Evolving Spiking Neural Networks for online learning over drifting data streams.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Data streams are increasingly large and non-stationary, necessitating adaptive online learning algorithms.
• Concept drift, or changes in data distribution, challenges traditional incremental learning methods.
• Spiking Neural Networks (SNNs) are well-suited for dynamic environments but underutilized in online learning.

Purpose of the Study:

• To adapt Spiking Neural Networks for efficient and scalable online learning.
• To address the challenge of concept drift in high-velocity data streams.
• To investigate SNNs' potential for real-time adaptation in changing environments.

Main Methods:

• Developing an online learning framework for Spiking Neural Networks.
• Implementing data reduction techniques to optimize limited neuron repository size.
• Evaluating adapted SNNs on synthetic and real-world non-stationary datasets.

Main Results:

• Adapted Spiking Neural Networks demonstrate superior performance in handling concept drift.
• Optimized neuron repository exploitation leads to higher accuracy scores compared to naive SNNs.
• The proposed methods enable SNNs to effectively adapt to changing data distributions.

Conclusions:

• Spiking Neural Networks, when tailored for online learning, offer a powerful solution for non-stationary data.
• Data reduction and strategic neuron repository use are key to SNN adaptability in concept drift scenarios.
• This research bridges a gap in utilizing SNNs for robust online learning and concept drift detection.