Related Experiment Videos
Information Transmitted From Bioinspired Neuron-Astrocyte Network Improves Cortical Spiking Network's Pattern
IEEE Transactions on Neural Networks and Learning Systems
|April 17, 2019
Summary
Two spiking neural networks (SNNs) were trained on digit datasets. Connecting them enabled one network to classify unseen capital letters with 70.57% accuracy, demonstrating effective information transfer.
Area of Science:
- Computational Neuroscience
- Artificial Intelligence
- Machine Learning
Background:
- Spiking Neural Networks (SNNs) offer a biologically plausible model for computation.
- Unsupervised learning methods are crucial for extracting patterns from complex data.
- Inter-network communication is key to enhancing model capabilities.
Purpose of the Study:
- To train and evaluate two SNNs, the cortical spiking network (CSN) and the cortical neuron-astrocyte network (CNAN), on digit recognition tasks.
- To investigate the transfer of information between CNAN and CSN using prolate spheroidal wave functions (PSWF).
- To assess the performance of the enhanced CSN in classifying previously unseen data, specifically capital letters.
Main Methods:
- Training of CSN and CNAN using a spike-based unsupervised learning approach on MNIST and alpha-digit datasets.
- Interconnecting CNAN and CSN while maximizing synchronization via prolate spheroidal wave functions (PSWF).
- Evaluating the classification accuracy of the integrated network on both trained and untrained data (capital letters).
Main Results:
- Achieved 96.1% accuracy for CSN and 77.35% for CNAN on digit datasets.
- The integrated network (CSN receiving information from CNAN) achieved 70.57% accuracy on capital letters without prior training.
- Demonstrated significant information transfer, with an overall contribution of 87.47% from the CNAN to CSN.
- Observed that CSN neurons classifying MNIST digits also supported alpha-digit classification, indicating cross-dataset similarity recognition.
Conclusions:
- The study successfully demonstrates effective unsupervised learning and information transfer between SNNs.
- Connecting CNAN to CSN via PSWF enhances CSN's capabilities, enabling classification of novel data without retraining.
- The findings highlight the potential of SNNs for robust pattern recognition and knowledge transfer across different datasets.