Energy-Performance Assessment of Oscillatory Neural Networks Based on VO₂ Devices for Future Edge AI Computing

Abstract

Oscillatory Neural Network (ONN) is an emerging neuromorphic architecture composed of oscillators that implement neurons and coupled by synapses. ONNs exhibit rich dynamics and associative properties, which can be used to solve problems in the analog domain according to the paradigm let physics compute. For example, compact oscillators made of VO 2 material are good candidates for building low-power ONN architectures dedicated to AI applications at the edge like pattern recognition. However, little is known about the ONN scalability and its performances when implemented in hardware. Before deploying ONN, it is necessary to assess its computation time, energy consumption, performance and accuracy for a given application. Here, we consider a VO 2 -oscillator as an ONN building block and we perform circuit-level simulations to evaluate the ONN performances at the architecture level. Notably, we investigate how ONN computation time, energy and memory capacity scale with the number of oscillators. We show that ONN energy grows linearly when scaling up the network, making it suitable for large-scale integration at the edge. Furthermore, we investigate the design knobs for minimizing the ONN energy. Assisted by TCAD simulations, we report on scaling of VO 2 devices in crossbar geometry to decrease the oscillator voltage and energy. We benchmark ONN versus state-of-the-art architectures and observe that the ONN paradigm is a competitive energy efficient solution for scaled VO 2 devices. Finally, we present how ONN can efficiently detect edges in images captured on low-power edge devices.

A mixed-signal oscillatory neural network for scalable analog computations in phase domain

“…Interestingly, the settling time (time to reach a steady phase state) seems to grow according to a logarithmic law with the ONN size. This result refines some previous scaling observations mentioning a quasi-constant settling time [9,17]. It also confirms the high ONN parallelism and ability to compute in a few tens of cycles, even for large graphs.…”

“…However, when the oscillation frequency increases to 1.2 MHz, the limited bandwidth of the hysteresis circuit causes a switching delay and a larger phase deviation δϕ = 13 • . For a given phase precision required by the application, this error could be mitigated by increasing the bandwidth of the hysteresis circuit or slowing down the oscillators, and constitute a trade-off with the energy consumption [17]. Interestingly, we have observed in experiments that having recurrent synapses W ij = W ji compensate the hysteresis delay induced in both neurons and the theoretical phase fixed point is reached in that case (see figures 5(c) and (d)).…”

“…In other words, there is no external control that makes the neuron state evolve. Instead, all the phases evolve in parallel and in continuous time, enabling fast and energy-efficient inferences in the analog domain [17].…”

Ultralow Power Electronic Analog of a Biological Fitzhugh–Nagumo Neuron

“…Purely electronic oscillatory neurons can extend the ability of FN neurons to different types of sensors and enable the construction of deep oscillator-based neural networks. In addition, electronic oscillatory neurons are instrumental to traditional oscillatory neural networks (ONNs), where a network of coupled oscillators can be used as associative memory − or an engine for convolution-like operations. ,,− However, there are three important challenges to achieving ultralow power oscillator-based computing: (1) scalability, (2) energy consumption, and (3) variability of the oscillatory neurons. These neurons serve as the smallest unit of computation in oscillator networks similar to the different logic gates in digital computing systems.…”

How fast can vanadium dioxide neuron-mimicking devices oscillate? Physical mechanisms limiting the frequency of vanadium dioxide oscillators

“…ONNs encode information into the stable phase difference between each oscillator of the network and a reference oscillator, achieved if/when the system settles into a synchronized state [21,22]. The energy that the ONN takes to compute is proportional to the ONN settling time [23], which is the number of cycles to synchronize multiplied by the oscillation period. The increase in oscillator frequency should bring about an overall reduction in energy consumption for ONN operativity.…”