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

Event-driven systems can operate either on discrete-time event streams or on analog signals transformed into the event domain by a predefined encoding scheme. This paper studies the problem of optimal event-based signal encoding if data are to be processed by a machine learning model, such as the spiking neural network (SNN). We introduce a method of encoding parameter selection that evaluates a k-Nearest Neighbor (k-NN) classifier operating on a measure of the event stream distance in multiple trials of a Bayesian optimization process. The efficiency of the proposed method is assessed by relating the classification performance with the number of events produced by a signal-to-event encoding scheme. The proposed method is validated for vehicle monitoring sensor data with three event-based encoding schemes: level-crossing encoding, send-on-delta, and leaky integrate-and-fire encoder. The best-performing sets of encoding parameters give an average accuracy of up to 0.912 for the k-NN classification, while producing 97.8% fewer number of samples than for the classical periodic discrete-time signal representation. Additionally, we train the SNN classifiers on data encoded according to the selected sets of parameters, achieving an average classification accuracy of up to 0.946, improving upon the k-NN baseline. This shows that the proposed model-agnostic signal-to-event encoding parameter selection is promising for training sophisticated machine learning models.