation and simple threshold comparisons, dramatically reducing calculation requirements while preserving learning capability.
Solution Approach 2:
The learning process becomes self-organizing through stochastic mechanisms. The system automatically adjusts synapse states based on probabilistic rules and input patterns without requiring intensive external control or computation, allowing the network to self-learn efficiently.
3Manufacturing precision
If existing deterministic methods are used, then precise weight control is possible, but learning speed decreases
Solution Approach 1:
The patent implements periodic learning updates where synapse states are adjusted at discrete time points based on accumulated input patterns. This periodic stochastic updating mechanism enables faster learning by allowing parallel updates across multiple synapses simultaneously, rather than requiring sequential precise adjustments.
Solution Approach 2:
The system transitions from static precise weight values to dynamic stochastic binary states that can rapidly flip between 0 and 1. This dynamic behavior allows the network to adapt quickly to new patterns, significantly increasing learning speed while the probabilistic nature maintains sufficient control precision for effective learning.
Data Source
AI summary
A neuron circuit performing synapse learning on weight values includes a first sub-circuit, a second sub-circuit, and a third sub-circuit. The first sub-circuit is configured to receive an input signal from a pre-synaptic neuron circuit and determine whether the received input signal is an active signal having an active synapse value. The second sub-circuit is configured to compare a first cumulative reception counter of active input signals with a learning threshold value based on results of the determination. The third sub-circuit is configured to perform a potentiating learning process based on a first probability value to set a synaptic weight value of at least one previously received input signal to an active value, upon the first cumulative reception counter reaching the learning threshold value, and perform a depressing learning process based on a second probability value to set each of the synaptic weight values to an inactive value.