Artificial sparse neuron dendrites for visual information inference

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computer Engineering

Background:

• Biological dendrites enable efficient visual processing through nonlinear integration and sparse parallel computation.
• Conventional neuromorphic devices lack the spatiotemporal processing capabilities of biological dendrites.
• Replicating dendritic properties is key to advancing efficient information inference in artificial systems.

Purpose of the Study:

• To develop an artificial neuron dendrite array that emulates biological dendritic spatiotemporal spike integration.
• To enable precise parallel computation and sparse spiking inference.
• To enhance computational efficiency and reduce power consumption in neuromorphic devices.

Main Methods:

• Integration of neurons, synapses, and dendrites into a single artificial array.
• Implementation of multigate threshold regulation for parallel sparse spiking inference.
• Formation of a sparse dendritic spiking neural network (SD-SNN) for various computational tasks.

Main Results:

• The SD-SNN demonstrated high-efficiency static and dynamic object processing.
• Achieved significant reductions in neuronal activity (99.5%) and power consumption (98% and 65%).
• Enabled parallel sparse inference with random spatial distribution.

Conclusions:

• The artificial neuron dendrite array effectively replicates biological dendritic functions for advanced computation.
• The developed SD-SNN offers a pathway to highly efficient and low-power neuromorphic systems.
• This research advances spatiotemporal computing capabilities and computational efficiency in artificial intelligence.