Synaptic Resource Model for Spiking Neural Network Hardware
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for modeling synaptic plasticity in spiking neural networks are inefficient in hardware implementation, particularly in handling multiple spike events and combining long-term potentiation (LTP) and long-term depression (LTD) rules, which are typically triggered separately.
Innovation Solution
A method and apparatus that model synaptic resources using a defined domain, modulate weight changes for multiple spike events upon a single trigger, and update resources based on these events, incorporating a circuit to save the resource state and project it back in time for pre-synaptic spike replay, allowing for combined LTP and LTD handling.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If separate LTP and LTD rules are used for synaptic plasticity, then biological accuracy is improved, but hardware complexity and operation count increase
Solution Approach 1:
The patent combines separate LTP and LTD rules into a unified resource model where a single resource variable r(w) modulates both potentiation and depression processes. This merging allows the hardware to handle multiple spike events and both LTP/LTD operations through one integrated circuit structure, reducing complexity while preserving biological accuracy in capturing synaptic behavior.
Solution Approach 2:
The resource model serves multiple functions simultaneously: it captures resource depletion during LTP, resource recovery during LTD, and handles multiple spike events within a single framework. This multi-functionality allows one hardware circuit to replace what would otherwise require separate dedicated circuits for each plasticity rule, reducing overall hardware complexity.
2Measurement precision
If multiple spike events are handled separately, then accuracy is improved, but operation count and processing time increase
Solution Approach 1:
The resource model maintains continuous tracking of synaptic resource levels across multiple spike events without requiring separate processing for each event. The resource variable r(w) is updated continuously based on the cumulative effect of multiple spikes, allowing the system to handle burst firing and multiple events efficiently while maintaining accurate learning dynamics.
Solution Approach 2:
The model pre-computes the resource level r(w) based on current weight and spike history, which then modulates all subsequent LTP and LTD operations. This preliminary computation of the resource state allows multiple spike events to be handled efficiently through a single trigger mechanism, reducing the operation count while preserving accuracy.
3Speed
If resource model is implemented in hardware, then speed is improved, but circuit complexity increases
Solution Approach 1:
The patent transforms the resource model into a hardware-friendly parameter representation where the resource level r(w) is derived from weight w through a defined function. This parameter transformation allows the complex resource dynamics to be implemented using simple arithmetic operations in hardware, achieving fast learning speed without excessive circuit complexity.
Data Source
AI summary
Certain aspects of the present disclosure support a method of designing the resource model in hardware (or software) for learning spiking neural networks. The present disclosure comprises accounting for resources in a different domain (e.g., negative log lack-of-resources instead of availability of resources), modulating weight changes for multiple spike events upon a single trigger, and strategically advancing or retarding the resource replenishment or decay (respectively) to overcome the limitation of single event-based triggering.