Neuromorphic Photonics Circuits: Contemporary Review

Abstract

Neuromorphic photonics is a cutting-edge fusion of neuroscience-inspired computing and photonics technology to overcome the constraints of conventional computing architectures. Its significance lies in the potential to transform information processing by mimicking the parallelism and efficiency of the human brain. Using optics and photonics principles, neuromorphic devices can execute intricate computations swiftly and with impressive energy efficiency. This innovation holds promise for advancing artificial intelligence and machine learning while addressing the limitations of traditional silicon-based computing. Neuromorphic photonics could herald a new era of computing that is more potent and draws inspiration from cognitive processes, leading to advancements in robotics, pattern recognition, and advanced data processing. This paper reviews the recent developments in neuromorphic photonic integrated circuits, applications, and current challenges.

Keywords: artificial intelligence; imaging; machine learning; neuromorphic computing; photonic integrated circuit.

Figure 7

(a) An illustration of the PPNN. It consists of several components, including M laser diodes (LDs), a MUX, a 3dB X-splitter, a bias branch denoted as W_b, and a reconfigurable Optical Linear Algebra Unit (OLAU) [97]. The OLAU comprises a 1-to-N splitting stage, input (X_n) and weight (W_n) modulator banks, and an N-to-1 combiner stage. The output from the combiner stage interferes with the bias signal within a 3dB X-coupler and is then sent to a DEMUX. A closer examination reveals details of (b) 1-to-N splitting and (d) its N-to-1 coupling stage [97], (c) view of the bias branch, which includes wavelength-selective weights and phase modulators [97], (e) a closer look at an axon of the OLAU, which consists of switches for signal routing and modulators for inputs (x_n,m) and weights (w_n,m) [97], layout of a Single Layer Coherent Optical Neural Network [106] (f) a tunable all-pass single RR functions as a phase tuning component, (g) tunable serially-coupled double RRs are employed as signal mixing components between the ports, (h) the nonlinear activation unit transforms input signal 𝑥𝑛 into 𝑓(𝑥𝑛), where 𝑓 represents a nonlinear function (with 𝑛 = 3 in this example). The black ring within the nonlinear activation unit acts as a directional coupler, directing a portion of the optical energy (𝛼) for electrical signal processing. The diode is a photodetector, and the blue ring modulates the signal. An electronic circuit (M) processes the electronic output from the photodetector to generate a modulation signal for the right ring [106] (i) displays the transmission and phase responses of a bus waveguide side-coupled with a ring, showcasing variations as a function of phase detuning, Δϕ. Over-coupling, indicated in green, is employed for phase-tuning components. At the same time, critical coupling, highlighted in blue, is crucial for achieving a larger amplitude tuning range in the nonlinear activation ring [106], (j) provides an example transmission and phase response of the coupled double ring, used as a signal mixing component [106].

Figure 13

(a) Diagram of typical structure of 2D layered materials [157]; (b) the communication architectures: α architecture, both the transmitter and the receiver are located within the same cuvette and phase, potentially with one component protected by micelles. In the β architecture, the transmitter and receiver are situated in the same cuvette but in two immiscible phases (water-ionic liquid). In the γ architecture, the transmitter(s) and receiver(s) are placed in separate cuvettes. The networks have been achieved by combining or enhancing the α, β, and γ architectures through hybridization or upgrades [162].

Figure 1

Neuromorphic computing market. Inspired by [14].

Figure 2

A comparison between specialized deep-learning digital electronic architectures and silicon photonic and nanophotonic platforms. In this context, photonic systems can support high on-chip bandwidth densities while maintaining low energy consumption during data transmission and computational tasks. The metrics for electronic architectures have been sourced from various references [21,22,23,24]. The metrics for silicon photonic platforms are estimated based on a contemporary silicon photonic setup operating at 20 GHz, comprising 100 channels with tightly packed micro rings. Meanwhile, the nanophotonic metrics are derived from the assumption of closely packed athermal microdisks [25], each occupying an area of approximately 20 µm, running at 100 GHz and operating close to the shot noise limit. Inspired by [17].

Figure 3

Traditional NN (left) versus DNN (right).

Figure 4

Conceptual framework of the newly proposed photonic neural network. The essential steps involved in achieving the desired task (a), an optical micrograph showcasing the distinctive structure of the proposed network, featuring nine input ports (i₁–i₉) and four output ports (o₁–o₄) [48] (b), an optical micrograph that zooms in on a single cell within the network, housing two phase shifters and a merging structure [48] (c).

Figure 5

The developed architectures of complex-valued optical neural architectures using: (a,b) MZIs, the circuits themselves realize a multiport interferometer with phase shifters (PSs) inserts used for phase tuning [49,50]; (c) MRRs for matrix-vector multiplication (MVM) applications using WDM [51].

Figure 6

CNN image classification pipeline.

Figure 8

Comparison of the von Neumann architecture with the neuromorphic architecture. Inspired by [1].

Figure 9

The fully integrated configuration of the curved neuromorphic imaging device is depicted in the following illustrations: (a) A photograph of the integrated imaging system, which comprises a plano-convex lens, cNISA, and housing. The inset provides a view of the components before they are assembled [130]. (b) An exploded diagram illustrating the components of the curved neuromorphic imaging device [130]. (c) A photograph of cNISA positioned on a concave substrate [130]. (d) A schematic representation of the custom-designed data acquisition system utilized for measuring the photocurrents of individual pixels in cNISA. (e–h) Demonstrations showcasing the process of obtaining a pre-processed image from a large set of noisy optical inputs. This includes the acquisition of a pre-processed C-shaped image (i), the gradual fading of the memorized C-shaped image (ii), the erasure of any residual afterimage (iii), and the acquisition of a pre-processed N-shaped image (iv), (e) displays the applied optical and electrical inputs, while (f) shows the obtained images at various time points [130], (g,h), the photocurrents recorded from specific pixels at each time point can be observed [130].

Figure 10

The contrast between the output generated by a neuromorphic vision sensor and a traditional frame-based camera when observing a spinning disk with a black dot. Compared to the regular frame-based camera, which transmits entire images with a consistent delay, the neuromorphic vision sensor emits events independently and without a fixed schedule, corresponding to the moments when they occur [132].

Figure 11

(a) Events gathered within a 10 ms time span [149]. (b) Events collected over a 20 ms time interval [149]. (c) Events compiled during a 30 ms time duration [149].

Figure 12

(a) Schematic diagram of the developed photonic circuit using IM-MRR technology [151]; (b) structural diagram of the photonic chip including all elements to realize the operation and illustration of the process of splitting the original image into pixel components and feeding data to each layer of the network [154].