Reservoir based spiking models for univariate Time Series Classification.

Area of Science:

• Neuromorphic Computing
• Artificial Intelligence
• Machine Learning

Background:

• Advanced machine learning and deep learning excel at temporal processing but are energy-intensive, relying on power-hungry CPUs and GPUs.
• Spiking Neural Networks (SNNs) offer energy efficiency on specialized neuromorphic hardware.

Purpose of the Study:

• To present two novel SNN architectures for Time Series Classification (TSC) inspired by Reservoir Computing and Legendre Memory Units.
• To demonstrate the energy efficiency and performance of these SNN models on neuromorphic hardware.

Main Methods:

• Developed two SNN architectures for TSC: one based on general Reservoir Computing deployed on Loihi, and a second with non-linearity in the readout layer.
• Utilized Surrogate Gradient Descent for training the second model, enabling non-linear decoding of temporal features.
• Conducted experiments on five TSC datasets and performed energy profiling on Loihi and CPU.

Main Results:

• The second SNN model achieved new state-of-the-art spiking results for TSC, with up to 28.607% accuracy improvement on one dataset.
• Demonstrated significant reduction in neuron count (over 40x) compared to existing spiking models, indicating low computational overhead.
• Energy profiling confirmed the energy-efficient nature of the proposed models on neuromorphic hardware.

Conclusions:

• The proposed SNN models effectively address TSC tasks in an energy-efficient manner, outperforming existing methods.
• Non-linear decoding in SNNs enhances performance while maintaining computational efficiency.
• These findings highlight the potential of SNNs for green AI applications in temporal data processing.