A ballistic transport model for an artificial neuron

TL;DR

This paper proposes a novel ballistic transport-based model for artificial neurons that integrates synaptic weights into the active region, simplifying input terminal complexity and enabling tunable transmission functions for neural network training.

Contribution

The paper introduces a ballistic transport model for artificial neurons that embeds weights into the device's active region, reducing complexity and allowing weight tuning via gate voltage.

Findings

The model effectively maps input wavefunctions to output transmission values.

Tuning gate voltage adjusts weights and transmission for training.

The design simplifies neuron architecture by embedding weights into the active region.

Abstract

We introduce a model for an artificial neuron which is based on ballistic transport in a multi-terminal device. Unlike standard configurations, the proposed design embeds the synaptic weights into the active region, thus significantly reducing the complexity of the input terminals. This is achieved by defining the basic elements of the ballistic artificial neuron as follows: the input values are set by the incoming wavefunctions amplitudes, while the weights correspond to the scattering matrix elements. Furthermore, the output value of the activation function of the artificial neuron is given by the transmission function. By tuning the gate voltage, the scattering potential and, consequently, the weights are changed so that the value of the transmission function gets closer to the target output, which is essential in the training process of artificial neural networks. Thus, we provide here the modus operandi of a ballistic artificial neuron.