Dendritic Computing with Multigate Ferroelectric Field-Effect Transistors

Area of Science:

• Neuromorphic Engineering
• Artificial Intelligence Hardware
• Computational Neuroscience

Background:

• Artificial neural networks typically use simplified point-neurons, lacking the complex processing capabilities of biological neurons.
• Biological neurons possess dendritic arbors that perform local computations, crucial for information processing and learning.

Purpose of the Study:

• To propose a novel artificial neuron design that mimics biological dendrites using multigate ferroelectric field-effect transistors.
• To leverage ferroelectric nonlinearity for local dendritic computations and transistor action for neuronal output.

Main Methods:

• Development of a multigate ferroelectric field-effect transistor-based neuron design.
• Implementation of an experimentally calibrated device-circuit-algorithm co-simulation framework.
• Comparison of network performance with and without dendritic neurons.

Main Results:

• The proposed dendritic neuron design mimics local nonlinear accumulation found in biological dendrites.
• Networks with dendritic neurons achieved superior performance with significantly fewer trainable parameters (approximately 17x reduction).
• Branched architecture facilitates smaller crossbar arrays for improved hardware integration efficiency.

Conclusions:

• Dendritic hardware, utilizing ferroelectric nonlinearity, can substantially enhance the computational efficiency of neuromorphic systems.
• This approach offers improved learning capacity, particularly for edge computing applications.
• The novel neuron design represents a significant step towards more biologically plausible and efficient artificial intelligence hardware.