Signal-to-event encoding parameter selection for multiple event classification with spiking neural networks.

Area of Science:

• Signal Processing
• Machine Learning
• Event-Driven Systems

Background:

• Event-driven systems process data via discrete-time event streams or encoded analog signals.
• Spiking Neural Networks (SNNs) are a class of machine learning models suited for event-based data.
• Efficient signal encoding is crucial for optimizing SNN performance and reducing computational load.

Purpose of the Study:

• To develop and validate a method for optimal event-based signal encoding parameter selection.
• To assess the efficiency of the proposed encoding method in terms of classification performance and data reduction.
• To demonstrate the method's effectiveness with different event-based encoding schemes and machine learning models.

Main Methods:

• A Bayesian optimization process was employed to select encoding parameters.
• A k-Nearest Neighbor (k-NN) classifier was used to evaluate event stream distance and guide parameter selection.
• The method was validated using vehicle monitoring sensor data and three encoding schemes: level-crossing, send-on-delta, and leaky integrate-and-fire.

Main Results:

• The optimized encoding parameters achieved up to 0.912 average accuracy for k-NN classification.
• The proposed method reduced the number of samples by 97.8% compared to classical periodic discrete-time signal representation.
• Spiking Neural Network (SNN) classifiers trained on the encoded data reached an average accuracy of up to 0.946, outperforming the k-NN baseline.

Conclusions:

• The model-agnostic signal-to-event encoding parameter selection method is effective for optimizing data representation.
• This approach shows significant potential for training sophisticated machine learning models, including SNNs, with reduced data requirements.
• The validated method offers a promising solution for efficient event-driven data processing in various applications.