Acceleration system and method for random Fourier feature learning algorithm

Abstract

The invention provides an acceleration system and method of a random Fourier feature learning algorithm, relates to the technical field of hardware acceleration of the random Fourier feature learning algorithm, and realizes hardware native calculation of a random Fourier feature kernel function through light wave interference. Mathematical mapping and complex operation are converted into a continuous physical process of optical phase modulation and interference, and the analog-to-digital conversion and digital operation overhead of a traditional electronic architecture is reduced. The dynamic reconfigurable light modulation unit is combined with a wavelength multiplexing technology, supports real-time adjustment and multi-probability distribution switching of RFF kernel function parameters, and flexibly adapts to diversified data characteristics and task requirements. The system-level photoelectric hybrid interconnection and multi-modal scheduling mechanism realizes hierarchical expansion from vector inner product to matrix operation, an attention mechanism and a Transform model, and considers edge end low-power-consumption reasoning and a data center-level training scene. The anti-interference characteristic of optical calculation and the hardware-algorithm closed-loop calibration design guarantee the calculation precision and reliability in a complex environment.

Description

Technical Field

[0001] The present invention relates to the technical field of hardware acceleration of random Fourier feature learning algorithms, and in particular to an acceleration system and method for random Fourier feature learning algorithms. Background Art

[0002] As an efficient approximation technology of kernel methods, the random Fourier feature (RFF) learning algorithm is widely used in machine learning, signal processing, and large model reasoning. However, existing RFF acceleration solutions mainly rely on GPU or ASIC implementation, which has significant bottlenecks: First, Feng Data movement under the Neumann architecture leads to a "memory wall" problem, limiting computing energy efficiency. Secondly, traditional electronic solutions require pseudo-random number generators (PRNGs) and digital multipliers to achieve random mapping, which dramatically increases hardware overhead and power consumption. Thirdly, dynamic task adaptability is poor, and fixed hardware is difficult to support multimodal computing requirements such as vector inner products, matrix operations, and attention mechanisms. For example, when deploying Transformer models at the edge, existing solutions are difficult to meet real-time requirements due to high power consumption and high latency. In data center scenarios, frequent optoelectronic signal conversion (such as silicon photonic interconnect systems) introduces additional latency and energy loss, limiting system scalability. In addition, due to the lack of algorithm-hardware co-design, traditional optical computing solutions have fixed RFF kernel function parameters and cannot adapt to non-stationary data distributions (such as time series signal analysis), which seriously restricts the progress of practical application.

[0003] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention

[0004] The purpose of the present invention is to provide an acceleration system and method for a random Fourier feature learning algorithm to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] An acceleration system for a random Fourier feature learning algorithm, specifically comprising: several groups of computing chips and connecting devices;

[0007] The computing chip includes:

[0008] A phase-modulated optical input module, wherein the phase-modulated optical input unit includes two groups of input ports and an optical modulation unit, and is used to respectively receive two groups of different input signals and generate corresponding phase-modulated optical signals through optical phase modulation operation;

[0009] A reconfigurable interference module, comprising an optical interference unit and an output port array, configured to perform interference modulation on a phase-modulated optical signal, generate an interference signal proportional to an RFF kernel function, and output the interference signal;

[0010] A wavelength multiplexing module, wherein the wavelength multiplexing module is used to realize parallel transmission of at least four wavelengths;

[0011] A photoelectric conversion module, configured to convert the interference signal into an electrical signal and transmit the signal;

[0012] The connecting device comprises:

[0013] An optoelectronic interconnection module is provided between computing chips and is used to perform optoelectronic signal conversion between different computing chips;

[0014] Dynamic task scheduling module, used to automatically select different working modes according to the matrix operation dimension.

[0015] Preferably, the phase-modulated optical signal and the interference signal have complex amplitudes.

[0016] And the phase of the phase-modulated optical signal after the optical phase-modulation operation is expressed as:

[0017]

[0018] In the formula Indicates the The phase of the phase-modulated optical signal, Indicates the The input data corresponding to the phase-modulated optical signal, Indicates the index of the phase-modulated optical signal, Indicates the Group satisfies probability distribution The random parameters of represents transpose;

[0019] The light intensity of the interference signal is expressed as:

[0020]

[0021]

[0022] In the formula Indicates the The light intensity of the group interference signal, represents the index of the interference signal, Represents an imaginary unit.

[0023] Preferably, the driving voltage of the light modulation unit and the random parameter satisfy the mapping relationship:

[0024]

[0025] In the formula represents the operating wavelength, represents the change in refractive index caused by the driving voltage, Indicates the length of the modulation area.

[0026] Preferably, the phase error compensation accuracy of the phase-modulated optical signal is less than 0.05π, and the phase offset caused by ambient temperature fluctuation is corrected in real time through a closed-loop feedback circuit.

[0027] Preferably, the working logic of the dynamic task scheduling module is:

[0028] When a single computing chip performs time multiplexing, it performs vector inner product calculations;

[0029] When multiple computing chips form a one-dimensional array, matrix-vector product calculations are performed;

[0030] When multiple computing chips form a multi-dimensional array, matrix-matrix product calculations are performed;

[0031] When multiple computing chips form a multi-dimensional array and perform time multiplexing, attention mechanism calculations are performed;

[0032] When multiple sets of time-multiplexed multi-dimensional arrays are cascaded with each other, high-dimensional transformer model calculations are performed.

[0033] Preferably, the logic of the attention mechanism calculation performed by the dynamic task scheduling module is:

[0034] The query vector With key vector They are loaded into the phase-modulated light input array to generate corresponding mapping vectors, and the approximate attention score matrix is ​​generated by light interference. , the approximate attention score matrix is ​​expressed as:

[0035]

[0036] In the formula 、 The mapping vectors representing query vector and key vector respectively, represents transpose;

[0037] After the optical signal is converted into an electrical signal, it is input into the nonlinear activation circuit to perform a soft maximization operation to generate the corresponding attention weight and make the dynamic range meet the preset linear interval. The calculation method is expressed as:

[0038]

[0039] In the formula Represents the approximate attention score matrix The attention weight, represents the soft maximization operation, represents the temperature parameter, Represents feature dimension;

[0040] The attention weights are normalized and mapped into optical signals, driving the next layer of computing chips to perform vector weighted summation.

[0041] Preferably, the mapping vectors of the query vector and the key vector are respectively expressed as:

[0042]

[0043]

[0044] In the formula Indicates optical phase modulation operation, 、 They represent the weight matrix, Indicates RFF calculation;

[0045] The weight matrix is ​​optimized by gradient descent method.

[0046] Preferably, the phase error of the optical phase modulation operation is calibrated in real time through a reference interferometer loop, and the corresponding mapping error satisfies:

[0047]

[0048] In the formula represents the L2 norm in the brackets, represents the imaginary unit, represents the error threshold, .

[0049] A method for accelerating a random Fourier feature learning algorithm, which is applicable to the above-mentioned acceleration system, comprises the following steps:

[0050] S1: Converts the input electrical signal into a phase-modulated optical signal and modulates its phase according to a preset random parameter distribution to achieve data-to-optical phase mapping.

[0051] S2: Perform interference superposition on the phase-modulated signals to generate a light intensity distribution proportional to the RFF kernel function, completing the physical calculation of the inner product of the high-dimensional vector;

[0052] S3: Converts the interference signal into an electrical signal and dynamically switches the working mode according to the task dimension;

[0053] S4: Broadcast the electrical signal converted from the interference signal in the computing chip array and perform multi-dimensional operations in blocks;

[0054] S5: Reorganize the output signal optoelectronic interconnection topology to implement the calculation of the attention mechanism or Transformer model.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] This invention uses optical wave interference to achieve hardware-native computation of the random Fourier feature (RFF) kernel function, transforming mathematical mapping and complex number operations into a continuous physical process of optical phase modulation and interference. This significantly reduces the analog-to-digital conversion and digital computing overhead of traditional electronic architectures, breaking through energy efficiency and computing power bottlenecks. The dynamically reconfigurable optical modulation unit, combined with wavelength multiplexing technology, supports real-time adjustment of RFF kernel function parameters and switching between multiple probability distributions, flexibly adapting to diverse data characteristics and task requirements, and resolving the issue of hardware functional rigidity. System-level optoelectronic hybrid interconnection and multimodal scheduling mechanisms enable hierarchical expansion from vector inner products to matrix operations, attention mechanisms, and Transformer models, taking into account both low-power inference at the edge and large-model training scenarios at the data center level. The anti-interference properties of optical computing and the hardware-algorithm closed-loop calibration design further ensure computational accuracy and reliability in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a schematic diagram of the overall module structure of the present invention;

[0058] Figure 2 This is a schematic diagram of the computing chip module structure of the present invention;

[0059] Figure 3 Schematic diagram of the overall method of the present invention. DETAILED DESCRIPTION

[0060] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to specific embodiments.

[0061] It should be noted that, unless otherwise defined, the technical or scientific terms used in the present invention should have the usual meanings understood by people with ordinary skills in the field to which the present invention belongs. The "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative position relationships. When the absolute position of the object being described changes, the relative position relationship may also change accordingly.

[0062] Example:

[0063] See also Figures 1 to 3 , the present invention provides a technical solution:

[0064] An acceleration system for a random Fourier feature learning algorithm, specifically comprising: several groups of computing chips and connecting devices;

[0065] The computing chip includes: phase modulation optical input module, reconfigurable interference module, wavelength multiplexing module, photoelectric conversion module,

[0066] The phase-modulated optical input unit comprises two groups of input ports and an optical modulation unit, and is used to respectively receive two groups of different input signals and generate corresponding phase-modulated optical signals through optical phase modulation operation.

[0067] The phase-modulated optical signal and the interference signal have complex amplitudes, and the phase of the phase-modulated optical signal after the optical phase-modulation operation is expressed as:

[0068]

[0069] In the formula Indicates the The phase of the phase-modulated optical signal, Indicates the The input data corresponding to the phase-modulated optical signal, Indicates the index of the phase-modulated optical signal, Indicates the Group satisfies probability distribution A random parameter of the distribution (e.g., uniform or Gaussian).

[0070] The light intensities of the two sets of interference signals are expressed as:

[0071]

[0072]

[0073] In the formula Indicates the The light intensity of the group interference signal, represents the index of the interference signal, Specifically, according to RFF theory, the translation-invariant kernel k can be represented by the inner product of the random mapping h(.) of two variables x and y:

[0074]

[0075] In this embodiment, the input data corresponding to the phase modulated optical signal 、 It is equivalent to the variables x and y in the formula. Correspondingly, the light intensity difference between the two sets of interference signals can be expressed as:

[0076]

[0077] Right now:

[0078]

[0079] From the expression of the light intensity difference between the two sets of interference signals, it can be seen that after the time domain accumulation, the calculation result directly corresponds to the imaginary part of the inner product in the RFF calculation, and the optical signal is applied by the phase modulator. After the offset, the real part of the RFF inner product can still be obtained, so it can be approximately regarded as an RFF calculation. In other words, through optical interference, RFF calculation is implemented at the hardware level, thus using RFF calculation to replace the matrix multiplication in the traditional attention mechanism.

[0080] It's understandable that in traditional RFF calculations, the complexity is proportional to the product of the square of the sequence length and the feature dimension. However, when using RFF calculations instead of the matrix multiplication in the traditional attention mechanism, the complexity is proportional to the product of the sequence length, the feature dimension, and the number of signal samples (i.e., the total number of phase-modulated optical signals). Because the number of signal samples is much smaller than the sequence length when performing more difficult calculations, the complexity of the latter is also much smaller than that of the former. In other words, using RFF calculations instead of the matrix multiplication in the traditional attention mechanism can significantly reduce computational complexity, improve computational response time, and achieve hardware-level acceleration.

[0081] Since the RFF calculation method is an existing technology, its specific calculation process is not described in detail here.

[0082] By directly implementing random mapping of the RFF kernel function through optical signal phase modulation, the mathematical complex multiplication and addition operations are converted into the physical process of light wave interference. Optical physical calculations replace digital operations, eliminating the analog-to-digital conversion and discrete operation steps in traditional electronic computing, thereby improving single-point performance indicators and helping to break through the bottleneck of the von Neumann architecture.

[0083] The driving voltage of the optical modulation unit and the random parameters satisfy the mapping relationship:

[0084]

[0085] In the formula represents the operating wavelength, represents the change in refractive index caused by the driving voltage, Indicates the length of the modulation area.

[0086] Specifically, the driving voltage here is generated by a digital-to-analog converter (DAC) according to a probability distribution, the operating wavelength is selected according to the laser configuration, the modulation zone length is determined by the chip layout design file (typical value 50μm~200μm), and the waveguide refractive index change caused by voltage is queried through the silicon-based waveguide Pockels coefficient table provided by the manufacturer.

[0087] The phase error compensation accuracy of the phase-modulated optical signal is less than 0.05π, and a closed-loop feedback circuit is used to correct phase offsets caused by ambient temperature fluctuations in real time. This closed-loop feedback circuit employs proportional-integral control, with an integral time constant that matches the thermal response time of the thermo-optical modulator.

[0088] By constructing the driving voltage, the frequency characteristics of the RFF kernel function (such as Gaussian kernel bandwidth and uniform distribution range) can be adjusted in real time on demand, avoiding the reliance on fixed hardware parameters or software reconfiguration in traditional methods. It also supports online switching of different probability distributions (such as switching from Gaussian to uniform) to adapt to diverse data characteristics, avoiding the scenario limitations caused by the functional rigidity of traditional ASIC solutions.

[0089] The reconfigurable interference module optical interference unit and output port array are used to modulate and interfere the phase-modulated optical signal, generate an interference signal proportional to the RFF kernel function, and output the interference signal.

[0090] Specifically, the optical modulation unit can adopt any one or more of the following: Mach-Jean-Arc interferometer modulator, Michelson interferometer modulator, microring modulator, electro-absorption modulator; the optical interference unit can adopt one or more of the following multimode interferometers, directional couplers, other components capable of achieving optical-optical interference, or any combination of any of the above devices.

[0091] In this embodiment, two input ports are used for input, and two output ports generate interference output to obtain a real signal corresponding to the difference in optical intensity between the two. Multiple sets of input phase-modulated optical signals are serialized to generate a corresponding output real signal sequence. The intensities of this real signal sequence are then summed to obtain the RFF calculation result. The optical signal intensity of a single output port, or the corresponding real signal generated by an appropriate linear combination of the signal intensities of multiple output ports, can also be used as the RFF or the calculation result of the RFF after some transformation.

[0092] Furthermore, the input phase-modulated optical signal can be multiplexed by multiple computing chips through broadcasting. The weight coefficients of the linear combination of the output port intensities of the computing chips can be multiple, and the corresponding RFF calculation results can also be differentiated. The output real signals of the computing chips can also be combined together to jointly calculate the RFF of a group of input phase-modulated signals, thereby reducing the RFF calculation clock delay.

[0093] The wavelength multiplexing module is used to realize parallel transmission of at least 4 wavelengths.

[0094] Specifically, the wavelength multiplexing module may adopt an arrayed waveguide grating, a multi-stage MZI filter, a microring resonator, a photonic crystal, or any combination thereof.

[0095] The photoelectric conversion module is used to convert the interference signal into an electrical signal and send it.

[0096] The connection device includes: an optoelectronic interconnection module and a dynamic task scheduling module.

[0097] The optoelectronic interconnection module is set between the computing chips and is used to perform optoelectronic signal conversion between different computing chips;

[0098] Used to automatically select different working modes according to the matrix operation dimension.

[0099] The working logic of the dynamic task scheduling module is:

[0100] When multiple wavelengths are parallelized within a single computing chip, vector inner product calculation is performed;

[0101] When multiple computing chips form a one-dimensional array, matrix-vector product calculations are performed;

[0102] When multiple computing chips form a multi-dimensional array, matrix-matrix product calculations are performed;

[0103] When multiple computing chips form a multi-dimensional array and perform time multiplexing, attention mechanism calculations are performed;

[0104] When multiple sets of time-multiplexed multi-dimensional arrays are cascaded with each other, high-dimensional transformer model calculations are performed.

[0105] That is to say, in the time multiplexing mode, the same physical unit processes computing tasks with different random parameters in a time-sharing manner, and the time slice length is dynamically allocated by the clock synchronization module. This allows the optimal mode to be automatically selected according to the task dimension (such as matrix-matrix multiplication enabling spatial multiplexing and small-scale tasks enabling time multiplexing), thereby improving hardware utilization.

[0106] The logic of the attention mechanism calculation in the dynamic task scheduling module is:

[0107] The query vector With key vector They are loaded into the phase-modulated light input array to generate the corresponding mapping vectors. In other words, the corresponding mapping vectors are generated after the optical phase modulation operation, and then the approximate attention score matrix is ​​generated by optical interference. , the approximate attention score matrix is ​​expressed as:

[0108]

[0109] In the formula 、 The mapping vectors representing query vector and key vector respectively, represents transpose;

[0110] After the optical signal is converted into an electrical signal, it is input into the nonlinear activation circuit to perform a soft maximization operation to generate the corresponding attention weight and make the dynamic range meet the preset linear interval. The calculation method is expressed as:

[0111]

[0112] In the formula Represents the approximate attention score matrix The attention weight, represents the soft maximization operation, represents the temperature parameter, Represents feature dimension;

[0113] The attention weights are normalized and mapped into optical signals, which drive the next layer of computing chips to perform vector weighted summation, that is, the weighted summation of the mapping vectors of the attention weights and the value vectors. Specifically, in this embodiment, the query vector and the key vector represent two sets of different phase-modulated optical signals, respectively. After the optical phase-modulation operation, the corresponding mapping vectors can be obtained, and the value vector is obtained by the input data and phase-modulated optical signals corresponding to the two sets of phase-modulated optical signals, including both the electrical and optical domains. Since the generation method of the three vector matrices Q, K, and V of the attention mechanism is a mature existing technology, the specific generation steps will not be repeated here.

[0114] The mapping vectors of query vector and key vector are expressed as:

[0115]

[0116]

[0117] In the formula Indicates optical phase modulation operation, 、 They represent the weight matrix, Indicates RFF calculation;

[0118] The weight matrix is ​​optimized by gradient descent.

[0119] The phase error of the optical phase modulation operation is calibrated in real time through a reference interferometer loop, and the corresponding mapping error satisfies:

[0120]

[0121] In the formula represents the L2 norm in the brackets, represents the imaginary unit, represents the error threshold, .

[0122] In summary, the present invention breaks through the energy efficiency and computing power bottlenecks of traditional electronic architectures through the collaborative innovation of optical physical computing reconstruction and dynamic task scheduling, achieving efficient, flexible, and scalable random Fourier feature (RFF) calculations. Based on the physical properties of light wave interference, the random mapping and complex number operations in the RFF kernel function are converted into a continuous process of optical signal phase modulation and interference, eliminating the redundant analog-to-digital conversion and digital operation links in traditional solutions, significantly reducing power consumption and improving computational parallelism. Through dynamically reconfigurable optical modulation units and wavelength multiplexing technology, it supports real-time adjustment of RFF kernel function parameters and switching of multiple probability distributions, adapting to diverse data characteristics and task requirements, and overcoming the limitations of traditional hardware functional rigidity. Combined with optoelectronic hybrid interconnection and multimodal task scheduling mechanisms, it achieves seamless expansion from vector inner product to matrix operations, attention mechanism, and Transformer model, meeting the scenario requirements of low-power inference at the edge and data center-level large model training. In addition, the anti-interference characteristics of optical computing and the hardware-algorithm closed-loop calibration design ensure computing reliability in complex environments, providing an end-to-end solution for optical intelligent acceleration systems that combines high energy efficiency, low latency, and strong generalization.

[0123] A method for accelerating a random Fourier feature learning algorithm is applicable to the above-mentioned acceleration system, and the specific steps include:

[0124] S1: Converts the input electrical signal into a phase-modulated optical signal and modulates its phase according to a preset random parameter distribution to achieve data-to-optical phase mapping.

[0125] S2: Perform interference superposition on the phase-modulated signals to generate a light intensity distribution proportional to the RFF kernel function, completing the physical calculation of the inner product of the high-dimensional vector;

[0126] S3: Converts the interference signal into an electrical signal and dynamically switches the working mode according to the task dimension;

[0127] S4: Broadcast the electrical signal converted from the interference signal in the computing chip array and perform multi-dimensional operations in blocks;

[0128] S5: Reorganize the output signal optoelectronic interconnection topology to implement the attention mechanism or Transformer model calculation

[0129] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0130] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed by hardware or software depends on the specific application and design constraints of the technical solution.

[0131] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, and may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment as needed.

[0132] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of the present application.

Claims

1. An acceleration system for a random Fourier feature learning algorithm, characterized in that: Specifically include: Several groups of computing chips and connecting devices; The computing chip includes: A phase-modulated optical input module, wherein the phase-modulated optical input unit includes two groups of input ports and an optical modulation unit, and is used to respectively receive two groups of different input signals and generate corresponding phase-modulated optical signals through optical phase modulation operation; A reconfigurable interference module, comprising an optical interference unit and an output port array, configured to perform interference modulation on a phase-modulated optical signal, generate an interference signal proportional to an RFF kernel function, and output the interference signal; A wavelength multiplexing module, wherein the wavelength multiplexing module is used to realize parallel transmission of at least four wavelengths; A photoelectric conversion module, configured to convert the interference signal into an electrical signal and transmit the signal; The connecting device comprises: An optoelectronic interconnection module is provided between computing chips and is used to perform optoelectronic signal conversion between different computing chips; Dynamic task scheduling module, used to automatically select different working modes according to the matrix operation dimension.

2. The acceleration system of the random Fourier feature learning algorithm according to claim 1, characterized in that: The phase-modulated optical signal and the interference signal have complex amplitudes, and the phase of the phase-modulated optical signal after the optical phase-modulation operation is expressed as: In the formula Indicates the The phase of the phase-modulated optical signal, Indicates the The input data corresponding to the phase-modulated optical signal, Indicates the index of the phase-modulated optical signal, Indicates the Group satisfies probability distribution The random parameters of represents transpose; The light intensity of the interference signal is expressed as: In the formula Indicates the The light intensity of the group interference signal, represents the index of the interference signal, Represents an imaginary unit.

3. The acceleration system of the random Fourier feature learning algorithm according to claim 2, characterized in that: The driving voltage of the optical modulation unit and the random parameters satisfy the mapping relationship: In the formula represents the operating wavelength, represents the change in refractive index caused by the driving voltage, Indicates the length of the modulation area.

4. The acceleration system of the random Fourier feature learning algorithm according to claim 1, characterized in that: The phase error compensation accuracy of the phase-modulated optical signal is less than 0.05π, and the phase offset caused by ambient temperature fluctuation is corrected in real time through a closed-loop feedback circuit.

5. The acceleration system of the random Fourier feature learning algorithm according to claim 1, characterized in that: The working logic of the dynamic task scheduling module is: When a single computing chip performs time multiplexing, it performs vector inner product calculations; When multiple computing chips form a one-dimensional array, matrix-vector product calculations are performed; When multiple computing chips form a multi-dimensional array, matrix-matrix product calculations are performed; When multiple computing chips form a multi-dimensional array and perform time multiplexing, attention mechanism calculations are performed; When multiple sets of time-multiplexed multi-dimensional arrays are cascaded with each other, high-dimensional transformer model calculations are performed.

6. The acceleration system of the random Fourier feature learning algorithm according to claim 5, characterized in that: The logic of the attention mechanism calculation in the dynamic task scheduling module is: The query vector With key vector They are loaded into the phase-modulated light input array to generate corresponding mapping vectors, and the approximate attention score matrix is ​​generated by light interference. , the approximate attention score matrix is ​​expressed as: In the formula 、 The mapping vectors representing query vector and key vector respectively; After the optical signal is converted into an electrical signal, it is input into the nonlinear activation circuit to perform a soft maximization operation to generate the corresponding attention weight and make the dynamic range meet the preset linear interval. The calculation method is expressed as: In the formula Represents the approximate attention score matrix The attention weight, represents the soft maximization operation, Represents feature dimension; The attention weights are normalized and mapped into optical signals, driving the next layer of computing chips to perform vector weighted summation.

7. The acceleration system of the random Fourier feature learning algorithm according to claim 6, characterized in that: The mapping vectors of the query vector and key vector are respectively expressed as: In the formula Indicates optical phase modulation operation, 、 They represent the weight matrix, Indicates RFF calculation; The weight matrix is ​​optimized by gradient descent method.

8. The acceleration system of the random Fourier feature learning algorithm according to claim 7, characterized in that: The phase error of the optical phase modulation operation is calibrated in real time through a reference interferometer loop, and the corresponding mapping error satisfies: In the formula represents the L2 norm in the brackets, represents the error threshold, .

9. A method for accelerating a random Fourier feature learning algorithm, characterized by: The acceleration method is applicable to the acceleration system according to any one of claims 1 to 8, and the specific steps include: S1: Converts the input electrical signal into a phase-modulated optical signal and modulates its phase according to a preset random parameter distribution to achieve data-to-optical phase mapping. S2: Perform interference superposition on the phase-modulated signals to generate a light intensity distribution proportional to the RFF kernel function, completing the physical calculation of the inner product of the high-dimensional vector; S3: Converts the interference signal into an electrical signal and dynamically switches the working mode according to the task dimension; S4: Broadcast the electrical signal converted from the interference signal in the computing chip array and perform multi-dimensional operations in blocks; S5: Reorganize the output signal optoelectronic interconnection topology to implement the calculation of the attention mechanism or Transformer model.

Images