High-efficiency reconfigurable all-optical neural network computing chip architecture for deep learning

Abstract

The invention discloses a high-efficiency reconfigurable all-optical neural network computing chip architecture for deep learning, comprising an input layer, a hidden layer and an output layer; the input layer is composed of an intensity-coded MZI array, and a plurality of input signals of different wavelengths are input into the hidden layer after being intensity-coded; the hidden layer comprises a programmable linear computing module and a nonlinear module; the programmable linear computing module is realized by an MZI array capable of constructing an arbitrary matrix; wherein the phase shifter in the MZI array is replaced by a plurality of tunable over-coupled micro-ring resonators to realize parallel computing, and the number of wavelength channels that can be calculated in parallel is determined by the number of micro-ring resonators; the nonlinear module is composed of a plurality of cascaded Add‑drop type micro-ring resonators, and nonlinear output is realized by the nonlinear effect of the micro-ring resonators; the output layer performs photoelectric conversion on the output signal of the hidden layer; the invention can solve the problems of insufficient integration and parallel computing capability of the current chip.

Description

Technical Field

[0001] The present invention belongs to the technical field of integrated optoelectronics, and specifically relates to a high-efficiency reconfigurable all-optical neural network computing chip architecture for deep learning. Background Art

[0002] Artificial intelligence, as one of the most active fields in information science, has important applications in machine vision, autonomous driving, target tracking and other fields. Neural networks, as one of the most important models in artificial intelligence, are used in various scenarios because of their good generalization ability and robustness. Integrated circuit (IC) chips are the hardware carriers for training and testing mainstream neural network models today, such as central processing units (CPUs), graphics processing units (GPUs), field programmable gate arrays (FPGAs) and application-specific integrated circuits (ASICs). However, no matter which IC chip is used, it is based on the "von Neumann" computing system, which has the characteristics of separation of computing and storage, resulting in the computing efficiency of neural networks being greatly limited. In some real-time control application scenarios, high power consumption and slow speed have become the main problems hindering the application of deep learning neural network technology.

[0003] Photonic technology is a technology that uses photons as the basic carrier for information transmission and processing. Compared with traditional electronic computing, photonic computing directly uses optical signals to achieve high-speed information processing and complex operations in the optical domain. Its main advantages include: (1) Information transmission using light as a carrier has the characteristics of large bandwidth, low latency and ultra-high speed; (2) It can achieve multi-dimensional parallel processing and has the potential for large-capacity processing; (3) Low power consumption. Therefore, hardware acceleration technology based on photonic computing is a very promising disruptive innovation technology.

[0004] In recent years, with the vigorous development of silicon-based photonic integration technology, scientific research teams at home and abroad have made great progress in the development of photonic computing chips. The photonic computing chip architecture based on the photonic integration platform is mainly divided into two categories. The first category is the Mach-Zehnder interferometer (MZI) array architecture based on optical interference; the second category is the architecture based on the microring resonator array. In 2012, researchers from the Institute of Semiconductors of the Chinese Academy of Sciences developed a microring resonator array device, and combined with wavelength division multiplexing technology (WDM), they realized high-speed matrix-vector multiplication operations. In 2016, a research team from the Massachusetts Institute of Technology developed the first all-optical neural network chip for deep learning. The linear computing module of this silicon-based photonic computing chip is constructed by cascaded Mach-Zehnder interferometers (MZIs), which can use matrix decomposition algorithms to meet general matrix-vector product operations, and its nonlinear computing module is implemented off-chip. In 2017, researchers at the University of Oxford in the UK developed a photonic synaptic structure based on phase change materials and photonic integrated devices to complete optical neuromorphic computing.

[0005] In 2019, a research team from the University of Münster in Germany developed the first all-optical integrated and scalable neuromorphic framework for building pulsed neural networks based on microring resonators and phase change materials, and proved that the system can complete typical artificial intelligence pattern recognition tasks. In 2020, researchers from Huazhong University of Science and Technology in my country realized multifunctional optical signal processing, including filtering, optical switching, matrix operations, etc., based on a universal MZI cascade network array structure. In addition, research teams from Shanghai Jiao Tong University and Swinburne University of Technology have also done a lot of representative research work in optical computing based on discrete optoelectronic devices.

[0006] At present, the MZI-based two-dimensional topology network structure has programmability, reconfigurability and certain versatility, but the nonlinear activation function module has not been integrated with the linear calculation module, and the power consumption is large, and parallel calculation cannot be achieved. The array structure based on microring resonators has high integration and low power consumption, but lacks versatility.

[0007] In summary, existing solutions have two key problems: (1) weak parallel computing capabilities and (2) insufficient chip architecture integrity. Summary of the invention

[0008] In view of this, the present invention provides a high-efficiency reconfigurable all-optical neural network computing chip architecture for deep learning, which can solve the problems of insufficient integration and parallel computing capabilities of current chips.

[0009] The technical solution for implementing the present invention is as follows:

[0010] High-efficiency reconfigurable all-optical neural network computing chip architecture for deep learning, including input layer, hidden layer and output layer connected from left to right;

[0011] The input layer is composed of an intensity-coded MZI array. Multiple input signals of different wavelengths are intensity-coded by the intensity-coded MZI array and then input into the hidden layer.

[0012] The hidden layer includes a programmable linear calculation module and a nonlinear module;

[0013] The programmable linear computing module is implemented by an MZI array that can build any matrix. The phase shifters in the MZI array are replaced by multiple tunable overcoupled microring resonators to achieve parallel computing. The number of wavelength channels that can be calculated in parallel is determined by the number of microring resonators.

[0014] The nonlinear module is composed of multiple cascaded Add-drop type micro-ring resonators, and the nonlinear output of the output signal of the programmable linear calculation module is realized through the nonlinear effect of the micro-ring resonator;

[0015] The output layer performs photoelectric conversion on the output signal of the hidden layer.

[0016] Furthermore, the intensity coding is implemented through matrix decomposition theory, and its signal transmission process is: out1 =M1·E in , E in represents the input signal, and M1 represents the intensity coding matrix.

[0017] Furthermore, the signal transmission process of the programmable linear calculation module is: out2 =M2·E out1 , M2 is an arbitrary matrix, E out1 Represents the output signal of the input layer, E out2 Represents the output signal of the programmable linear calculation module.

[0018] Furthermore, the arbitrary matrix is ​​split into a unitary matrix, a diagonal matrix, and a complex conjugate of the unitary matrix through singular value decomposition, and the programmable linear calculation module is divided into three parts: a unitary matrix, a diagonal matrix, and a complex conjugate of the unitary matrix.

[0019] Furthermore, the output layer is implemented by a photodetector.

[0020] Beneficial effects:

[0021] 1. The present invention is based on the MZI two-dimensional topology network array structure, and realizes multi-wavelength multiplexing on the basis of versatility, solves the problem of parallel computing, and compared with traditional phase shifters, overcoupled microring resonators are more sensitive to thermo-optic and electro-optic modulation, and the device length and power consumption can be reduced by 1-2 orders of magnitude, greatly improving the computing energy efficiency of the network.

[0022] 2. The nonlinear effect of the Add-drop type microring resonator of the present invention makes it possible to integrate the linear computing module and the nonlinear functional module on the same chip, which is beneficial to the further integration and miniaturization of the photonic computing chip. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a schematic diagram of the overall structure of the high-performance reconfigurable all-optical neural network computing chip architecture for deep learning of the present invention.

[0024] Figure 2 Schematic diagram of the MZI unit structure of the programmable linear computing module of the high-efficiency reconfigurable all-optical neural network computing chip architecture for deep learning of the present invention.

[0025] Figure 3 Schematic diagrams of amplitude spectrum and phase spectrum of three coupling states of the microring resonator used in the present invention; (a) schematic diagram of amplitude spectrum, (b) schematic diagram of phase spectrum.

[0026] Figure 4 This is a schematic diagram of the linear module light output results of the present invention that realizes independent regulation and parallel calculation when light of different wavelengths passes through different microring resonators.

[0027] Figure 5 It is a schematic diagram of the nonlinear functional module structure of the microring resonator of the present invention.

[0028] Figure 6 Schematic diagram of the nonlinear effect of the microring resonator of the present invention. DETAILED DESCRIPTION

[0029] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0030] like Figure 1 As shown, the present invention provides a high-efficiency reconfigurable all-optical neural network computing chip architecture for deep learning, including an input layer, a hidden layer (including a programmable linear computing function module and a nonlinear module), and an output layer connected in sequence from left to right.

[0031] The input layer consists of an intensity-coded MZI array, λ1, λ2, λ3, ... j They respectively represent signals of different wavelengths input to the intensity-coded MZI array, which are intensity-coded by the intensity-coded MZI array and then input to the hidden layer.

[0032] The intensity coding module is also implemented by matrix decomposition theory (see the hidden layer programmable linear calculation module for the specific principle), and its signal transmission process can be written as E out1 =M1·E in , E in represents the input signal, and M1 represents the intensity coding matrix.

[0033] The hidden layer is composed of programmable linear calculation function modules and nonlinear modules.

[0034] The programmable linear calculation function module is implemented by the MZI array that can construct any matrix. Its signal transmission process is E out2 =M2·E out1 , M2 is an arbitrary matrix, E out1 Represents the output signal of the input layer; wherein, the phase shifter in the MZI array is replaced by multiple tunable over-coupled microring resonators to achieve parallel computing, and the number of wavelength channels that can be calculated in parallel is determined by the number of microring resonators.

[0035] According to matrix analysis theory, an arbitrary matrix can be decomposed by singular value decomposition, that is Where U is an m×m unitary matrix, ∑ is an m×n diagonal matrix, is the complex conjugate of the n×n unitary matrix V. Therefore, the programmable linear computing function module in the hidden layer can be divided into three parts: the unitary matrix, the diagonal matrix, and the complex conjugate of the unitary matrix. Each MZI consists of two beam splitters with a splitting ratio of 50% and two phase shifters (θ, φ). By controlling the MZI with two phase shifters, the unitary matrix transformation between the two input and two output ports can be realized:

[0036]

[0037] Based on the tunable MZI basic unit, a unitary matrix array structure can be constructed. Parallel computing is achieved by replacing the phase shifter with a tunable overcoupled microring resonator. The number of wavelength channels that can be calculated in parallel is determined by the number of microrings in the unit structure.

[0038] Figure 2 The schematic diagram of the MZI unit structure of the programmable linear computing module of the high-efficiency reconfigurable all-optical neural network computing chip architecture for deep learning of the present invention is shown. The MZI unit structure assisted by microrings can realize wavelength division multiplexing and then perform parallel computing. Here, three rings are taken as an example (the number of channels that can be calculated in parallel is equal to the number of microrings). Each microring resonant cavity is independently controlled by electrodes, and its corresponding working wavelength can be flexibly adjusted. In order to achieve as dense wavelength division multiplexing as possible, the size of the microring resonator can be further reduced. After preliminary simulation and calculation, its FSR can reach a maximum of 90nm.

[0039] Figure 3 (a) and (b) are schematic diagrams of the amplitude spectrum and phase spectrum of the three coupling states of the microring resonator used in the present invention. Figure 3 It can be seen that when the microring resonator is in an overcoupled state, its through output spectrum has a 2π phase shift near the resonant wavelength, and the amplitude attenuation in the overcoupled state is low. Therefore, by virtue of its wavelength-dependent phase and amplitude distribution characteristics in the overcoupled state, an independent phase shift can be introduced for each wavelength channel. According to sufficient research and theoretical estimation, compared with the straight waveguide phase shifter, the proposed overcoupled microring resonator device is more sensitive to thermo-optical and electro-optical modulation, and the ring-assisted phase shifter can reduce the device length and power consumption by 1-2 orders of magnitude.

[0040] Figure 4 This is a diagram of the light output result of the linear module of the present invention when different wavelengths of light pass through different microring resonators to achieve independent regulation and parallel calculation. It can be seen that the light intensity output by different wavelengths of light after independent regulation of their corresponding microring resonators is different, indicating that parallel calculation has been achieved.

[0041] The nonlinear function module is composed of a plurality of cascaded Add-drop type micro-ring resonators, and the nonlinear output of the output signal of the programmable linear calculation function module is realized through the nonlinear effect of the micro-ring resonator.

[0042] The nonlinear function module is based on the nonlinear effect dynamic mechanism of silicon-based microrings, including the thermo-optical effect (TOE-FCA, TOE-TPA) caused by carrier absorption and two-photon absorption, which causes the change of the complex refractive index of the microring in the entire optical power range. The nonlinear optical phenomenon caused by this nonlinear effect can be directly reflected in the output spectrum of the microring resonator, and the Relu6 function can be simulated. The signal is output after passing through the nonlinear module.

[0043] Figure 5 The nonlinear functional module structure diagram of the microring resonant cavity of the present invention is shown in FIG. After light of different wavelengths and different light intensities enters the corresponding microring cavity, nonlinear output light can be achieved due to the nonlinear effect of the microring cavity at the resonant wavelength and the difference in input light power.

[0044] Figure 6 The figure is a schematic diagram of the nonlinear effect of the microring resonant cavity of the present invention. The thermo-optical effect (TOE-FCA, TOE-TPA) caused by carrier absorption and two-photon absorption causes the change of the complex refractive index of the microring in the entire optical power range. The nonlinear optical phenomenon caused by this nonlinear effect can be directly reflected in the output spectrum of the microring resonant cavity: when the input pump light power is small, the transmission spectrum of the microring presents a symmetrical inverse Lorentzian line distribution; when the input pump light power is large, the resonance peak drift and asymmetric spectrum can be observed. Theoretical simulation found that at the resonant wavelength, the microring transmission light power changes nonlinearly with the input light power. Figure 6 This is the curve of the transmission light power at the micro-ring drop end at the resonant wavelength measured in the experiment versus the input light power. This curve has good mathematical properties and is similar to the Relu6 nonlinear activation function in the entire domain, which can meet the nonlinear calculations in the all-optical neural network.

[0045] The output layer is realized by a photodetector, which performs photoelectric conversion on the output signal of the hidden layer.

[0046] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A high-performance reconfigurable all-optical neural network computing chip for deep learning, characterized by: It includes the input layer, hidden layer and output layer connected from left to right; The input layer is composed of an intensity-coded MZI array. Multiple input signals of different wavelengths are intensity-coded by the intensity-coded MZI array and then input into the hidden layer. The hidden layer includes a programmable linear calculation module and a nonlinear module; The programmable linear computing module is implemented by an MZI array that can build any matrix. The phase shifters in the MZI array are replaced by multiple tunable overcoupled microring resonators to achieve parallel computing. The number of wavelength channels that can be calculated in parallel is determined by the number of microring resonators. The nonlinear module is composed of multiple cascaded Add-drop type micro-ring resonators, and the nonlinear output of the output signal of the programmable linear calculation module is realized through the nonlinear effect of the micro-ring resonator; The output layer performs photoelectric conversion on the output signal of the hidden layer.

2. The high-performance reconfigurable all-optical neural network computing chip for deep learning according to claim 1, characterized in that: The intensity coding is realized by matrix decomposition theory, and its signal transmission process is E out1 =M1·E in , E in represents the input signal, M1 represents the intensity coding matrix, E out1 Represents the output signal of the input layer.

3. The high-performance reconfigurable all-optical neural network computing chip for deep learning according to claim 1, characterized in that: The signal transmission process of the programmable linear calculation module is: out2 =M2·E out1 , M2 is an arbitrary matrix, E out1 Represents the output signal of the input layer, E out2 Represents the output signal of the programmable linear calculation module.

4. The high-performance reconfigurable all-optical neural network computing chip for deep learning according to claim 1, characterized in that: The arbitrary matrix is ​​split into a unitary matrix, a diagonal matrix, and a complex conjugate of the unitary matrix by singular value decomposition, and the programmable linear calculation module is divided into three parts: a unitary matrix, a diagonal matrix, and a complex conjugate of the unitary matrix.

5. The high-performance reconfigurable all-optical neural network computing chip for deep learning according to claim 1, characterized in that: The output layer is implemented by a photodetector.

Images