The impact of encoding-decoding schemes and weight normalization in spiking neural networks

Area of Science:

• Computational Neuroscience
• Machine Learning
• Artificial Intelligence

Background:

• Spike-timing dependent plasticity (STDP) is crucial for learning and memory in biological neural networks.
• Understanding STDP's learning window in spiking neural networks (SNNs) is key for developing advanced AI.
• Previous research has explored STDP's functionality, but its interaction with various parameters requires further investigation.

Purpose of the Study:

• To investigate the interplay of encoding/decoding schemes, STDP learning windows, and normalization rules in SNN classifiers.
• To evaluate classifier performance on benchmark datasets like MNIST, NIST, and ETH80-Contour.
• To determine optimal configurations for efficient and accurate SNN-based learning.

Main Methods:

• Trained and tested SNN classifiers on MNIST, NIST, and ETH80-Contour datasets.
• Systematically varied encoding/decoding schemes, STDP learning window parameters, and normalization rules.
• Analyzed classifier performance metrics including accuracy and decoding time.

Main Results:

• Classical STDP achieved optimal performance when no normalization rules were applied.
• First-spike decoding classifiers demonstrated significantly faster decoding times compared to spike count classifiers.
• Classifier accuracy decreased with increased encoding duration (10ms to 34ms) using the count decoding scheme without normalization.
• Normalization of output weights enhanced the performance of first-spike decoding classifiers.

Conclusions:

• The choice of normalization significantly impacts SNN classifier performance, with no normalization often yielding the best results for classical STDP.
• First-spike decoding offers a computationally efficient alternative to spike count decoding in SNNs.
• Weight normalization is critical for optimizing first-spike decoding SNN classifiers, highlighting its importance in SNN design.