Exploiting the stimuli encoding scheme of evolving Spiking Neural Networks for stream learning

Area of Science:

• Machine Learning
• Data Stream Processing
• Computational Neuroscience

Background:

• Traditional machine learning struggles with data stream challenges like incremental learning and non-stationarity.
• Spiking Neural Networks (SNNs), particularly Evolving SNNs, show promise for stream learning.
• SNNs utilize population encoding, such as Gaussian Receptive Fields (GRFs), to convert stimuli into temporal spikes.

Purpose of the Study:

• To investigate the predictive potential of Gaussian Receptive Fields as a preprocessing step for data stream learning.
• To determine the conditions under which GRFs provide prediction gains in stream classification.
• To assess the computational efficiency and model-agnostic nature of GRFs.

Main Methods:

• Utilizing Gaussian Receptive Fields as a preprocessing technique for input data streams.
• Applying GRFs to various stream learning models.
• Evaluating performance on both synthetic and real-world stream datasets.
• Comparing results against stream learning models without GRF preprocessing.

Main Results:

• Gaussian Receptive Fields significantly boost the predictive performance of multiple stream learning methods.
• GRFs act as a computationally lightweight and model-agnostic preprocessing step.
• Empirical evidence supports the effectiveness of GRFs across diverse datasets and models.

Conclusions:

• Gaussian Receptive Fields are a valuable tool for enhancing data stream classification.
• The findings encourage further research into applying GRFs to broader machine learning problems.
• GRFs offer a practical approach to improving stream learning efficiency and accuracy.