Event-Based, Timescale Invariant Unsupervised Online Deep Learning With STDP.

Area of Science:

• Neuromorphic Engineering
• Artificial Intelligence
• Computational Neuroscience

Background:

• Spiking neural networks (SNNs) offer potential for efficient, brain-inspired computation.
• Traditional deep learning often relies on batch processing, limiting real-time adaptation.
• Neuromorphic systems excel at processing continuous, event-based sensor data.

Purpose of the Study:

• To develop a deep spiking convolutional neural network (SCNN) for unsupervised online learning.
• To enable simultaneous training of all network layers for efficient inference.
• To create an event-driven architecture adaptable to real-world, continuous data streams.

Main Methods:

• Implemented a deep SCNN using integrate-and-fire (IF) neurons.
• Utilized spike-timing dependent plasticity (STDP) for unsupervised learning.
• Developed a simultaneous layer training mechanism and an event-driven STDP rule based on relative spike timings.

Main Results:

• The network performs unsupervised online deep learning from asynchronous, continuous event-based data.
• Simultaneous layer training allows for approximate online inference during learning.
• The network learns effectively without prior knowledge of data structure (e.g., number of classes, stimulus duration).
• The event-driven nature and lack of absolute timescale dependency make it robust and hardware-friendly.

Conclusions:

• The proposed SCNN architecture is suitable for online learning and inference from event-based sensors.
• The simultaneous training approach enhances efficiency and adaptability for real-world applications.
• This work advances the development of neuromorphic systems capable of direct feature learning from continuous sensory input.