Feature Representations for Neuromorphic Audio Spike Streams.

Area of Science:

• Neuromorphic engineering
• Spiking neural networks
• Bio-inspired computing

Background:

• Event-driven neuromorphic sensors generate asynchronous spike streams.
• Integrating these sensors with deep neural networks is challenging due to data processing limitations.
• Current pre-processing methods for spike streams require further investigation.

Purpose of the Study:

• To investigate frame-based feature extraction methods for spike streams.
• To evaluate synchronous and asynchronous features using spike count and event binning.
• To propose and validate a novel pre-processing technique for enhanced inter-spike timing preservation.

Main Methods:

• Utilized the N-TIDIGITS18 dataset, recorded from a spiking silicon cochlea sensor.
• Compared spike count and constant event binning for feature generation.
• Implemented a recurrent neural network for classification.
• Proposed an exponential kernel pre-processing method to preserve inter-spike timing.

Main Results:

• The proposed exponential features outperformed spike count features.
• Achieved over 91% accuracy on the digit classification task.
• Demonstrated a performance improvement of at least 2.5% compared to spike count features.
• Established a new state-of-the-art performance on the N-TIDIGITS18 dataset.

Conclusions:

• The exponential kernel pre-processing method effectively preserves inter-spike timing information.
• This approach significantly enhances classification accuracy for spike-based data.
• The findings pave the way for more effective integration of neuromorphic sensors with deep learning models.