An all-optical control reservoir computing system and method

Abstract

This invention discloses a fully optically controlled reservoir computing system and method. The system includes a reservoir layer and an output layer. The reservoir layer includes an encoding module and a fully optically controlled synaptic array. The encoding module performs bipolar or parallel encoding on the input signal using suppressed and enhanced light to obtain encoded light pulses. These encoded light pulses are then alternately or simultaneously irradiated onto the fully optically controlled synaptic array. The fully optically controlled synaptic array receives the corresponding encoded light pulses and obtains multiple photocurrent values, thereby completing feature extraction of the input signal. The output layer is a neural network trained using gradient descent, used to receive the multiple photocurrent values ​​and complete the identification or prediction of the input signal. This system significantly improves the expressive power and multi-source information fusion capability of reservoir computing.

Description

Technical Field

[0001] This invention belongs to the field of optoelectronic technology and neuromorphic computing, specifically relating to a fully optically controlled reservoir computing system and method. Background Technology

[0002] With the rapid development of the Internet of Things (IoT) and smart terminals, edge computing, as a distributed computing paradigm, is receiving increasing attention. Its main goal is to load computational tasks onto front-end hardware at the data generation point, thereby reducing latency, saving bandwidth, and improving the real-time processing performance of the system. Pool computing is a computational framework inspired by recurrent neural networks, replacing traditional trainable recurrent layers with fixed, nonlinear pool layers. This pool layer, as a dynamic system, maps input signals to a high-dimensional computational space, making pool computing extremely efficient when processing temporal signals. Since only the output layer weights need to be trained, this framework significantly reduces computational complexity and hardware resource requirements, making it a highly promising technology in the field of edge intelligence.

[0003] To further improve the energy efficiency of reservoir systems in resource-efficient information processing, researchers are increasingly turning to physical hardware implementations, utilizing the inherent dynamic evolution of physical systems to achieve high-dimensional and efficient mapping of input signals. Among these, optoelectronic reservoir computing offers advantages such as wide photonic bandwidth, low crosstalk, and strong electronic compatibility, allowing for seamless integration with in-sensor computing. However, traditional optoelectronic reservoir computing faces two major limitations: First, most devices exhibit unipolar photoresponse characteristics. Chinese patent CN 119152909 A discloses a method for reservoir computing using memristor unipolar photoresponse, but unipolar photoresponse devices result in insufficient reservoir state diversity and limited nonlinear mapping capabilities, affecting the accuracy of complex task processing. Second, multi-source information processing relies on the serial operation of multiple independent reservoirs, leading to high hardware resource consumption and energy consumption, making it difficult to adapt to edge environments with limited hardware resources.

[0004] While existing technologies employ hybrid photoelectric signal encoding strategies to improve performance—for example, Chinese Patent CN116883713 A discloses a method for reservoir computing using hybrid encoding of memristor unipolar optical and electrical signals—this increases system complexity and energy consumption. Furthermore, a single device cannot simultaneously meet the demands of rich nonlinear dynamics and multi-source signal processing, limiting the application of photoelectric reservoir computing in edge intelligence. Therefore, developing a fully optically controlled reservoir computing system that requires no external electrical control, possesses rich dynamic characteristics, and supports multi-source fusion is of great significance. Summary of the Invention

[0005] This invention provides a fully optically controlled reservoir computing system, which significantly improves the expressive power and multi-source information fusion capabilities of reservoir computing.

[0006] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention achieves bipolar encoding by alternately irradiating a fully photocontrolled synaptic array with suppression and enhancement light. Compared with unipolar encoding, the bipolar encoding provided by this invention utilizes the positive continuous photoconductivity of the enhancement light and the negative continuous photoconductivity of the suppression light to obtain positive and negative photocurrent values, thus broadening the distribution range of the reservoir state, reducing the probability of feature overlap, and extracting richer features, which is beneficial to improving the calculation accuracy.

[0007] This invention achieves parallel encoding by simultaneously irradiating a fully optically controlled synaptic array with both suppressing and enhancing light, thereby enabling the simultaneous processing of more digital signals, improving computational efficiency, and outputting relatively less data. This reduces the size requirements of the output layer neural network and significantly reduces the number of neurons in the output layer compared to bipolar encoding, resulting in a substantial reduction in hardware resource overhead. Attached Figure Description

[0008] Figure 1 (a) shows an optical microscope image (scale bar: 1 mm) of the 16×16 ZnO fully optically controlled memristor array of the present invention on the left. Figure 1 (a) shows an enlarged view of the 3×3 subarray of the present invention (scale bar: 50 micrometers) on the right; Figure 1 (b) shows a schematic diagram of the working mode of the fully optically controlled memristor array of the present invention on the left and a schematic diagram of the change of photocurrent (ΔI) of the fully optically controlled memristor of the present invention with time (t) on the right. The device exhibits bidirectional optical response. Figure 1 (c) A comparison diagram of unipolar encoding, serial encoding and bipolar encoding and parallel encoding provided by the present invention in traditional reservoir calculation; Figure 2 (a) The photoresponse of the device provided in a specific embodiment of the present invention after being irradiated with light of different wavelengths. The illustration shows an enlarged view of the device under 405 nm and 650 nm illumination. Figure 2 (b) The device provided in a specific embodiment of the present invention achieves bidirectional photoresponse diagrams using independent illumination modes at different optical power densities; Figure 2 (c) The relaxation behavior of photocurrent of the device provided in the specific embodiment of the present invention in the alternating illumination mode. After applying a 405 nm light pulse, the device applies a 650 nm light illumination with gradually increasing light power density. The curve corresponding to w / o in the figure represents the photoresponse under 405 nm light illumination only. P in the figure is the light power density. Figure 2(d) is a photoresponse diagram of the device provided in a specific embodiment of the present invention in the simultaneous illumination mode, wherein the device simultaneously applies a 405 nm light pulse with a fixed light power density and a 650 nm light pulse with a different light power density; Figure 2 (e) is a photocurrent diagram of the device provided in a specific embodiment of the present invention under continuous light pulse alternating irradiation mode, where the light pulse width is 0.15 s; Figure 2 (f) is a photocurrent diagram of the device provided in a specific embodiment of the present invention in the mode of simultaneous illumination by continuous light pulses, where the light pulse width and pulse interval are both 0.15 s.

[0009] Figure 3 (a) A schematic diagram of bipolar coding for word recognition provided in a specific embodiment of the present invention; Figure 3 (b) A reservoir state diagram obtained under bipolar coding mode for 16 different input combinations, provided for a specific embodiment of the present invention; Figure 3 (c) A reservoir state diagram obtained under unipolar coding mode for 16 different input combinations, provided for a specific embodiment of the present invention; Figure 3 (d) A comparison chart of recognition accuracy under bipolar and unipolar coding modes provided in a specific embodiment of the present invention; Figure 4 (a) A schematic diagram of time series bipolar coding provided for a specific embodiment of the present invention; Figure 4 (b) A diagram illustrating the bipolar coding process provided in a specific embodiment of the present invention; Figure 4 (c) and Figure 4 (d) are the prediction results of x, y and z dimensions under bipolar and unipolar encoding, respectively, provided by specific embodiments of the present invention.

[0010] Figure 5 (a) is a schematic diagram of multi-source information fusion in the parallel coding mode provided by the present invention; Figure 5 (b) and Figure 5 (c) is a graph showing the variation of ΔI in the parallel coding mode provided by the present invention. The graph uses different combinations of 405 nm and 650 nm light pulses, and the optical power density of the 405 nm light is 0.8 mW / cm². 2 The optical power density of 650 nm light is 8 mW / cm². 2 The pulse width was 1 second and the pulse time interval was 0.3 seconds. For comparison, the figure also shows the variation of ΔI under a single 405 nm or 650 nm light pulse. Figure 5 (d) ΔI distribution diagram of all 4-bit parallel coding combinations. The statistical value of ΔI is the photocurrent 0.3 seconds after parallel coding. Figure 5 (e) Distribution of ΔI standard deviation obtained after three parallel coding tests for all combinations; Figure 5 (f) is a comparison chart of the recognition accuracy under unipolar, bipolar and parallel coding of the present invention. Detailed Implementation

[0011] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Operating methods not specifically specified in the following embodiments are generally performed under conventional conditions or as recommended by the manufacturer.

[0012] A specific embodiment of the present invention provides a fully optically controlled reservoir computing system, including a reservoir layer and an output layer. The reservoir layer includes a fully optically controlled synaptic array. The conductance of the fully optically controlled synaptic array is mapped to the reservoir state, i.e., multiple photocurrent values, and then transmitted to the output layer for identification or prediction. The output layer is implemented by artificial synaptic hardware or neural network algorithms.

[0013] The reservoir layer provided in a specific embodiment of the present invention includes an encoding module and a fully optically controlled synaptic array. The encoding module performs bipolar or parallel encoding on the input signal by suppressing and enhancing light to obtain encoded light pulses. These encoded light pulses are then alternately or simultaneously irradiated onto the fully optically controlled synaptic array. The fully optically controlled synaptic array receives the corresponding encoded light pulses and obtains multiple photocurrent values, thereby completing the feature extraction of the input signal. The output layer is a neural network trained using the gradient descent method, used to receive the multiple photocurrent values ​​and complete the recognition or prediction of the input signal.

[0014] The fully optically controlled synapse provided in this invention is an optoelectronic device with a bidirectional continuous photoconductivity effect. It simulates the "memory-learning-computation" function of a biological synapse by controlling the device's conductivity state through optical signals, with the core being the integration of "sensing-storage-computation." This fully optically controlled synapse device includes fully optically controlled memristors, fully optically controlled transistors, and fully optically controlled heterojunction synapse devices, etc. This invention uses a fully optically controlled memristor as an example. Artificial synapses also include devices with structures such as memristors, transistors, and heterojunction synapse devices.

[0015] In one specific embodiment, the enhanced light is ultraviolet light or blue light, the wavelength of the enhanced light is 300~495nm, and the power density is 0.01~10 mW / cm². 2 The suppressing light is green or red light, with a wavelength of 500-750 nm and a power density of 0.1-100 mW / cm².2 .

[0016] The fully optically controlled memristor provided in the specific embodiments of the present invention is composed of a bottom electrode layer, an intermediate dielectric layer and a top electrode layer from bottom to top.

[0017] The enhanced light irradiation of the fully optically controlled memristor array exhibits positive continuous photoconductivity, while the suppressed light irradiation of the fully optically controlled memristor array exhibits negative continuous photoconductivity. By alternately or simultaneously irradiating the bottom or top electrode of the fully optically controlled memristor array with enhanced and suppressed light, feature values ​​of the input signal can be extracted.

[0018] like Figure 1 (a) The left side shows an optical microscope image of the all-optically controlled memristor array. Figure 1 (a) The right side shows an enlarged view of the 3×3 subarray in the fully optically controlled memristor array. The fully optically controlled memristor provided in this embodiment consists of a bottom electrode layer 1, an intermediate dielectric layer 2 and a top electrode layer 3 from bottom to top.

[0019] The fully optically controlled memristor provided in this embodiment includes two electrode ports: a top electrode and a bottom electrode, and an intermediate dielectric layer located between the two electrodes. Both the top electrode and the bottom electrode are made of Pt material, and the intermediate dielectric layer is a ZnO thin film.

[0020] like Figure 1 (b) The left side shows a schematic diagram of the operating mode of the all-optically controlled memristor array. Figure 1 (b) The right side shows a schematic diagram of the change of ΔI of the fully optically controlled memristor with t. Light of different wavelengths is irradiated onto the device, and the reading voltage V is... r Used for real-time monitoring of current changes.

[0021] In one specific embodiment, this embodiment provides a method for obtaining multiple photocurrent values ​​by irradiating a fully optically controlled memristor array with bipolar encoded light pulses. This includes: encoding the input signal into enhanced and suppressed light pulses using enhanced and suppressed light of different wavelengths, or enhanced and suppressed light of different power densities at a constant wavelength; and dynamically evolving the photocurrent value in the positive and negative ranges by alternately irradiating the fully optically controlled memristor array with enhanced and suppressed light pulses. This invention broadens the conductance range through this bipolar encoding method. Compared with unipolar encoding, which only has a positive or negative conductance range, it achieves both positive and negative conductance ranges, enriching the extraction of feature information and improving prediction accuracy.

[0022] Specifically, the bipolar encoding provided in the specific embodiments of the present invention requires converting the input signal into an encodeable feature signal. The input signal includes a binary signal, a numerical signal, or a spatial pattern signal. The feature signal contains at least two logical states or numerical polarities. Preferably, the specific embodiments of the present invention select a binary signal ("1 / 0") as the input signal.

[0023] The input signal provided in this embodiment is a binary signal. A "1" bit in the binary signal corresponds to an enhanced light pulse, and a "0" bit corresponds to a suppressed light pulse; alternatively, a "0" bit in the binary signal corresponds to an enhanced light pulse, and a "1" bit corresponds to a suppressed light pulse. Furthermore, this embodiment assigns a "1" bit to blue light with a wavelength of 405 nm and a "0" bit to red light with a wavelength of 650 nm. By alternately inputting these two codes, the state of the reservoir dynamically evolves within the positive and negative conductance ranges.

[0024] Another bipolar encoding method provided in a specific embodiment of the present invention is as follows: a constant wavelength is used to select an optical signal with a large optical power density to correspond to an enhanced optical pulse, and an optical signal with a small optical power density to correspond to a suppressed optical pulse; a constant wavelength is used to select an optical signal with a small optical power density to correspond to an enhanced optical pulse, and an optical signal with a large optical power density to correspond to a suppressed optical pulse.

[0025] In one specific embodiment, the method provided in this embodiment, which obtains multiple photocurrent values ​​by irradiating a fully optically controlled memristor array with parallel encoded optical signals, includes: encoding the input signals corresponding to the multi-source signals into enhanced optical pulses and suppressed optical pulses respectively, and then simultaneously irradiating the corresponding fully optically controlled memristor arrays with the enhanced optical pulses and suppressed optical pulses to achieve parallel encoding. Through the optical coupling effect, multiple photocurrent values ​​are obtained, thereby completing the fusion and feature extraction of multi-source signals.

[0026] The parallel encoding provided in this embodiment utilizes enhancement light and suppression light with different wavelengths or the same wavelength but different power densities to encode the input signal as enhancement light and suppression light pulses. By simultaneously inputting enhancement light and suppression light pulses, the state of the reservoir can be dynamically evolved in the positive and negative conductance range. By using a constant read voltage to monitor the change in device conductance in real time, the nonlinear mapping of the input signal to the high-dimensional reservoir state can be achieved.

[0027] Specifically, when the input signal is a binary signal and the encoding method is parallel encoding, the "1" bit in the binary signal corresponding to each source signal corresponds to the illumination light pulse, and the "0" bit in the binary signal corresponds to the unilluminated light pulse.

[0028] like Figure 1 As shown in (c), the unipolar coding in traditional reservoir calculation uses a single wavelength of light for coding, and the serial coding refers to coding independently in time order. The bipolar coding provided in this application uses different wavelengths of light for coding, and the parallel coding uses different wavelengths of light for coding simultaneously.

[0029] Specifically, the optoelectronic performance of the devices was characterized using a semiconductor parameter analyzer equipped with a monochromatic light source, a 405 nm laser, and a 650 nm laser. The monochromatic light source provided 350 nm illumination. An arbitrary waveform generator controlled the alternating and simultaneous illumination of 405 nm and 650 nm light pulses. During testing, a read voltage was applied to the top electrode of the device, and the bottom electrode was grounded. The optical signal was input through the top electrode layer, inducing a bidirectional continuous photoconductivity effect at the synapses. The memristor in this embodiment exhibited positive continuous photoconductivity under 405 nm blue light illumination and negative continuous photoconductivity under 650 nm red light illumination. Furthermore, when illuminating the two wavelengths alternately or simultaneously, these devices exhibited dynamically tunable bidirectional photoresponse behavior, a characteristic that enabled us to demonstrate bipolar coding and parallel coding schemes on a single device platform for the first time.

[0030] like Figure 2 As shown in (a), the memristor exhibits a significant positive photoresponse under ultraviolet light irradiation, such as Figure 2 (a) As shown in the illustration, after exposure to ultraviolet light, the device exhibits positive continuous photoconductivity to blue light and negative continuous photoconductivity to red light.

[0031] like Figure 2 As shown in (b), the dependence of the bidirectional photoresponse of the device on the optical power density (P) under illumination at 405 nm and 650 nm is illustrated. The bidirectional photoresponse gradually increases with increasing power density at both wavelengths. When illumination stops, the photocurrent exhibits a significant spontaneous relaxation phenomenon.

[0032] like Figure 2 As shown in (c), after a positive photoresponse is induced by 405 nm light, 650 nm light irradiation will accelerate the decay of photocurrent. The relaxation rate can be controlled by adjusting the power density of 650 nm light.

[0033] like Figure 2 As shown in (d), when the device is simultaneously irradiated with 405 nm and 650 nm light, it exhibits diverse photoresponses and relaxation dynamics. Notably, compared to single-beam irradiation, dual-beam irradiation produces a strong nonlinear photoresponse with fluctuating characteristics, which may stem from the dynamic competition between oxygen vacancy ionization and neutralization processes under 405 nm and 650 nm light irradiation.

[0034] like Figure 2 As shown in (e), the photoresponse of the device is demonstrated when it is alternately irradiated with light pulses of different power densities of 405 nm and 650 nm.

[0035] like Figure 2 As shown in (f), the photoresponse of the device is demonstrated when it is simultaneously irradiated with light pulses of different power densities of 405 nm and 650 nm.

[0036] On the other hand, the present invention also provides a method for recognizing words using the above-mentioned all-optical-controlled reservoir computing system, comprising: Convert the word into a binary image, which is the input signal.

[0037] The encoding module encodes the "1" bits in the input signal into enhanced optical pulses and the "0" bits into suppressed optical pulses. The encoded optical pulses are then irradiated onto the fully optically controlled memristor array to complete bipolar encoding and obtain multiple photocurrent values.

[0038] The multiple photocurrent values ​​are transmitted to the output layer, and the output layer neural network, trained by gradient descent, yields the word recognition result.

[0039] In one specific embodiment, this embodiment utilizes the mapping relationship of "binary bit → dual-wavelength light pulse" in bipolar encoding to transform the pixel features of words into high-dimensional states that can be recognized by the reservoir, thus solving the recognition confusion problem caused by insufficient state space in the unipolar encoding reservoir.

[0040] like Figure 3 As shown in (a), (1) signal preprocessing is performed: the 13 letters are designed as 3×3 pixel binary patterns and combined into 20 four-letter words. Each word corresponds to 9 groups of four-bit binary data streams, thus obtaining the input signal. The 13 letters include: 'C', 'H', 'I', 'J', 'K', 'L', 'O', 'T', 'U', 'V', 'X', 'Y', and 'Z'. These form 20 four-letter words: 'COZY', 'HOLL', 'JOLT', 'KILO', 'LICK', 'LUCK', 'COOK', 'CITY', 'HOLY', 'OILY', 'YOLY', 'ITCH', 'TUCK', 'LOCK', 'COOL', 'TICK', 'HULK', 'KILL', 'VOLT', 'TOXI'.

[0041] (2) Bipolar encoding mapping: The encoding module maps the "1" bit in the input signal to a 405 nm blue light pulse, and the "0" bit to a 650 nm red light pulse. The optical power density of the 405 nm light is 0.8 mW / cm². 2 The optical power density of 650 nm light is 8 mW / cm². 2 The pulse width is 1 second, with alternating illumination without time intervals. Nine sets of four-bit binary data streams are converted into nine sets of coded optical pulses, thus completing bipolar encoding. The nine sets of coded optical pulses are then applied to a 3×3 subarray memristor to generate nine photocurrent values.

[0042] (3) Output layer training: Nine photocurrent values ​​are input into a 9×20 neural network readout layer, with the Sigmoid function as the activation function and 20 four-letter words as 20 labels. The weights are optimized using gradient descent based on the loss function to obtain the trained neural network readout layer. In application, the photocurrent value ΔI is input into the trained neural network readout layer to obtain the word prediction value. The above process is implemented using Python. Performance analysis: Bipolar encoding broadens the distribution range of the reservoir state and improves the accuracy of word recognition. For example, the confusion probability between "LOCK" and "LUCK" is reduced from 48% in unipolar encoding to 5%, and the final accuracy reaches 93%.

[0043] like Figure 3 As shown in (b), under bipolar coding, the reservoir state diagrams obtained by 16 different input combinations are distributed in the positive and negative conductance ranges.

[0044] like Figure 3 As shown in (c), in contrast, unipolar coding restricts the reservoir state to the positive conductance range, thereby reducing its separability.

[0045] like Figure 3 As shown in (d), the recognition accuracy of four-letter words under bipolar encoding reaches about 93% after 100 training cycles, while unipolar encoding can only reach about 76%.

[0046] On the other hand, the present invention also provides a method for predicting time-series signals using the above-mentioned all-optically controlled reservoir computing system, comprising: The time-series signal is transformed into a high-dimensional sequence based on a binary mask, thus obtaining the input signal; The negative values ​​in the input signal are mapped to the suppression light, and the optical power density of the suppression light is mapped to the magnitude of the negative values ​​in the input signal. The positive values ​​are mapped to the enhancement light, and the optical power density of the enhancement light is mapped to the magnitude of the positive values. This converts the timing signal into encoded optical pulses. The encoded optical pulses illuminate the all-optically controlled memristor array, thereby completing bipolar encoding and generating multiple photocurrent values. The multiple photocurrent values ​​are transmitted to the output layer, and the output layer neural network trained by the gradient descent method is used to obtain the prediction results of the time series signal.

[0047] In one specific embodiment, this embodiment utilizes the dynamic adjustment capability of bipolar coding "numerical features → dual-wavelength power mapping" in conjunction with a time-division multiplexing strategy to predict timing signals and improve prediction accuracy.

[0048] (1) Dataset generation: Based on the ordinary differential equation of the Lorentz system, 800 continuous data points were generated, 600 for training and 200 for testing, and the data was scaled to the interval [-1,1].

[0049] (2) such as Figure 4 As shown in (a), a binary mask of length 5, such as [1,-1,-1,1,-1], is used to convert the original single-valued signal (e.g., 0.8) into a high-dimensional sequence (0.8,-0.8,-0.8,0.8,-0.8); positive values ​​are defined as 405 nm light pulses, and the optical power density is 0.2~1.0 mW / cm². 2 The increment was 0.2 mW / cm 2 A negative value for a 650 nm light pulse has an optical power density of 2~10 mW / cm². 2 The increment is 2 mW / cm 2 Other original single-value signals, such as 0.4, -1, and -0.8, are processed in the same way to obtain the encoded optical pulses.

[0050] The encoded light pulses are then directed onto an all-optically controlled memristor array, generating multiple photocurrent values, such as... Figure 4 As shown in (b).

[0051] The method for obtaining the final predicted value is as follows: Based on the ordinary differential equation of the Lorentz system, 800 continuous data points are generated, 600 of which are used for training and 200 for testing. The data is scaled to the interval [-1,1].

[0052] Training involves using a binary mask of length 5 (e.g., [1,-1,-1,1,-1]) to convert the original single-valued signal (e.g., 0.8) into a high-dimensional sequence (0.8,-0.8,-0.8,0.8,-0.8), and then inputting the ΔI corresponding to different mask signals for training. Testing involves testing the ΔI corresponding to different input mask signals, outputting the network to obtain predicted values, and comparing these predicted values ​​with the original data to determine if the predictions are correct.

[0053] Comparative verification: such as Figure 4 (c) and Figure 4 As shown in (d), the red curve in the figure represents the predicted value, and the black curve represents 200 data points obtained based on the ordinary differential equation of the Lorentz system. By comparing the two, the prediction results of bipolar and unipolar coding are obtained. Their average errors (NRMSE) in the (x, y, z) dimensions are approximately 0.31 and 0.12, respectively, which confirms the superiority of the bipolar coding scheme.

[0054] On the other hand, the present invention also provides a method for recognizing facial and fingerprint signals using the aforementioned all-optical-controlled reservoir computing system, comprising: The facial and fingerprint information is converted into binary images to obtain the first source input signal and the second source input signal. The first source input signal is encoded into an enhancement light pulse by an encoding module, where the "1" bit in the first source input signal is mapped to enhancement light and the "0" bit is mapped to unirradiated enhancement light. The second source input signal is encoded into a suppression light pulse, where the "1" bit in the second source input signal is mapped to suppression light and the "0" bit is mapped to unirradiated suppression light. The enhancement light pulse and the suppression light pulse are simultaneously irradiated onto the all-optically controlled memristor array to generate multiple photocurrent values, thereby completing the feature extraction of the input signal through parallel encoding. The multiple photocurrent values ​​are transmitted to the output layer, and the output layer neural network trained by the gradient descent method is used to complete the identification of multi-source signals.

[0055] In one specific embodiment, this embodiment utilizes the coupling characteristics of parallel encoding's "multi-source signal → dual-wavelength simultaneous illumination" to complete the feature extraction and fusion of face and fingerprint signals within a single fully optically controlled memristor array, solving the problem of high hardware overhead in traditional serial encoding. Figure 5 As shown.

[0056] (1) such as Figure 5 As shown in (a), multi-source data preprocessing: 10 pairs of face and fingerprint images were collected from the FVC fingerprint database and the CASIA-WebFace face database, and the images were converted into 32×64 pixel binary images.

[0057] (2) Parallel coding mapping: A four-bit coding scheme is adopted, defining the first source signal (face) as a 405 nm blue light pulse, the second source signal (fingerprint) as a 650 nm red light pulse, and the optical power density of 405 nm light as 0.8 mW / cm². 2 The optical power density of 650 nm light is 8 mW / cm². 2 The pulse width is 1 second, and two light pulses are simultaneously applied to the fully optically controlled memristor array.

[0058] In this process, the "1" bit in the first source input signal is mapped to an irradiated blue light pulse, and the "0" bit is mapped to an unirradiated blue light pulse. The second source input signal is encoded as a red light pulse, where the "1" bit in the second source input signal is mapped to an irradiated red light pulse, and the "0" bit is mapped to an unirradiated red light pulse, thus completing the parallel encoding.

[0059] (3) Coupling and fusion mechanism: When illuminated by dual wavelengths of light simultaneously, the ionization (405 nm) and neutralization (650 nm) of oxygen vacancies within the fully optically controlled memristor dynamically compete, generating a coupled photoconductive response, directly outputting the fused reservoir state, such as... Figure 5 As shown in (b) and (c) in the figure.

[0060] Dataset and readout layer design: Generate 1000 training samples and 9000 test samples. The output layer uses a 512×100 neural network. Bipolar encoding requires a 1024×100 neural network, reducing hardware resource overhead by 50%.

[0061] The training process involves collecting 10 pairs of face and fingerprint images, converting them into 32×64 pixel binary images, and inputting four pixel values ​​each time. Faces are illuminated with blue light, and fingerprints with red light. Two light pulses are simultaneously applied to an all-optically controlled memristor array to obtain ΔI, which is then input into the output layer neural network for training. The testing process involves inputting ΔI into the output layer neural network to predict the correct combination of face and fingerprint images, ultimately obtaining the multi-source information recognition accuracy.

[0062] (5) Performance verification: The parallel coding accuracy reached 90.5%, which is comparable to bipolar coding (91%) and significantly better than unipolar coding (70%), achieving the goal of "high precision + low hardware consumption" of multi-source fusion.

[0063] We comprehensively evaluated the performance of 4-bit configuration parallel encoding by testing all 256 possible encoding combinations, such as... Figure 5 As shown in (d), the distribution of ΔI measured 0.3 seconds after encoding is completed is displayed.

[0064] like Figure 5 As shown in (e), the standard deviation (σ) of each encoding combination is presented. This data comes from three repeated tests, demonstrating the operational reliability of parallel encoding. The smaller σ value indicates that the parallel encoding strategy has strong reliability.

[0065] like Figure 5 As shown in (f), the recognition performance of different coding schemes was compared. Bipolar coding and parallel coding both achieved high accuracy rates of 91% and 90.5%, respectively, significantly outperforming unipolar coding, which only achieved 70%.

[0066] Furthermore, it should be understood that after reading the above description of the present invention, those skilled in the art can make various alterations or modifications to the present invention, and these equivalent forms also fall within the scope defined by the appended claims.

Claims

1. A fully optically controlled storage pool computing system, characterized in that, include: The reservoir layer includes an encoding module and a fully optically controlled synaptic array. The encoding module performs bipolar or parallel encoding on the input signal by suppressing and enhancing light to obtain encoded light pulses. The encoded light pulses are then alternately or simultaneously irradiated onto the fully optically controlled synaptic array. The fully optically controlled synaptic array receives the corresponding encoded light pulses and obtains multiple photocurrent values, thereby completing the feature extraction of the input signal. The output layer is a neural network trained by gradient descent, used to receive the multiple photocurrent values ​​and to identify or predict the input signal through the neural network.

2. The all-optical-controlled reservoir computing system according to claim 1, characterized in that, Multiple photocurrent values ​​were obtained by irradiating a fully optically controlled synaptic array with bipolar-coded light pulses, including: By using enhancement and suppression light of different wavelengths, or enhancement and suppression light of different power densities at a constant wavelength, the input signal is encoded into enhancement and suppression light pulses; by alternately irradiating the fully optically controlled synaptic array with enhancement and suppression light pulses, the dynamic evolution of the photocurrent value within the positive and negative range is achieved.

3. The all-optical-controlled reservoir computing system according to claim 2, characterized in that, The input signal is a binary signal. When the encoding method is bipolar encoding, the "1" bit in the binary signal corresponds to the enhanced light pulse, and the "0" bit in the binary signal corresponds to the suppressed light pulse. Alternatively, the "0" bit in the binary signal corresponds to the enhanced light pulse, and the "1" bit in the binary signal corresponds to the suppressed light pulse.

4. The all-optical-controlled reservoir computing system according to claim 1, characterized in that, Multiple photocurrent values ​​are obtained by illuminating a fully optically controlled synaptic array with parallel encoded optical signals. This process involves encoding the input signals corresponding to the multi-source signals into enhancement and suppression optical pulses, and then simultaneously illuminating the corresponding fully optically controlled synaptic arrays with the enhancement and suppression optical pulses to achieve parallel encoding. Through optical coupling, multiple photocurrent values ​​are obtained, thus completing the fusion and feature extraction of the multi-source signals.

5. The all-optical-controlled reservoir computing system according to claim 4, characterized in that, The input signal is a binary signal. When the encoding method is parallel encoding, the "1" bit in the binary signal corresponding to each source signal corresponds to the illumination light pulse, and the "0" bit in the binary signal corresponds to the unilluminated light pulse.

6. The all-optical-controlled reservoir computing system according to claim 1, characterized in that, The enhancement light is ultraviolet or blue light, with a wavelength of 300-495 nm and a power density of 0.01-10 mW / cm². 2 The suppressing light is green or red light, with a wavelength of 500-750 nm and a power density of 0.1-100 mW / cm². 2 ; The fully optically controlled synapse includes a fully optically controlled memristor, a fully optically controlled transistor, or a fully optically controlled heterojunction synapse device. The fully optically controlled memristor is composed of a bottom electrode layer, an intermediate dielectric layer, and a top electrode layer from bottom to top. The enhanced light irradiation of the fully optically controlled memristor array exhibits positive continuous photoconductivity, while the suppressed light irradiation of the fully optically controlled memristor array exhibits negative continuous photoconductivity. By alternately or simultaneously irradiating the bottom or top electrode of the fully optically controlled memristor array with enhanced and suppressed light, feature values ​​of the input signal can be extracted.

7. The all-optical-controlled reservoir computing system according to claim 6, characterized in that, The bottom electrode layer and the top electrode layer are both made of Pt, and the intermediate dielectric layer is made of ZnO.

8. A method for recognizing words using the all-optical-controlled reservoir computing system according to any one of claims 1-7, characterized in that, include: Convert the word into a binary image, i.e., the input signal; The encoding module encodes the "1" bits in the input signal into enhanced light pulses and the "0" bits into suppressed light pulses. The encoded light pulses are then irradiated onto the fully photocontrolled synaptic array to complete bipolar encoding and obtain multiple photocurrent values. The multiple photocurrent values ​​are transmitted to the output layer, and the output layer neural network, trained by gradient descent, yields the word recognition result.

9. A method for predicting time-series signals using the all-optically controlled reservoir computing system according to any one of claims 1-7, characterized in that, include: The time-series signal is transformed into a high-dimensional sequence based on a binary mask, thus obtaining the input signal; The negative values ​​in the input signal are mapped to the suppression light, and the optical power density of the suppression light is mapped to the magnitude of the negative values ​​in the input signal. The positive values ​​are mapped to the enhancement light, and the optical power density of the enhancement light is mapped to the magnitude of the positive values. This converts the timing signal into an encoded optical pulse. The encoded optical pulse illuminates the fully optically controlled synaptic array, thereby completing bipolar encoding and generating multiple photocurrent values. The multiple photocurrent values ​​are transmitted to the output layer, and the output layer neural network trained by the gradient descent method is used to obtain the prediction results of the time series signal.

10. A method for recognizing facial and fingerprint signals using the all-optical-controlled reservoir computing system according to any one of claims 1-7, characterized in that, include: The facial and fingerprint information is converted into binary images to obtain the first source input signal and the second source input signal. The first source input signal is encoded into an enhancement light pulse by an encoding module, where "1" bits in the first source input signal are mapped to enhancement light and "0" bits are mapped to unirradiated enhancement light. The second source input signal is encoded into a suppression light pulse, where "1" bits in the second source input signal are mapped to suppression light and "0" bits are mapped to unirradiated suppression light. The enhancement light pulse and the suppression light pulse are simultaneously irradiated onto the fully optically controlled synaptic array to generate multiple photocurrent values, thereby completing the feature extraction of the input signal through parallel encoding. The multiple photocurrent values ​​are transmitted to the output layer, and the output layer neural network trained by the gradient descent method is used to complete the identification of multi-source signals.

Images