Weighted Population Spike Coding for Bandwidth-Limited Neuromorphic Chips
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current stochastic spike coding schemas in neuromorphic systems, such as stochastic rate coding, are inefficient in terms of bandwidth and power consumption, requiring a large number of spikes to represent values, which exceeds the limited input and output bandwidth of neuromorphic chips like IBM TrueNorth.
Innovation Solution
The implementation of a weighted population code schema that represents values using a combination of axonal input lines and weights, allowing for efficient data ingestion and conversion to stochastic code, independent of the number of classifiers or dynamic range, thereby reducing the required number of spikes and maintaining high accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If stochastic rate coding is used to represent values, then classification accuracy can be maintained, but the number of spikes required increases significantly, exceeding the limited input and output bandwidth of neuromorphic chips
Solution Approach 1:
The patent segments the representation of a single value into multiple parallel spike streams, each carrying a portion of the information. Instead of using one large population of spikes to represent a value, the system divides the population into multiple smaller groups that work together, reducing the spike burden on any single channel while maintaining overall representation accuracy.
Solution Approach 2:
The patent introduces a temporal dimension to the spike coding scheme by using sequential time windows for different classifiers. Multiple classifiers process the same input value at different time steps rather than simultaneously, allowing the system to handle multiple classification tasks without proportionally increasing the spatial bandwidth requirements.
2Loss of information
If a large number of spikes are used to represent values in stochastic coding, then more information can be conveyed, but power consumption and bandwidth requirements increase beyond the capabilities of neuromorphic chips
Solution Approach 1:
The patent implements periodic action by organizing spike generation into discrete time windows or frames, where each window corresponds to a specific classifier. Spikes are generated periodically for each time window rather than continuously, allowing the system to control information throughput while managing power consumption through rhythmic, predictable spike patterns that can be efficiently processed by the neuromorphic hardware.
3Adaptability or versatility
If the number of classifiers is increased to handle multiple classification tasks, then system versatility improves, but the required input bandwidth increases proportionally
Solution Approach 1:
The patent resolves this contradiction by moving from a spatial to a temporal organization of classifiers. Instead of providing separate input channels for each classifier simultaneously (spatial dimension), the system sequences classifier processing across multiple time windows (temporal dimension). This allows N classifiers to share a single input channel over time, maintaining versatility while keeping bandwidth requirements constant.
Data Source
AI summary
Weighted population code in neuromorphic systems is provided. According to an embodiment, a plurality of input values is received. For each of the plurality of values, a plurality of spikes is generated. Each of the plurality of spikes has an associated weight. A consumption time is determined for each of the plurality of spikes. Each of the plurality of spikes is sent for consumption at its consumption time.