Oscillatory neural networks for energy-efficient computing

In-vehicle artificial intelligence requires very substantial computing power to process large quantities of data in real time from many different types of device, such as cameras, proximity sensors for an autonomous vehicle, and so on. Today, data processing takes place mainly in the cloud or in data centers. But this requires high bandwidth for data transfer, data security, data processing delays and high energy consumption. To solve these problems, the scientific and industrial community has focused on developing low-power architectures to enable local data processing. This becomes all the more crucial as, in the coming years, billions of peripheral IoT devices will be connected

So how can we drastically reduce the power consumption of these electronic systems? Research is now turning to new computing units to enable more data to be processed in less time, while consuming less energy. Conventional processors (CPUs) are based on the von Neumann architecture, in which the computing unit and memory are separated into two distinct units. The transfer between the memory and the calculation unit is therefore time- and energy-consuming. To speed up processing, new processors (GPUs) use several of these architectures (calculation unit and memory unit) to perform calculations in parallel. However, this type of processor still consumes far too much energy when transferring data between processor and memory. To reduce data transfer between processor and memory, in-memory computing architectures have also been proposed, offering memory cells with simple calculation capabilities. More recently, research has turned to new architectures inspired by the nature and functioning of the human brain to reduce the energy consumption of electronic systems. These circuits are known as neuromorphic. Researchers have estimated that the human brain requires around 20 watts of power to perform extremely complex tasks.

This is how Artificial Neural Networks (ANNs) came into being. ANNs are inspired by neural networks and their ability to reproduce and model non-linear processes. An ANN contains a multitude of neurons connected by synapses. The calculations are performed by the neurons, while the synapses represent the system's memory through their value.

Several types of ANN have been designed, the most widely used being single or multi-layer perceptron neural networks, convolutional neural networks, pulse neural networks and recurrent neural networks. Among recurrent neural networks, oscillatory neural networks (ONNs) have recently emerged as an interesting alternative for solving complex problems at low power consumption.

This article gives an in-depth description of the ONN-based computing paradigm, implementation, and applications. First,...