Learning Spatiotemporally Encoded Pattern Transformations in Structured Spiking Neural Networks.

Area of Science:

• Computational Neuroscience
• Artificial Neural Networks
• Spiking Neuron Models

Background:

• Precise spike timing in neurons underpins neural information encoding.
• Mechanisms for forming these temporal representations in neural networks are not fully understood.

Purpose of the Study:

• To investigate how multilayered spiking neural networks can learn to encode input patterns using temporal coding.
• To introduce and validate a new supervised learning rule for training such networks.

Main Methods:

• Development of a supervised learning rule named MultilayerSpiker.
• Training spiking networks with hidden layers to transform spatiotemporal input and output spike patterns.
• Evaluation of the rule's performance on pattern mapping capacity, network complexity, and classification accuracy.

Main Results:

• MultilayerSpiker successfully trains spiking networks for temporal coding tasks.
• The learning rule demonstrates robustness against input noise and good generalization capabilities.
• Performance was validated across various network structures and classification tasks using multispike encodings.

Conclusions:

• The MultilayerSpiker learning rule offers a technically capable method for training spiking neural networks.
• This work contributes to a systematic understanding of neural computation and learning in spiking networks.