Mixed Neural Coding for Spiking Neurons
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current spiking neural networks (SNNs) face limitations in encoding multiple variables efficiently, leading to increased energy consumption and hardware size requirements, and struggle with unsupervised learning from sequential and multi-timescale data.
Innovation Solution
The method involves generating spikes that encode two variable values at each time instant, using mixed neural coding that combines rate-based and timing-based coding, and optimizing synaptic weights through a synapse system with both short-term and long-term plasticity components, allowing for efficient processing of data features across different timescales.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If current spiking neural networks encode single variable per neuron, then hardware implementation is simpler, but information capacity is limited and more neurons are required
Solution Approach 1:
The patent combines rate-based coding and timing-based coding into a unified mixed coding framework, allowing each neuron to encode multiple variables simultaneously. This merging of coding strategies enables the network to represent complex multi-dimensional data with fewer neurons while maintaining manageable hardware complexity through systematic integration of encoding mechanisms.
Solution Approach 2:
The mixed coding scheme enables neurons to perform multiple encoding functions concurrently - representing both the identity of encoded items and their temporal relationships. This multi-functionality allows a single neuron to handle what would otherwise require multiple specialized neurons, reducing overall network size while preserving computational capabilities.
2Adaptability or versatility
If spiking neural networks process multi-timescale data, then learning capability improves, but energy consumption increases
Solution Approach 1:
The patent segments the processing of multi-timescale data by introducing distinct timescale parameters and separate learning rates for different temporal scales. This segmentation allows the network to handle multiple timescales systematically without requiring uniformly high computational resources across all processing operations, thereby reducing overall energy consumption while maintaining learning capability.
Solution Approach 2:
The invention employs parameter changes by dynamically adjusting learning rates and timescale parameters based on the specific processing requirements. This allows the network to optimize energy consumption by using appropriate parameter settings for different learning tasks and data characteristics, rather than maintaining maximum computational capacity continuously.
3Adaptability or versatility
If synapse system uses only long-term plasticity, then learning from sequential data is limited, but system complexity is reduced
Solution Approach 1:
The patent merges long-term plasticity (LTP) and short-term plasticity (STP) mechanisms into a unified synapse system. This combination enables the system to learn from sequential data across multiple timescales simultaneously - LTP capturing long-term patterns and STP capturing short-term temporal dynamics - while maintaining a cohesive architectural framework that manages complexity through integrated design.
Solution Approach 2:
The synapse system implements dynamics by allowing plasticity parameters to vary over different timescales. The system dynamically adjusts between long-term and short-term plasticity effects based on the temporal characteristics of incoming data, enabling adaptive learning from sequential patterns while maintaining system complexity through principled temporal differentiation rather than multiple independent systems.
Data Source
AI summary
The present disclosure relates to a method of generating spikes by a neuron of a spiking neural network. The method comprises generating at each time, wherein the spike generation encodes at each time instant at least two variable values at the neuron. Synaptic weights may be optimized for a spike train generated by a given presynaptic neuron of a spiking neural network, wherein the spike train being indicative of features of at least one timescale.