Spiking Autoencoders With Temporal Coding.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Spiking neural networks (SNNs) process information using the timing of neural spikes.
• Temporal coding in SNNs enables supervised learning via backpropagation, matching conventional neural network accuracy.
• Autoencoders are key for learning compact data representations and dimensionality reduction.

Purpose of the Study:

• Introduce spiking autoencoders with temporal coding for image storage and reconstruction.
• Evaluate their performance on neuromorphic datasets (MNIST, FMNIST).
• Investigate the impact of parameters like latency, noise, and embedding size on performance.

Main Methods:

• Implemented spiking autoencoders utilizing temporal coding and pulse-based communication.
• Trained the networks using backpropagation for image reconstruction tasks.
• Analyzed the role of inhibition and explored parameter variations (latency, noise, embedding size).

Main Results:

• Single-layer spiking autoencoders effectively represented and reconstructed images from MNIST and FMNIST datasets.
• Performance was comparable or superior to conventional non-spiking autoencoders.
• Inhibition was found to be crucial for memorization and reconstruction, especially at higher target latencies.

Conclusions:

• Spiking autoencoders with temporal coding show significant potential for high-fidelity image reconstruction.
• These networks offer a biologically plausible alternative to traditional autoencoders.
• The findings suggest their utility as foundational components for complex, brain-inspired AI architectures.