An unsupervised neuromorphic clustering algorithm.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Hardware Systems

Background:

• Conventional computers require significant power for complex tasks compared to the human brain.
• Neuromorphic hardware systems, designed to mimic the brain's efficiency and parallelism, are emerging.
• Developing specialized neuromorphic algorithms is crucial for leveraging these new hardware capabilities.

Purpose of the Study:

• To develop a spiking neural network model for unsupervised clustering on neuromorphic hardware.
• To enable mapping of time-invariant, rate-coded datasets into a feature space with adjustable resolution.
• To demonstrate the practical application of neuromorphic clustering as a preprocessing module.

Main Methods:

• Developed a spiking neural network model incorporating spike-timing-dependent plasticity and lateral inhibition.
• Implemented and tested the model on the SpiNNaker neuromorphic system and GPUs using the GeNN framework.
• Evaluated clustering performance against conventional algorithms like self-organizing maps, neural gas, and k-means.

Main Results:

• The neuromorphic clustering algorithm successfully mapped datasets into a specified feature space resolution.
• Performance was comparable to established conventional clustering algorithms.
• Integration with a supervised neuromorphic classifier demonstrated its utility as a preprocessing step.

Conclusions:

• The developed spiking neural network model provides an effective neuromorphic solution for unsupervised clustering.
• This algorithm facilitates the use of neuromorphic hardware for efficient data preprocessing and analysis.
• The findings support the advancement of brain-inspired computing for complex AI tasks.