Single Neurons with Delay-Based Learning Can Generalise Between Time-Warped Patterns
- Publisher:
- Springer International Publishing
- Publication Type:
- Conference Proceeding
- Citation:
- Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 2021, 12894 LNCS, pp. 127-138
- Issue Date:
- 2021-01-01
Closed Access
| Filename | Description | Size | |||
|---|---|---|---|---|---|
| Arnold2021_Chapter_SingleNeuronsWithDelay-BasedLe.pdf | Published version | 687.37 kB |
Copyright Clearance Process
- Recently Added
- In Progress
- Closed Access
This item is closed access and not available.
Spatiotemporal patterns, such as words in speech, are rarely precisely the same duration, yet a word spoken faster or slower is still easily recognisable. Neural mechanisms underlying this ability to recognise stretched or compressed versions of the same spatiotemporal pattern are not well understood. Recognition of time-varying patterns is often studied at the network level, however here we propose a single neuron using learnable spike delays for the task. We characterise the response of a single neuron to stretched and compressed versions of a learnt pattern and show that using delays leads to pattern recognition above 99% accuracy for patterns morphed by up to 50%. Additionally, we demonstrate a significantly reduced response when the pattern is reversed, a property that is often difficult to reproduce in synaptic efficacy (synaptic weight) based learning systems. With appropriate settings of the neuron membrane time constant and spike threshold, we show that a single neuron is able to generalise to time-warped patterns while discriminating temporally reversed patterns. Together, these results highlight the potential of synaptic delay-based learning rules as a robust mechanism for learning time-warped spatiotemporal patterns.
Please use this identifier to cite or link to this item: