Flexible Transmitter Network

Area of Science:

• Computational Neuroscience
• Artificial Intelligence

Background:

• Current artificial neural networks predominantly use the McCulloch-Pitts (MP) model, which simplifies neuron function.
• The MP model's limitations hinder biological realism and the ability to process complex data like spatiotemporal information.

Purpose of the Study:

• To propose a novel, biologically plausible neuron model, the flexible transmitter (FT) model.
• To develop a new neural network architecture, the flexible transmitter network (FTNet), utilizing the FT model.
• To demonstrate the FTNet's capability in handling complex data, particularly spatiotemporal data.

Main Methods:

• Introduced the flexible transmitter (FT) model, incorporating neurotransmitter dynamics and neurotrophin density.
• Formulated the FT model as a two-variable, two-valued function, encompassing the MP model as a special case.
• Developed the flexible transmitter network (FTNet) using a fully connected feedforward architecture with FT models.
• Implemented FTNet using an improved backpropagation algorithm in the complex-valued domain for gradient calculation.

Main Results:

• The FT model offers greater biological realism compared to the MP model.
• FTNet demonstrated significant potential and power in processing spatiotemporal data across various tasks.
• The FT model serves as a viable alternative building block for artificial neural networks.

Conclusions:

• The flexible transmitter (FT) model enhances biological plausibility in artificial neurons.
• Flexible transmitter networks (FTNets) show promise for advanced spatiotemporal data processing.
• This research validates the development of artificial neural networks incorporating neuronal plasticity.