Correlative Time Coding for Spiking Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Spiking neural networks face challenges in efficiently processing large datasets and addressing correlated input elements, as existing methods rely on independent spike patterns that limit the network's ability to learn from datasets with thousands of elements and fail to account for correlations between input elements.
Innovation Solution
A computer-implemented method and spiking neural network architecture that generates pseudo-random spikes based on feature values, allowing for classification by measuring the distance between output spiking patterns and predefined target patterns, which enables efficient processing of large datasets and handling of correlated input elements through a correlative temporal encoding scheme.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If independent Poisson spike trains are generated for each input neuron, then a relatively large number of patterns can be encoded and learned, but only information on small data sets (about 100 elements) can be reliably learned and correlations between input elements are not addressed
Solution Approach 1:
The patent applies dynamics by making the spike generation process adaptive rather than static. The input neurons generate spikes based on dynamically computed feature values derived from the input data, allowing the system to adapt to different data sets and their correlations. This dynamic approach enables reliable learning of large numbers of patterns (hundreds of thousands or millions) while capturing correlations between input elements, resolving the contradiction between pattern capacity and adaptability to large datasets.
Solution Approach 2:
The patent changes the parameters of spike generation from fixed Poisson processes to feature-dependent spike times. By computing feature values from input data and using these to determine spike generation parameters, the system can encode correlations between inputs. This parameter transformation allows the network to handle large datasets effectively while maintaining the ability to encode and learn from complex pattern relationships.
2Ease of operation
If rate encoding is used for spike patterns, then a dominant paradigm for information processing is established, but many operations are needed for reliable estimation of neural firing rates
Solution Approach 1:
The patent replaces the complex rate encoding mechanism with a temporal encoding approach. Instead of relying on firing rate estimation that requires many operations, the system uses precise spike timing relative to feature values. This substitution eliminates the need for complex rate estimation operations while maintaining effective information processing, resolving the contradiction between operational simplicity and computational complexity.
3Adaptability or versatility
If each input element is encoded with an independent pattern, then encoding and learning of multiple patterns is enabled, but the network cannot reliably learn large data sets with thousands of elements
Solution Approach 1:
The patent merges the encoding of multiple input features into a coordinated spike generation process. By computing feature values from all input elements and using these combined features to determine spike times across the input neuron population, the system encodes correlations between inputs. This merging approach enables reliable learning of large data sets (hundreds of thousands or millions of patterns) while maintaining multi-pattern encoding capability, resolving the contradiction between encoding versatility and data set size.
Data Source
AI summary
A computer-implemented method for classification of an input element to an output class in a spiking neural network may be provided. The method comprises receiving an input data set comprising a plurality of elements, identifying a set of features and corresponding feature values for each element of the input data set, and associating each feature to a subset of spiking neurons of a set of input spiking neurons of the spiking neural network. Furthermore, the method comprises also generating, by the input spiking neurons, spikes at pseudo-random time instants depending on a value of the feature for a given input element, and classifying an element into a class depending on a distance measure value between output spiking patterns at output spiking neurons of the spiking neural network and a predefined target pattern related to the class.