Energy-Efficient Neural Network Inference with Microcavity Exciton Polaritons

Abstract

We propose all-optical neural networks characterized by very high energy efficiency and performance density of inference. We argue that the use of microcavity exciton polaritons allows one to take advantage of the properties of both photons and electrons in a seamless manner. This results in strong optical nonlinearity without the use of optoelectronic conversion. We propose a design of a realistic neural network and estimate energy cost to be at the level of attojoules per bit, also when including the optoelectronic conversion at the input and output of the network, several orders of magnitude below state-of-the-art hardware implementations. We propose two kinds of nonlinear binarized nodes based either on optical phase shifts and interferometry or on polariton spin rotations.

Microcavity Exciton‐Polariton Quantum Spin Fluids

“…These are desirable attributes to construct spinoptronics as a future computing architecture. Lastly, we draw attention to recent cuttingedge research activities of microcavity exciton-polariton spin SFs in the development of high performance computing technologies in several modern platforms: quantum computers, [193][194][195] neural networks, [196][197][198] and neuromorphic computing. [199,200] These new computing technologies target to solve computationally challenging problems in various scientific and engineering fields.…”

Neural networks based on ultrafast time-delayed effects in exciton-polaritons

“…One of the main practical obstacles in implementing optical data processing is the weakness of the nonlinear response of optical media, or equivalently photon-photon interactions, which is necessary to implement an activation function of a neuron or a transistor. From this point of view, semiconductor exciton-polaritons, where photons and matter excitations (excitons) coexist in a quantum superposition state [15], are an exceptionally promising candidates [16]. The excitonic component is providing strong interactions required for low-threshold nonlinear operation [17,18].…”

“…Neural networks based on exciton-polaritons have been recently investigated both theoretically [16,[22][23][24] and experimentally [25][26][27]. Results presented so far, relied on reservoir computing [24] and binarized extreme learning machines [28] approaches where most of the network connections are static and only the output layer is trainable.…”

Явление Самозахвата В Системе Экситон-Поляритонов При Накачке Нижней Поляритонной Ветви Двумя Лазерными Импульсами С Близкими Частотами

“…Свойства экситонполяритонов предполагается использовать в различных оптических и квантовых приборах, таких как оптические транзисторы [9][10][11], диоды [12], интерферерометры [13], маршрутизаторы [14], ответвители [15][16][17], лазеры [18]. Энергоэффективный нейросетевой логический вывод с использованием экситон-поляритонов в микрорезонаторе был получен в [19].…”