All-optical neural network

Abstract

The present disclosure relates to an all-optical neural network that uses light beams and optical components to implement the layers of a neural network. The all-optical neural network includes an input layer, zero or more hidden layers, and an output layer. Each layer of the neural network is configured to simulate the linear and non-linear operations of a conventional artificial neural network neuron on an optical signal. In one embodiment, the optical linear operation is performed by a spatial light modulator and an optical lens. The optical lens performs a Fourier transform on the set of light beams and sums the light beams having similar propagation directions. The optical non-linear operation is implemented using a non-linear optical medium that has the property of electromagnetically induced transparency, and the transmission of the probe beam of the non-linear optical medium is controlled by an intermediate output of the coupled beam from the optical linear operation.

Description

[0001] Priority Claim

[0002] This application claims the benefit of U.S. Provisional Application No. 62 / 834,005, titled "All-Optical Neural Network," filed on April 15, 2019, the entire content of which is incorporated herein by reference. Technical Field

[0003] This disclosure relates to deep learning artificial neural networks. More specifically, this disclosure is directed to artificial neural networks implemented using optical components to perform computations using light as a medium. Background Art

[0004] In the past few decades, machine learning based on artificial neural networks (ANNs) has seen significant growth. Machine learning provides a general technique for systems to learn from data and make decisions with minimal human intervention. As a machine learning algorithm, an ANN is a computational model based on the neural structure of the brain, in which a set of connected nodes called artificial neurons are implemented. With the extensive interconnection of a large number of artificial neurons, an ANN can perform complex tasks in a powerful way. It has been demonstrated that machine learning based on artificial neural networks is very powerful in various fields such as image recognition, medical diagnosis, and machine translation. In the field of scientific research, ANNs also show great potential, especially in discovering new materials, classifying phases of matter, representing changing wave functions, accelerating Monte Carlo simulations, and various other applications. Moreover, ANNs can be used to solve problems that are difficult to solve by traditional methods.

[0005] An ANN can be implemented by software simulation in an electronic computer, where various complex algorithms can be applied. However, an ANN with a large number of artificial neurons and interconnections requires huge computational resource requirements, such as high energy consumption and long training times for the learning process. As the scale of the neural network increases, the computational complexity may increase exponentially.

[0006] Hardware solutions for implementing an ANN may be able to significantly reduce the execution time. For example, a circuit built with transistors can perform a large number of computations through a physical process of summing currents or charges. However, circuit solutions are vulnerable to noise and process parameter variations, which may limit the computational accuracy.

[0007] Photons, as non-interacting bosons, can be naturally used to achieve multiple interconnections at the speed of light and perform parallel computing simultaneously once applied to the implementation of ANNs. The key component of an ANN is the artificial neuron, which performs linear and non-linear transformations on signals. In a hybrid optical neural network (ONN), optics has been used to implement the linear transformation. However, the non-linear transformation function is usually implemented electronically because it has been proven challenging to implement the non-linear transformation function optically. Therefore, with the research on optical neural networks, it is necessary to solve this problem or other problems. Summary of the Invention

[0008] The all-optical neural network includes an input layer, zero or more hidden layers, and an output layer. Each layer includes one or more optical neurons, which are configured to process one or more light beams as inputs to the optical neurons through linear and non-linear transformations. The activation signal of each optical neuron can be weighted by a set associated with the interface between this layer of the neural network and the subsequent layer of the neural network. Both the linear and non-linear transformations are processed by optical components and non-linear optical media. In particular, the non-linear transformation is achieved by controlling the transmission of the second (probe) light beam according to the transmission of the first (coupling) light beam by utilizing the electromagnetically induced transparency (EIT) property of the non-linear optical medium.

[0009] In a first aspect of the present disclosure, a system for implementing an optical neuron is disclosed. The system includes one or more light beams as inputs to the optical neuron, a linear subsystem, a non-linear subsystem, and one or more additional light beams as outputs of the optical neuron. A light source is configured to generate one or more light beams as inputs to the optical neuron. The linear subsystem is configured to perform an optical summation operation, which combines one or more light beams to generate a coupled light beam as an intermediate signal. The non-linear subsystem is configured to perform an optical non-linear operation based on the coupled light beam to generate an activation response signal. The non-linear subsystem includes a non-linear optical medium having an electromagnetically induced transparency (ETI) property, and the activation response signal includes a probe light beam transmitted through the non-linear optical medium, which is non-linearly controlled by the intermediate signal. One or more additional light beams are separated from the activation response signal.

[0010] In some embodiments, the non-linear subsystem includes a probe laser configured to generate the probe light beam directed to the non-linear optical medium such that the transmission of the probe light beam through the non-linear optical medium is controlled based on the coupled light beam.

[0011] In some embodiments, the linear subsystem includes an optical lens configured to perform a Fourier transform that combines one or more light beams to generate an intermediate signal.

[0012] The optical lens generates at least two intermediate signals for two or more optical neurons by combining light beams having similar propagation directions. Each optical neuron corresponds to a specific propagation direction, and each intermediate signal is located at a different position on the focal plane of the optical lens.

[0013] In some embodiments, the system further includes a spatial light modulator (SLM) configured to modulate one or more additional light beams through a set of weights to generate one or more weighted light beams as the output of the optical neurons.

[0014] In some embodiments, the SLM is modulated using a weighted Gerchberg - Saxton (GSW) algorithm. The system further includes a photoelectric sensor to measure the output from the SLM.

[0015] In some embodiments, the set of weights is learned by a trained neural network based on a set of training data. The neural network includes an input layer, one or more hidden layers, and an output layer. The set of weights is associated with the interface between the input layer and the first hidden layer of the one or more hidden layers, between the hidden layers of the one or more hidden layers and the subsequent hidden layer of the one or more hidden layers, or between the hidden layer and the output layer.

[0016] In some embodiments, the propagation of the probe light beam in the nonlinear optical medium is controlled at least by the intensity or frequency of the coupled light beam.

[0017] In some embodiments, the nonlinear optical medium includes at least one of atoms, molecules, quantum dots, or solid - state materials.

[0018] In another aspect of the present disclosure, a system for implementing an all - optical neural network (AONN) is disclosed. The system includes: an input layer including one or more optical neurons; zero or more hidden layers; and an output layer including one or more optical neurons. Each hidden layer includes one or more optical neurons

[0019] In some embodiments, each optical neuron in the input layer includes one light beam received as an input to the optical neuron. Additionally, each optical neuron in the output layer includes one light beam transmitted as the output of the optical neuron.

[0020] In some embodiments, the system is configured to implement a method for realizing an AONN. The method includes the following steps: generating one or more light beams as inputs to optical neurons in an input layer; performing an optical linear operation on the outputs of optical neurons in a layer to generate one or more light beams as inputs to optical neurons in a subsequent layer of the AONN; and performing an optical nonlinear operation to generate a nonlinear activation response signal for the optical neurons. The optical nonlinear operation is implemented using a nonlinear optical medium having electromagnetically induced transparency (EIT) characteristics.

[0021] In some embodiments, the method for realizing an AONN further includes: for each optical neuron in a layer of optical neurons, using a spatial light modulator (SLM) to modulate the activation response signal of the optical neuron through a set of weights to generate a weighted output signal as an input to a subsequent layer of the AONN.

[0022] In some embodiments, the method for realizing an AONN further includes: using a photoelectric sensor to capture the weighted output signal; and constructing a light source to generate one or more additional light beams according to the weighted output signal to realize a subsequent layer of the AONN, where the subsequent layer is a hidden layer or an output layer in one or more hidden layers.

[0023] In some embodiments, one or more light beams are generated by a spatial light modulator (SLM), and the spatial light modulator is configured to spatially modulate at least one of the amplitude and phase of the coupled light beam incident on the surface of the SLM. In one embodiment, a weighted Gerchberg-Saxton (GSW) algorithm is used to tune the SLM.

[0024] In some embodiments, a probe light beam is directed at the nonlinear optical medium such that the transmission of the probe light beam through the nonlinear optical medium is controlled based on the coupled light beam. The power of the portion of the probe light beam transmitted through the nonlinear optical medium corresponds to the activation response signal.

[0025] In some embodiments, the optical linear operation is at least partially performed by an optical lens, and the optical lens is configured to combine one or more light beams having similar propagation directions on the focal plane of the optical lens.

[0026] In another aspect of the present disclosure, an apparatus for establishing a neural network model using light beams and an optical medium is disclosed. The apparatus includes: one or more light beams; at least one optical component; a nonlinear optical medium; a probe beam; and one or more additional light beams. The at least one optical component is configured to combine the one or more light beams to generate a coupled beam as an intermediate signal. The nonlinear optical medium has an electromagnetically induced transparency (ETI) characteristic that is controlled according to the intermediate signal. The probe beam is directed at the nonlinear optical medium such that the transmission of the probe beam through the nonlinear optical medium is controlled based on the coupled beam.

[0027] In some embodiments, the at least one optical component includes at least one of a lens, a wave plate, or a diffraction grating.

[0028] In some embodiments, the neural network model is established by iteratively simulating the layers of the neural network such that the weighted output signals of one or more optical neurons of a particular layer correspond to the inputs of one or more optical neurons of a subsequent layer. Each layer includes one or more optical neurons, and the neural network includes an input layer, zero or more hidden layers, and an output layer. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1A An optical neuron according to some embodiments is shown.

[0030] Figure 1B A general layered optical neural network according to some embodiments is shown.

[0031] Figure 1C An implementation of an optical neuron using a series of optical components according to some embodiments is shown.

[0032] Figure 2 A flowchart of a method for operating an optical neuron according to some embodiments is shown.

[0033] Figure 3 An optical linear subsystem configured to perform a linear operation within an optical medium according to some embodiments is shown.

[0034] Figure 4A An optical nonlinear subsystem configured to perform a nonlinear operation within an optical medium according to some embodiments is shown.

[0035] Figure 4B A corresponding Λ-shaped energy level diagram of a nonlinear optical medium according to one embodiment is depicted.

[0036] Figure 4C A corresponding Λ-shaped energy level diagram of a nonlinear optical medium according to another embodiment is depicted.

[0037] Figure 4D depicts the corresponding V-shaped energy level diagram of a nonlinear optical medium according to one embodiment.

[0038] Figure 4E depicts the corresponding trapezoidal energy level diagram of a nonlinear optical medium according to one embodiment.

[0039] Figure 5 illustrates a system configured to establish an all-optical neural network model according to some embodiments.

[0040] Figure 6 illustrates a control system configured to operate an optical system to establish an all-optical neural network model according to some embodiments.

[0041] Figure 7 illustrates an exemplary computer system according to some embodiments.

[0042] Figure 8A illustrates a flowchart of a method for training an all-optical neural network according to some embodiments.

[0043] Figure 8B illustrates a method for constructing an SLM using a weighted Gerchberg-Saxton algorithm according to some embodiments.

[0044] Figure 9 illustrates according to some embodiments Figure 5 a timing diagram of sampling the system of Detailed Description

[0045] Optical components can be used to simulate a layer of artificial neurons in an optical path to perform linear and nonlinear transformations within an optical medium. The optical amplitude or phase can be modulated at various positions of the signal (e.g., multiplying the components of the input signal by weights), and then the modulated optical signals are combined to achieve linear summation (e.g., combining the weighted components to produce a weighted sum). Different diffraction gratings can be used to implement different optical neurons of the layer to separate the components of the input signal according to various optical neurons in the layer. The optical nonlinear transformation can be performed by adjusting the nonlinear optical medium to have electromagnetically induced transparency (EIT) characteristics, which have a function similar to a traditional activation function when illuminated by the output of the linear transformation. In other words, the coupled beam representing the output of the linear transformation excites the nonlinear optical medium such that the probe beam can be transmitted through the nonlinear optical medium according to the energy of the coupled beam.

[0046] Compared with a hybrid optical neural network, which can perform an optical linear combination before digitizing intermediate results to apply a non - linear transformation to a conventional electronic or software implementation of a neural network, a fully optical neural network can be designed such that both linear and non - linear operations can be optically implemented for the neural network. This also allows for more complex optical neural networks to be implemented as multiple cascaded layers, such that the output of the neural network is available at the speed of light. For example, an optical path can be designed where instances of two or more artificial neuron layers are simulated by alternating different linear and non - linear subsystems along the path. As light propagates through the path, the output of one layer is processed by the next layer, and so on, until the output of the last layer is produced, without having to store the intermediate results of the hidden layers. Of course, other implementations can also implement a single layer of artificial neurons, and this single layer of artificial neurons can be re - constructed multiple times to implement subsequent layers of the neural network. In such an embodiment, each time a different layer of the neural network is implemented, and the optical image at the output of that layer is captured and reproduced for subsequent times.

[0047] Figure 1A FIG. 100 shows an optical neuron 100 according to some embodiments. The optical neuron includes a linear subsystem 102 and a non - linear subsystem 104. The optical neuron 100 receives one or more input signals u j , and combines the input signals to produce an intermediate output signal z i . The input signals and the intermediate output signal are light beams. The non - linear subsystem 104 operates on the intermediate output signal z i and produces an activation response signal of the optical neuron Then, the activation response signal is transmitted to a subsequent layer of the neural network, and a separate and distinct weighting W i,k is applied to the activation response signal corresponding to the target optical neuron in the subsequent layer of the neural network.

[0048] In one embodiment, the linear subsystem 102 is implemented as a summation that can be performed by one or more optical lenses that combine separate and distinct light beams for each of the one or more input signals. In other words, the linear subsystem 102 implements the following equation function:

[0049] z i =∑u j (Equation 1)

[0050] The non - linear subsystem 104 operates on the intermediate output signal z i to produce what is referred to as an activation response signal Nonlinear output. In one embodiment, the characteristics of the nonlinear optical medium are utilized to achieve a nonlinear transformation by based on the energy state conversion of the nonlinear optical medium caused by the incident coupled beam. The transmission of the probe beam controlled by the coupled beam through the nonlinear optical medium is called electromagnetically induced transparency (EIT). In the absence of the coupled beam, the nonlinear optical medium is opaque and blocks the transmission of the probe beam through the nonlinear optical medium. The coupled beam corresponding to the output of the linear subsystem 102 is used to cause an energy conversion in the nonlinear optical medium, making the nonlinear optical medium transparent and allowing the probe beam to be transmitted through the nonlinear optical medium. Utilizing the EIT characteristic to achieve a nonlinear transformation in the optical medium is similar to the conventional activation function implemented in the software implementation of an ANN. Therefore, the combination of the linear subsystem and the nonlinear subsystem in a single optical path can achieve the all-optical implementation of an ANN.

[0051] Generally, those skilled in the art should understand that the basic structure of the optical neuron can be different from Figure 1A the structure shown. For example, in some embodiments, the weighting W i,j can be applied to the input u j entering the neuron, rather than the activation response signal sent from the neuron As another example, the summation implemented by the linear subsystem 102 can include the bias value or the previous state of the optical neuron in addition to one or more input signals (e.g., for the recurrent neural network (RNN) implementation of a neural network).

[0052] Figure 1B FIG. shows a general layered optical neural network 150 according to some embodiments. The all-optical neural network (AONN) 150 includes an input layer, zero or more hidden layers, and an output layer. Each layer includes a plurality of optical neurons 110, 120, and 130 in the input layer, one or more hidden layers, and the output layer, respectively. The optical neuron 110 in the input layer is different from the optical neuron 120 in the hidden layer in that each optical neuron 110 has only one input signal u j . The optical neuron 130 in the output layer is different from the optical neuron 120 in the hidden layer in that each optical neuron 130 has only one output signal It should be understood that the number of optical neurons 110 in the input layer corresponds to the number of discrete input signals, and the number of optical neurons 130 in the output layer corresponds to the number of discrete output signals. In addition, the number of optical neurons 120 in each hidden layer can vary between hidden layers.

[0053] Figure 1CShows an implementation of an optical neuron 180 utilizing a series of optical components according to some embodiments. The linear summation operation transforms the coupled laser beam 190 using an optical lens 184. As Figure 1C shown, multiple light beams are combined at positions on the focal plane 186 to produce a combined intermediate light beam that represents the sum of the multiple input light beams. It should be understood that in other embodiments, more than one optical lens or other optical elements may be implemented to shape the incident light according to the optical system requirements.

[0054] The optical lens 184 sums light beams having similar propagation directions into one light spot on its front focal plane 186. The size of the output light spot is determined by the size of the input light beams and the focal length of the optical lens 184. The interference of multiple light beams modifies the spatial profile of the output light spot. The power of the output light spot is the integral over the entire light spot area, which, according to the law of conservation of energy, is the linear sum of the total power of the interfering light beams.

[0055] As Figure 1C shown, the optical lens 184 performs a Fourier transform and sums all diffracted light beams corresponding to the same propagation direction onto the light spot at its front focal plane 186 as a linear sum z i = ∑u j . At the front focal plane 186, multiple light spots z i can be presented at different positions (x, y) on the plane 186, with each light spot corresponding to a different optical neuron in the layer.

[0056] Then, the combined intermediate output signal light beam incident on the nonlinear optical medium 188 serves as the coupling light beam for the EIT of the nonlinear optical medium 188. The probe light beam 191 passes through the nonlinear optical medium, the transparency of the nonlinear optical medium is controlled by the combined intermediate output signal light beam, and the transmitted probe light beam becomes the nonlinear activation response signal

[0057] The spatial light modulator (SLM) 182 is configured to spatially modulate the light beam of the activation response signal to divide the activation response signal into one or more weighted light beams having different orientations. The SLM 182 can be programmable.

[0058] In one embodiment, the SLM 182 includes a plurality of light modulation elements (pixels) that modulate the amplitude, phase, and / or polarization of light incident thereon. A transmissive SLM and a reflective SLM can be implemented in the optical neuron 180. In some embodiments, the SLM includes a screen formed by microdisplay pixels composed of liquid crystal molecules. By setting a voltage signal for each pixel, the orientation of the liquid crystal molecules of each pixel can be rotated relative to a fixed angle. In such an embodiment, the amplitude of the incident light transmitted through different pixels of the SLM 182 or reflected by different pixels of the SLM 182 can be modulated based on the relative rotation of the liquid crystal molecules with respect to one or more polarization films (e.g., a polarized glass substrate) of adjacent liquid crystal molecules. In other embodiments, a phase-only SLM such as a liquid crystal on silicon SLM is utilized in an optical setup to modulate the phase of the light transmitted or reflected by each pixel. By modulating the phase of the light, different light beams can be combined to produce constructive or destructive interference, rather than directly modulating the amplitude of a discrete coherent light source.

[0059] In one embodiment, each pixel of the SLM 182 includes a phase grating (e.g., a diffraction grating). The multi-phase grating inside the SLM 182 modulates the incident light transmitted through each pixel to generate superimposed plane waves with different propagation directions, where each direction corresponds to a different optical neuron in the AONN 150 layer. In this way, more than one optical neuron model can be established within the optical medium in the same optical path.

[0060] In one embodiment, the SLM 182 is utilized to modulate the phase of the probe light. Assuming the SLM 182 is placed on the xy plane and illuminated by one or more light beams simultaneously, the complex amplitude of the reflected or transmitted light is where E p (x,y) is the amplitude of the incident light, φ0 is the phase of the incident light, and φ(x,y) is the phase change caused by the SLM 182. The phase modulation φ of the pixels in the SLM 182 can be controlled independently. In Figure 1C the embodiment, the complex amplitude of the coupled laser beam 190 at the rear focal plane of the lens 184 is The intermediate output plane is at the front focal plane 186 of the lens 184. In this configuration, the lens 184 performs a linear transformation. The output complex amplitude of the light beam at the front focal plane 186 is given below as

[0061] in Equation 2 as

[0062]

[0063] Determine the phase modulation φ(x,y) with a predetermined weighting coefficient w ij to satisfy the target distribution |E(x,y)| 2The initial phase distribution setting φ(x,y) can be achieved by superimposing multiple different phase gratings, where the calculation is performed using the grating equation. The phase distribution φ(x,y) can be fine-tuned by applying a weighted Gerchberg-Saxton (GSW) algorithm, which includes the following steps.

[0064] The target intensity of the field incident on the SLM is modeled as I0(x,y). In an iterative process, the electric field in the plane of the SLM is transformed using a Fast Fourier Transform (FFT) Propagate through the effective lens to compute the field in the focal plane It is also possible to capture (e.g., measure) the output beam intensity in practice. If the difference between the target image intensity and the measured intensity is small enough, the phase mode can be used to drive the SLM; otherwise, the amplitude A n Replaced by using the correction factor g n The corrected target amplitude is where g n is the adaptive factor and is defined as g1=1. In the expression, δ is the feedback parameter from 0 to 1, I m is the measured intensity distribution. Then the field is transformed into propagates back to the SLM plane, thus obtaining the field in the SLM plane The calculated phase φ n+1 The new phase mode in the SLM plane is maintained, while the amplitude is incident on a Then, another iteration starts with the field

[0065] Figure 2 A flow chart of a method 200 for operating an optical neuron according to some embodiments is shown. The steps of method 200 are described herein as software (e.g., instructions) executed by a processor, a processing unit, or any other controller or device capable of directing the operation of one or more components (configured to model the layers of optical neurons in the all-optical neural network 150). However, in some embodiments, the steps can be performed by hardware or a combination of hardware and software, such as software executed on a processor that causes hardware to perform operations (e.g., control actuators or control SLMs, couple lasers, etc.). Of course, it should be appreciated that any system capable of performing the steps of method 200 is considered to be within the scope of the present disclosure.

[0066] In step 202, one or more light beams corresponding to the input signals of the optical neurons are generated. In one embodiment, the SLM is configured to modulate a coupled laser beam to generate multiple coherent light beams as the inputs of the optical neurons. In one embodiment, the multiple coherent light beams represent the activation levels of the previous layer of the optical neurons. The activation levels can be weighted according to the parameters of the AONN. In some embodiments, a single light beam or multiple light beams can be split into multiple coherent light beams via optical components, and each individual light beam can be modulated by the first SLM to match the desired level of the input signal.

[0067] In step 204, a linear operation is performed on the one or more light beams. According to some embodiments, one or more optical components are used to implement the linear operation, and the optical components include optical lenses, wave plates, diffraction gratings, and / or other linear optical components and systems. The linear operation combines the one or more light beams representing the inputs of the optical neurons to generate an intermediate output signal z i 。

[0068] In step 206, a non - linear operation is performed on the intermediate output signal generated by the linear operation. In some embodiments, a non - linear optical medium 188 is utilized to apply a non - linear activation function to the optical signal. In some embodiments, the non - linear activation function is implemented based on the EIT property of the non - linear optical medium, which is a coherent optical non - linearity. The non - linear operation generally involves two highly coherent light sources (e.g., lasers), which are tuned to interact with three quantum states of the material. In other embodiments, the non - linear activation function can also be implemented by controlling the number of particles (e.g., atoms and molecules) in a specific quantum or classical state. The non - linear activation function generates an activation signal (e.g., output signal) for the optical neuron, and the activation signal can be expressed as

[0069] In step 208, the activation signal can be modulated according to one or more weightings to generate a weighted output signal, which is transmitted to one or more optical neurons in the subsequent layer of the AONN. In certain optical neurons, such as the optical neurons in the output layer of the AONN, step 208 can be optional.

[0070] It should be understood that steps 202 to 208 can be repeated for one or more layers of the AONN, and the output of one layer as defined by the structure of the AONN is transmitted to the input of the subsequent layer.

[0071] Figure 3 An optical linear subsystem 300 configured to implement input layer generation and linear operations within an optical medium is shown according to some embodiments. As Figure 3As shown, for illustrative purposes, the operation from M to N is provided as an example. In other words, M input signals (e.g., M = 8) are processed by an optical neuron layer including N optical neurons to produce N intermediate output signals z i (e.g., N = 4). In one embodiment, a coherent light source 302 (such as a laser diode, a gas laser, a single-mode fiber (SMF) laser, etc.) emits a laser beam that propagates through an optical coupler 304 and is collimated by a collimating lens L1 306. The collimated laser illuminates the surface of the first SLM SLM1 308, on which the incident light is selectively reflected from the pixels of SLM1 308 to produce M individual light beams representing the M input signals. The patterns encoded on SLM1 308 can be a combination of diffraction gratings in different directions. Each pixel can be used to modulate the amplitude, phase, and / or polarization of the reflected light beam such that each light beam represents the activation level of the input signal provided to the layer of optical neurons.

[0072] The M light beams propagate through a 4-f optical lens system (shown as two lenses L2 310 and L3 312 in the Figure 3 embodiment), which has an aperture at the focal point to block unwanted light beams, and the M light beams are imaged on a second SLM SLM2 314. The stray light from the first SLM1 308 is blocked at the Fourier plane of L2 310. Each laser beam is imaged (e.g., directed) on a certain part of SLM2 314, where each part corresponds to a different optical neuron of the layer of optical neurons. Each SLM part can modulate the light such that different weights w ij are applied to the same input signal, where the weights w ij correspond to different optical neurons.

[0073] In one embodiment, each laser beam incident on the N individual parts of SLM2 314 diffracts into N plane waves in N different directions corresponding to N established optical neuron models. By Fourier transform, the plane waves in similar directions are summed by constructive or destructive interference to produce N light beams output on the back side of the optical lens L4 316, which are reflected by a mirror 318 onto a receiving plane 320. In this embodiment, a mirror 318 is placed on the optical path to direct the laser beam to the receiving plane 320, thereby forming an image of the output light beam, where the receiving plane 320 is placed at the effective focal plane of L4 316.

[0074] It should be understood that changes can be made in different embodiments Figure 3Optical components. For example, the mirror 318 can be omitted from the system, and the receiving plane 320 can be located at the rear focal plane of the optical lens L4 316. Additionally, the input signal can be generated by different light sources and modulated separately, rather than using the SLM1 308 to divide a single light beam into M individual light beams. For example, a backlight array can be placed behind the SLM2 314, where the backlight array includes multiple different light sources for illuminating one or more pixels, and each light source represents a different input signal. However, care must be taken to generate coherent (e.g., in-phase) light from all individual light sources, which may require the ability to tune the light sources.

[0075] In some embodiments, the receiving plane 320 includes a photoelectric sensor (e.g., a CCD sensor, a CMOS sensor, etc.), a digital camera, etc., for imaging the image transmitted through the optical lens L4 316. Then the input of the nonlinear subsystem of the image generation can be used. Optionally, the receiving plane 320 can be replaced by a system or component that directly performs a nonlinear transformation based on the output light beam.

[0076] Figure 4A An optical nonlinear subsystem 400 configured to implement a nonlinear operation in an optical medium 408 is shown according to some embodiments. The coupled laser beam 402 represents an intermediate output signal generated by the optical linear subsystem 300 for the optical neuron, and the probe laser beam 404 realizes the nonlinear transformation of the intermediate output signal based on the transparency of the nonlinear optical medium 408 caused by the coupled laser beam 402. The angle 406 between the coupled laser beam 402 and the probe laser beam 404 can vary between 0 degrees and 360 degrees. In one embodiment, the nonlinear optical medium 408 is cold atoms of a specific element, such as laser-cooled rubidium atoms ( 85 Rb). In one example, the rubidium atoms are laser-cooled to a temperature of about 0.00001 to 0.0001 Kelvin. In other embodiments, the nonlinear optical medium 408 can include hot atoms (e.g., atoms at about room temperature or above room temperature), or some other medium that can be used to implement an optical nonlinear transformation. In other embodiments, the EIT nonlinear activation function can be implemented in quantum dots and other solid-state materials.

[0077] Figure 4B A corresponding ∧-shaped energy level diagram of the nonlinear optical medium 408 is depicted according to one embodiment. As Figure 4BAs shown, the relative energies E1, E2, and E3 of states |1> (482), |2> (484), and |3> (486) are shown to illustrate the energy transformation between the states, where E1 < E2 < E3. The particles of the nonlinear optical medium 408 are prepared in state 482. The coupled laser beam 402 from the optical linear subsystem 300 operates to effect a transition between state |2> (484) and state |3> (486). The probe laser beam 404 operates to effect a transition between state |1> (482) and state |3> (486). The transmission of the probe laser beam 404 through the nonlinear optical medium 408 is controlled by the power of the coupled laser beam 402. In the absence of the coupled laser beam 402, the particles in the nonlinear optical medium 408 are opaque to the probe laser beam 404. When the coupled laser beam 402 is present, quantum interference between the transition paths results in an EIT transparency spectral window, where the peak transmission and bandwidth are controlled by the intensity of the coupled laser beam 402 (e.g., the output signal from the linear subsystem). The output of the probe laser beam 404 is given in Equation 3 as

[0078]

[0079] where, I p,输入 and I p,输出 are the intensities of the input coupled laser beam 402 and the output probe laser beam 404, OD is the depth of the nonlinear optical medium on the |1>-|3> (482 to 486) transition, γ ij is the phase shift rate between states |i>-|j>. Ω c is the coupling field Rabi frequency, the square of which is proportional to the coupled laser intensity . As shown in Equation 3, the intensity of the probe laser beam 404 is nonlinearly controlled by the coupled beam intensity. A nonlinear activation function is implemented by taking the coupled laser beam 402 intensity as the input and the intensity of the transmitted probe laser beam 404 as the output. Equation 3 shows that the nonlinear activation function is determined by OD and γ 12 .

[0080] Figure 4C depicts the corresponding ∧-shaped energy level diagram of the nonlinear optical medium 408 according to another embodiment. In Figure 4C the relative energies E1, E2, and E3 of states |1> (482), |2> (484), and |3> (486) are shown, where E2 < E1 < E3. Contrary to the energy level diagram of Figure 4B , the transition between state |2> (484) and state |3> (486) requires more energy than the transition between state |1> (482) and state |3> (486).

[0081] Figure 4Ddepicts the corresponding V-shaped energy level diagram of the nonlinear optical medium 408 according to one embodiment. In Figure 4D the relative energies E1, E2, and E3 of the states |1> (482), |2> (484), and |3> (486) are shown, where {E2, E1} > E3. Contrary to the Figure 4B and Figure 4C ∧-shaped energy level diagrams, the energies of the state |1> (482) and the state |2> (484) are higher than the energy of the state |3> (486), where the energy of the state |1> (482) can be less than, equal to, or greater than the energy of the state |2> (484) (e.g., E1 < E2, or E1 = E2, or E1 > E2).

[0082] Figure 4E depicts the corresponding trapezoidal energy level diagram of the nonlinear optical medium 408 according to one embodiment. In Figure 4E the relative energies E1, E2, and E3 of the states |1> (482), |2> (484), and |3> (486) are shown, where E1 < E3 < E2. Contrary to the Figure 4B , Figure 4C and Figure 4D energy level diagrams, the transition between the state |2> (484) and the state |3> (486) is opposite in direction to the transition between the state |1> (482) and the state |3> (486).

[0083] A dilute gas, a solid solution, or an atomic system of more unique states (e.g., a magneto-optical trap or a Bose-Einstein condensate) can be implemented to achieve EIT characteristics. EIT characteristics can be achieved in an electromechanical system, an optomechanical system, or a semiconductor nanostructure (e.g., a quantum well, a quantum wire, a quantum dot, and other solid-state materials), each of which can be incorporated into the nonlinear subsystem 400 to achieve a nonlinear transformation within the optical medium.

[0084] Figure 5 illustrates a system 500 configured to establish a model of an all-optical neural network (AONN) according to some embodiments. The system 500 is designed to establish a model of a layer of a neural network that includes N optical neurons that generate N output signals. Each of the N optical neurons is designed to receive M input signals. Each of the N outputs can be multiplied by a K weight to generate N x K weighted output signals, which can be propagated to a subsequent layer of the AONN.

[0085] An optical signal propagates along an optical path through various optical components (sometimes referred to as optical computing units). The coupled laser beam 522 from the fiber coupler is projected onto the surface of the SLM1 518 through the collimating lens L5 520, and the surface is divided into a plurality of sub-regions, each sub-region corresponding to one or more light modulation elements. The SLM1 518 is configured to modulate the coupled laser beam 522 to generate M x N input signals. The lens L4 516 acts as a linear subsystem 102 of the optical neuron by combining the M x N input signals into N intermediate output signals.

[0086] Then, the N intermediate output signals pass through the lens L3 514, the beam splitter 512, and the lens L2 510 before reaching the nonlinear optical medium 508. In one embodiment, the N intermediate output signals cause the nonlinear optical medium 508 to exhibit an EIT property that modulates the probe laser beam 502 according to a nonlinear function. The probe laser beam 502 passes through the collimating lens L1 506 before impinging on the side of the nonlinear optical medium 508 opposite to the N intermediate output signals. The portion of the probe laser beam 502 transmitted through the nonlinear optical medium 508 passes through the lens L2 510 and impinges on the beam splitter 512, which redirects the probe laser beam 502 to the second SLM SLM2 528. This portion of the probe laser beam 502 represents the activation signal for the optical neuron, and separate and distinct nonlinear optical media 508 can be implemented for the N optical neurons, such that the N separate and distinct activation signals correspond to the N optical neurons. The probe laser beam 502 passes through the lenses L6 523, L7 524, and L8 526 before impinging on the surface of the SLM2 528.

[0087] The SLM2 528 is configured to modulate the N activation signals, each activation signal being modulated by K weights such that the K-weighted activation signals correspond to K different optical neurons in the subsequent layer of the AONN. The N×K output signals from the SLM2 528 are guided through the lens 530 and focused on the receiving plane 532, where the signals can be captured and digitized by a photoelectric sensor or measured / recorded by any technically feasible technique for measuring light.

[0088] In one embodiment, the size of the beam directed at the surface of the SLM1 518 covers (e.g., overlaps or illuminates) a plurality of light modulation elements. This enables modulation of a single light source at different positions of the SLM1 518 to generate multiple input beams. Additionally, each light modulation element in the SLM1 518 can be associated with a specific diffraction grating that corresponds to one or more propagation directions that allow multiple optical neuron models to be established simultaneously on the same optical path.

[0089] It should be understood that during subsequent passes through the system 500, an AONN having an input layer, one or more hidden layers, and an output layer can be constructed by injecting the output of the system 500 at the receiving plane 532 as the modulated input signal generated at the SLM1 518. Thus, the control system can be used to configure the SLM1 518 and the SLM2 528 to sequentially process the inputs through each layer of the AONN in order to generate the output of the output layer of the AONN after multiple passes through the system 500.

[0090] Figure 6 Shown is a control system 600 configured to operate an optical system 500 to establish an all-optical neural network model according to some embodiments. As Figure 6 shown, a controller 610 can be implemented in the optical system to control and tune optical components to perform linear operations and non-linear operations. Instructions can be executed by the controller 610 to control and tune the optical system.

[0091] As an example, the linear operation 620 is implemented by a lens associated with the SLM, SLM1 622 (corresponding to the SLM1 518 in the system 500). The controller 610 generates a signal transmitted to the SLM 622 to adjust the light modulation elements in the SLM 622 to encode the input signal processed by the linear operation, and when decoding the signal, the SLM 622 adjusts the voltage applied to each light modulation element according to the decoded signal. In some embodiments, an image formed by the SLM can be sampled using, for example, a movable mirror and a photoelectric sensor to provide a feedback loop for ensuring the accuracy of the combined input signals in the linear operation. The image can be fed back to the controller 610 to tune the light modulation elements using a weighted Gerchberg-Saxton (GSW) algorithm.

[0092] In some embodiments, the non-linear operation 630 is implemented by one or more components associated with the non-linear optical medium 508, for example, by adjusting the intensity of the coupled laser beam 632 to control the output of the probe laser beam 634 to implement the non-linear operation 630. The controller 610 can be configured to generate control signals for each of these various components to implement non-linear transformations using the non-linear optical medium 508. It should be understood that there can be other optical components / subsystems controlled by the controller, such as tunable optical devices (e.g., gimbal mirrors, adjustable lenses, etc.) or cameras (e.g., photoelectric sensors and optical components) to capture images with certain exposure settings. Additionally, although not explicitly shown in Figure 6 the controller can also provide control signals to a second SLM, such as the SLM2 528, for applying weights to the activation signals of one or more optical neurons.

[0093] The controller 610 can be implemented using software executed by a processor (i.e., instructions stored in a memory or other computer-readable medium), such as a central processing unit (CPU), a digital signal processor (DSP), a microcontroller, an embedded processor, etc. In other embodiments, the controller 610 can be implemented in hardware, such as within an application-specific integrated circuit (ASIC) or a field-programmable gate array (FPGA). In other embodiments, the controller 610 can be implemented in any combination of software and / or hardware. In some embodiments, the hardware can include mechanical and / or electromechanical actuators, such as stepper motors, hydraulic cylinders, or pneumatic cylinders, etc. It is contemplated that any type of control system 600 for managing the operation of one or more components including the AONN is within the scope of the present disclosure.

[0094] Figure 7 An exemplary computer system 700 is shown in accordance with some embodiments. As Figure 6 shown, the controller 610 can be implemented by a computer system 700 as Figure 7 shown. The computer system 700 can include one or more CPUs 702 and an arithmetic logic unit 706 that performs arithmetic and logic operations, and the CPU 702 includes a control unit 704 that directs the operation of the processor 704. The computer 700 also includes a volatile memory 708 (e.g., DRAM) for storing data, and input peripherals 710 and output peripherals 712 that receive measurement data and send control and tuning signals.

[0095] It should be understood that the computer system 700 can include additional components, including, for example, non-volatile memories, such as hard disk drives (HDDs), solid-state drives (SDDs), flash memories, graphics processing units (GPUs), network interface controllers (NICs), etc.

[0096] In some embodiments, the computer system 700 is a programmable logic controller (PLC). The inputs 710 and outputs 720 can be I / O modules attached to the PLC.

[0097] Figure 8A A flowchart of a method for training an AONN in accordance with some embodiments is shown. With a conventional computer-implemented neural network, a neural network can be trained using a training data set. The training data set (i.e., a combination of training data) includes various instances of sample input signals and corresponding target output signals. Each instance of the input signal is processed by the neural network, and the output signal is compared with the target output signal to calculate a value representative of the difference based on a loss function. Then, the parameters of the neural network are adjusted in various ways (such as by gradient descent (e.g., backpropagation)) to minimize the difference. These techniques are known for conventional neural networks.

[0098] In step 802, a set of training data is received. The training data includes multiple instances of input signals provided to the input layer of the AONN and the true target output signals generated by the output layer of the AONN. For example, in one embodiment, the input signals are vectors of these values: these values indicate the modulation levels of the input laser beams for multiple optical neurons in the input layer of the AONN. The target output signals can include vectors of these values: these values represent the desired intensities of the detection laser beams for each of the multiple optical neurons in the output layer of the AONN.

[0099] In step 804, the input signals are processed by the AONN to produce output signals. In one embodiment, system 500 is configured to sequentially process each layer of the AONN using a set of stored weights associated with the interfaces between each subsequent layer. The weights can be initialized using random or pseudo-random values prior to the first training pass.

[0100] In step 806, the output signals from the output layer of the AONN are captured. In one embodiment, capturing the output signals involves sampling a photoelectric sensor arranged in the optical path after all layers of the AONN have been processed.

[0101] In step 808, a loss function is calculated. The loss function can take various forms, such as L1 loss (e.g., least absolute deviation) or L2 loss (e.g., least square error). The goal is to minimize the value of the loss function over the entire training set by adjusting the weights associated with each layer of the AONN.

[0102] In step 810, the parameters of the AONN are updated. In one embodiment, conventional techniques for updating the weights can be employed, such as backpropagation with gradient descent.

[0103] In step 812, if the training data set includes another training sample, the process described in steps 804 to 810 can be repeated for the additional training sample. If the training set has been exhausted, the training process is complete.

[0104] Step 804, as described above, involves constructing one or more SLMs based on a set of parameters (e.g., weights, input signal levels, etc.) such that the AONN learns the function of converting the input signals into the desired output signals. In the case of a conventional neural network implemented in digital logic executed by a processor, computational accuracy is usually not an issue. However, when these operations are performed in an optical medium, the accuracy is imperfect and in fact non-negligible. One technique for correctly encoding the SLM relies on an iterative method using the GSW algorithm.

[0105] Figure 8BShows a method for constructing an SLM using a weighted Gerchberg-Saxton (GSW) algorithm according to some embodiments. It should be understood that the linear operation depends to some extent on the phase difference of the light waves combined by the optical lens. Two light sources will combine constructively or destructively due to interference based on their phase difference. Therefore, the encoding pattern for the modulation of the SLM can be different from the target vector representing the desired modulation value. This difference is caused by the phase difference associated with the light transmitted or reflected through different pixels of the SLM due to different path lengths, optical aberrations in the components, etc. To compensate for these differences, an iterative process of constructing the SLM and then measuring the resulting output can be employed to ensure that the differences are reduced below an acceptable level.

[0106] In step 852, the SLM is constructed based on the target output vector. In one embodiment, the SLM is initially constructed by encoding the modulation signals of different components of the SLM. The target output vector can represent the desired input signal level to be generated by the SLM.

[0107] In step 854, the output of the SLM is measured. In one embodiment, a rotating mirror and a photoelectric sensor can be used to capture the intensity of the signal generated by the SLM after the optical lens. It should be understood that due to the phase difference of different light sources, the deflection of the optical path, and the optical aberrations in various optical components, the measured output may again be different from the target output vector.

[0108] In step 856, the SLM is updated based on the feedback parameter. In one embodiment, the encoding parameter is updated based on the following equation:

[0109]

[0110] where I t is the target output intensity, I m is the measured output intensity, g n is the iteration vector (e.g., the updated encoding parameter of the SLM). δ is the feedback parameter, such as 0.2, ranging from 0 to 1, for controlling the speed at which the encoding parameter converges during multiple iterations. In other embodiments, other techniques for controlling the convergence speed can be utilized, such as, where δ > 1.

[0111] In step 858, the error is calculated and the calculated error is compared with a threshold. The error can be calculated as follows:

[0112]

[0113] where j ranges from 0 to n. When the error is less than the threshold (e.g., ∈ jWhen (<0.05), the iteration stops. Otherwise, the method repeats steps 854 and 856, makes another measurement in steps 854 and 856, and updates the SLM using the newly calculated iteration vector in Equation 4.

[0114] Figure 9 A timing diagram for sampling the system 500 according to some embodiments is shown. Measuring the output signal (e.g., the intensity of the probing laser transmission across a 2D plane) requires following a certain procedure to ensure that the nonlinear optical medium is in the correct energy state at the start of the measurement. Although not explicitly shown in the drawings, in some embodiments, the nonlinear optical medium may require additional trap lasers and repump lasers. The trap lasers are used to cool the particles in the nonlinear optical medium, and the repump lasers are used to keep the particles in the cooling cycle. A decoupling laser is also included, and after the cooling cycle, the decoupling laser is used to depopulate other ground states that are not involved in the transition process related to the EIT characteristics.

[0115] In one embodiment, the cooling cycle is started by turning on the trap lasers and the repump lasers for a period of time. This period of time should be sufficient to stabilize the nonlinear optical medium in a known state. An example of the duty cycle time is 100 ms. In one embodiment, this period of time is 88 ms, but the exact period of time can vary. The repump lasers are turned off 50 μs before the end of the cooling cycle, and at this time, the trap lasers are also turned off. In addition, as the repump lasers are turned off, the decoupling laser is turned on for 50 s, and a trigger is sent to the photoelectric sensor configured to capture the output signal.

[0116] After a short delay (depending on the response time of the photoelectric sensor to the trigger signal), the exposure starts at the end of the cooling cycle. Among other considerations, the exposure time is set based on the dynamic range of the photoelectric sensor and the intensity of the coupling laser. The example exposure time is shown as 2 ms. After the exposure starts, the coupling laser and the probing laser are turned on so as to perform linear operations and nonlinear operations in the optical medium. The coupling laser and the probing laser can be turned on for, for example, 880 μs, and there may be a delay between the start of the exposure and the turning on or activation of the coupling laser and the probing laser. Alternatively, the coupling laser and the probing laser can be turned on before the start of the exposure. For example, as Figure 9 shown, after deactivating the repump lasers, the probing laser and the coupling laser are activated.

[0117] Exposure starts in response to a trigger signal and is delayed based on an electronic delay in the optoelectronic sensor logic. As long as the coupling laser and the detection laser are in an active state during the exposure and the overlapping time period is sufficient to reduce or minimize the noise ratio (SNR) of the measured output signal, the precise timing of the exposure associated with a sampling time period (e.g., 12 ms) is not critical. Reducing the activation time during the exposure will reduce the SNR and result in a lower measurement certainty. In addition, the overlapping time period may affect the dynamic range of the output signal, and the dynamic range of the output signal may be limited by the characteristics of the optoelectronic sensor. It should be understood that the timing diagram 900 is provided as a means to illustrate how measurements are made to capture the results of linear and non-linear operations performed entirely within the optical medium. Depending on the exact configuration of the system 500, other timing diagrams are possible.

[0118] Note that the techniques described herein can be embodied in executable instructions stored in a computer-readable medium, which is used by or in conjunction with a processor-based instruction execution machine, system, device, or apparatus. Those skilled in the art should understand that for some embodiments, various types of computer-readable media can be included to store data. As used herein, "computer-readable medium" includes one or more of any suitable medium for storing executable instructions of a computer program such that the instruction execution machine, system, device, or apparatus can read (or obtain) the instructions from the computer-readable medium and execute the instructions for implementing the described embodiments. Suitable storage formats include one or more of electronic, magnetic, optical, and electromagnetic formats. A non-exhaustive list of conventional exemplary computer-readable media includes: portable computer disks; random access memory (RAM); read-only memory (ROM); erasable programmable read-only memory (EPROM); flash memory devices; and optical storage devices, including portable optical discs (CDs), portable digital video discs (DVDs), etc.

[0119] It should be understood that the arrangement of the components shown in the drawings is for illustrative purposes and other arrangements are possible. For example, one or more of the elements described herein can be implemented in whole or in part as electronic hardware components. Other elements can be implemented in software, hardware, or a combination of software and hardware. Moreover, some or all of these other elements can be combined, some other elements can be completely omitted, and additional components can be added while still implementing the functions described herein. Therefore, the subject matter described herein can be embodied in many different variations, and all such variations are considered to be within the scope of the claims.

[0120] To aid understanding of the subject matter described herein, many aspects are described in terms of action sequences. Those skilled in the art will recognize that various actions can be performed by dedicated circuitry or circuitry, program instructions executed by one or more processors, or a combination of both. The description of any action sequence herein is not intended to imply that a particular order must be followed to perform the sequence. Unless otherwise indicated herein or clearly contradicted by context, all of the methods described herein can be performed in any suitable order.

[0121] Unless otherwise specified herein or clearly contradicted by context, the use of the terms "a," "an," and "the" and similar references in the context of describing the subject matter (especially in the context of the claims) should be construed to cover both the singular and the plural. Unless otherwise specified herein or clearly contradicted by context, the term "at least one" (e.g., "at least one of A and B") following a list of one or more items should be understood to mean either one item (A or B) selected from the listed items or any combination of two or more of the listed items (A and B). Further, the foregoing description is for illustrative purposes only and not for purposes of limitation, since the scope of the sought protection is defined by the claims and their equivalents. Unless otherwise stated, the use of any and all examples or exemplary language (e.g., "such as") provided herein is only intended to better illustrate the subject matter and does not limit the scope of the subject matter. In the claims and the description of the specification, the use of the term "based on" and other similar phrases indicating conditions that produce a result is not intended to exclude any other conditions that produce that result. No language in the specification should be construed as indicating any non-claimed element as essential to the practice of the claimed invention.

Claims

1. A system for implementing an optical neuron, the system comprising: One or more light beams, which serve as inputs to the optical neuron; A linear subsystem configured to perform an optical summation operation that combines one or more light beams to produce a coupled light beam as an intermediate signal; A non-linear subsystem configured to perform an optical non-linear operation based on the coupled light beam to produce an activation response signal, wherein the non-linear subsystem includes a non-linear optical medium having electromagnetically induced transparency (EIT) characteristics, and wherein the activation response signal includes a probe light beam transmitted through the non-linear optical medium, and the non-linear optical medium is non-linearly controlled by the intermediate signal; and One or more additional light beams, which serve as outputs of the optical neuron, wherein the one or more additional light beams are separated from the activation response signal.

2. The system according to claim 1, wherein, The non-linear subsystem includes: A probe laser configured to generate the probe light beam directed at the non-linear optical medium such that the transmission of the probe light beam through the non-linear optical medium is controlled based on the coupled light beam.

3. The system according to claim 1, wherein, The linear subsystem includes: An optical lens configured to perform a Fourier transform that combines the one or more light beams to produce the intermediate signal.

4. The system according to claim 3, wherein, The optical lens produces at least two intermediate signals for two or more optical neurons by combining light beams having similar propagation directions, each optical neuron corresponding to a specific propagation direction, and each intermediate signal being located at a different position on the focal plane of the optical lens.

5. The system according to claim 1, further comprising: A spatial light modulator (SLM) configured to produce one or more weighted light beams as outputs of the optical neuron by modulating the one or more additional light beams with a set of weights.

6. The system according to claim 5, wherein, The spatial light modulator is tuned using a weighted Gerchberg-Saxton (GSW) algorithm, and the system further includes a photoelectric sensor for measuring the output from the spatial light modulator.

7. The system according to claim 5, wherein, Learning the set of weights by training a neural network based on a set of training data, and wherein the neural network includes an input layer, one or more hidden layers, and an output layer, and the set of weights is associated with an interface between the input layer and a first hidden layer of the one or more hidden layers, between hidden layers of the one or more hidden layers and subsequent hidden layers of the one or more hidden layers, or between the hidden layer and the output layer.

8. The system according to claim 1, wherein, The transmission of the probe light beam in the non-linear optical medium is controlled at least by the intensity or frequency of the coupled light beam.

9. The system according to claim 1, wherein, The non-linear optical medium includes at least one of atoms, molecules, quantum dots, or solid-state materials.

10. A system for implementing an all - optical neural network (AONN), the system comprising: An input layer, which includes one or more optical neurons; Zero or more hidden layers, wherein each hidden layer includes one or more optical neurons; And An output layer, which includes one or more optical neurons, wherein a method for implementing the all-optical neural network includes: Generating one or more light beams as inputs to the optical neurons in the input layer; Perform an optical linear operation on the output of the optical neurons in a layer to generate one or more light beams as inputs to the optical neurons in a subsequent layer of the all-optical neural network; and Perform an optical nonlinear operation to generate a non-linear activation response signal for the optical neurons, wherein the optical nonlinear operation is implemented using a non-linear optical medium having electromagnetically induced transparency (EIT) characteristics.

11. The system according to claim 10, wherein, Each optical neuron in the input layer includes a light beam received as an input to the optical neuron, and wherein each optical neuron in the output layer includes a light beam transmitted as an output of the optical neuron.

12. The system according to claim 10, wherein, The method for implementing the all-optical neural network further includes: For each optical neuron in a layer of optical neurons, use a spatial light modulator (SLM) to modulate the activation response signal of the optical neuron through a set of weights to generate a weighted output signal as an input to a subsequent layer of the all-optical neural network.

13. The system according to claim 12, wherein, The method for implementing the all-optical neural network further includes: Capture the weighted output signal using a photoelectric sensor; and Construct a light source to generate one or more additional light beams according to the weighted output signal to implement a subsequent layer of the all-optical neural network, wherein the subsequent layer is a hidden layer or an output layer in the one or more hidden layers.

14. The system according to claim 10, wherein, The one or more light beams are generated by a spatial light modulator (SLM) configured to spatially modulate at least one of the amplitude and phase of an incident coupled light beam on the surface of the spatial light modulator.

15. The system according to claim 14, wherein, Use a weighted Gerchberg-Saxton (GSW) algorithm to tune the spatial light modulator.

16. The system according to claim 14, wherein, Direct a probe light beam at the non-linear optical medium such that the transmission of the probe light beam through the non-linear optical medium is controlled based on the coupled light beam, and wherein the power of the portion of the probe light beam transmitted through the non-linear optical medium corresponds to the activation response signal.

17. The system according to claim 10, wherein, The optical linear operation is at least partially performed by an optical lens configured to combine one or more light beams having similar propagation directions on a focal plane of the optical lens.

18. An apparatus for establishing a neural network model using a light beam and an optical medium, the apparatus comprising: One or more light beams; At least one optical component configured to combine the one or more light beams to generate a coupled light beam as an intermediate signal; A non-linear optical medium having electromagnetically induced transparency (ETI) characteristics controlled according to the intermediate signal; A probe light beam directed at the non-linear optical medium such that the transmission of the probe light beam through the non-linear optical medium is controlled based on the coupled light beam; And One or more additional light beams as outputs of the optical neurons of the neural network, wherein the one or more additional light beams are separated from the activation response signal of the optical neurons.

19. The apparatus according to claim 18, wherein, The at least one optical component includes at least one of a lens, a wave plate, or a diffraction grating.

20. The apparatus according to claim 18, wherein, A model of the neural network is established by iteratively simulating the layers of the neural network such that the weighted activation signals of one or more optical neurons of a particular layer correspond to the inputs of one or more optical neurons of a subsequent layer, wherein each layer includes one or more optical neurons, and the neural network includes an input layer, zero or more hidden layers, and an output layer.

Images