Character recognition from trajectory by recurrent spiking neural networks

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Spiking neural networks (SNNs) offer biological plausibility and power efficiency on neuromorphic hardware.
• Recurrent neural networks (RNNs) excel at processing time series data.
• Integrating RNN properties into SNNs for improved performance remains a challenge.

Purpose of the Study:

• To propose a recurrent spiking neural network (RSNN) for character recognition using trajectory data.
• To enhance the generalization ability of SNNs by incorporating recurrent properties.
• To achieve higher recognition accuracy compared to existing methods.

Main Methods:

• Development of a novel recurrent spiking neural network architecture.
• Introduction of a new encoding method utilizing varying time ranges for input streams across different recurrent layers.
• Experimental validation on character datasets from the University of Edinburgh.

Main Results:

• The proposed RSNN demonstrated improved generalization capabilities.
• The novel encoding method outperformed general encoding techniques.
• Experiments showed higher average recognition accuracy compared to existing methods.

Conclusions:

• The developed recurrent spiking neural network effectively leverages recurrent properties for character recognition.
• The proposed encoding strategy enhances model generalization and recognition performance.
• This approach offers a promising direction for improving SNNs in time-series recognition tasks.