High-precision analog light calculation method and system based on bit slicing

Abstract

The invention discloses a high-precision analog optical calculation method and system based on bit slicing, and the method comprises the steps: decomposing a high-precision signal into low-precision signal segments, decomposing a high-precision weight into a plurality of low-precision weight segments, and carrying out the multiplication of the low-precision signal and the weight in an optical domain, and finally, combining the low-precision calculation results to recover high-precision output. According to the method, high-precision data are split into a plurality of low-precision bit slices for processing, and a plurality of input bit slices and weight bit slices are processed at the same time by utilizing the high-speed parallel characteristic of photon calculation, so that the calculation efficiency of a large-scale calculation task is improved, and the optical calculation hardware requirement is reduced. Besides, by adjusting the values of M and N, the method can flexibly adapt to calculation tasks of different scales and accuracies, has strong adaptability and flexibility, realizes balance between calculation accuracy and hardware complexity, fully utilizes parallel processing advantages of photons, and provides a solution for large-scale data processing and efficient optical domain calculation.

Description

Technical Field

[0001] The present invention relates to the field of optical computing, and in particular to a high-precision analog optical computing method and system based on bit slicing. Background Art

[0002] Optical computing is a new technology that uses light as an information carrier and realizes computing through the physical properties of light. Its core principle relies on the propagation, interference, diffraction and refraction of light waves. Unlike traditional electronic computers that rely on the flow of electrons to perform logical operations in semiconductors, optical computing completes similar operations through the behavior of photons. The basic units of optical computing include optical devices such as optical crystals, waveguides, modulators and lasers. These devices realize data transmission and logical operations by manipulating the phase, amplitude, polarization and wavelength of light. At the same time, optical computing, with its unique parallelization advantages, can realize simultaneous computing of multi-channel radars through time multiplexing, space multiplexing, wavelength multiplexing, mode multiplexing, etc.

[0003] In recent years, a good research foundation has been formed in the field of optical computing technology at home and abroad, and the high-throughput and low-latency advantages of optical processing methods have been preliminarily verified. Among them, the University of Wisconsin-Madison used an optical frequency comb as a light source and adopted the principle of optical dispersion to achieve ultra-high-speed optical convolution calculations, with a clock frequency of 62.9GHz (X. Xu, M. Tan, B. Corcoran, et al. 11TOPS photonic convolutional accelerator for optical neural networks. Nature589, 44–51 (2021)). Cornell University used spatial light modulators to build large-scale neural networks, proving that less than one photon of energy is required to complete a multiplication (T. Wang, SY. Ma, LG Wright, et al. An optical neural network using less than 1 photon per multiplication. Nat Commun 13,123 (2022)).

[0004] However, although optical computing has achieved preliminary verification in individual performance, further improvement of its practical application performance still faces challenges. Among them, factors such as optical link insertion loss, environmental interference, and active noise are the main reasons for the decline in optical computing accuracy. These problems not only affect the stability and reliability of optical computing, but also limit its application in more fields.

[0005] At present, the research on improving the accuracy of optical computing mainly focuses on the control of weight precision. In order to solve the problem of high sensitivity and limited weight precision of microring resonators, the research team of Princeton University proposed to use feedforward and feedback methods to achieve microring resonance locking and continuous weight configuration control, and achieved 9-bit microring control precision. They successfully developed a high-precision control scheme for microring resonators that can be extended to large optical computing networks (W. Zhang, A. Tait, C. Huang, et al. Broadband physical layer cognitive radio with an integrated photonicprocessor for blind source separation. Nat Commun 14, 1107 (2023)). The research team of Peking University also conquered the calibration of multi-wavelength parallel optical computing systems and ultra-high precision weight loading. The calibration method developed for multi-wavelength computing systems can realize accurate control of each part of the optoelectronic devices and achieve 9-bit weight control precision (B. Bai, Q. Yang, H. Shu et al. Microcomb-based integrated photonicprocessing unit. Nat Commun 14, 66 (2023)).

[0006] However, although these weight accuracy improvement methods have improved the accuracy of optical calculations to a certain extent, they have not fundamentally solved the problem of improving the final calculation accuracy. In particular, the optimization effect on the final calculation results of simulated optical calculations is not obvious. When processing complex tasks, the accuracy and stability of simulated optical calculations are often restricted by various factors, such as nonlinear effects, optical loss, and error accumulation. Therefore, a comprehensive solution for improving the final calculation accuracy is needed to achieve high-precision simulated optical calculations. Summary of the invention

[0007] The purpose of the present invention is to address the deficiencies of the above-mentioned prior art and propose a high-precision analog optical computing method and system based on bit slicing. The scheme adopts a multiple calculation method. First, the input signal and weight signal are bit sliced ​​according to their accuracy requirements to generate multiple low-bit precision signal combinations; then, these low-bit precision signals and weights are independently calculated using analog optical computing technology; finally, multiple low-precision calculation results are accumulated and combined according to weights to generate high-precision matrix multiplication or convolution calculation results.

[0008] The technical solution of the present invention is as follows: In one aspect, the present invention provides a high-precision analog light calculation method based on bit slicing, which is characterized in that it includes the following steps: S1. Decomposing the input signal and the weight signal by bit slicing according to the predetermined precision requirements, and generating a plurality of signal combinations with low bit precision, wherein each signal combination with low bit precision contains a part of the bits of the input signal or the weight signal; S2. Using analog optical computing technology, each low-bit precision signal combination generated in step S1 and the corresponding low-bit precision weight signal are independently calculated to obtain multiple low-precision calculation results; S3. Accumulate and combine the multiple low-precision calculation results obtained in step S2 according to predetermined weights to generate a high-precision matrix multiplication or convolution calculation result.

[0009] Furthermore, the bit slice decomposition in step S1 is to decompose the input signal and the weight signal into a plurality of binary bit segments respectively, and the number of bits of each binary bit segment is determined according to the required calculation accuracy.

[0010] Further, the high-precision input data stream with a word length of K bits is split into M low-precision input bit slices with a word length of K / M bits through an input bit slice reconstruction module, wherein K is a positive integer multiple of M, and M is a positive integer ≥ 2; The high-precision weight with a word length of L bits is split into N low-precision weight bit slices with a word length of L / N bits through the weight bit slice reconstruction module, and M copies of the output are copied, where L is a positive integer multiple of N and N is a positive integer ≥2.

[0011] Furthermore, the input bit slice reconstruction module includes an electrical input port, M optical output ports and M electro-optical converters, and the electro-optical converter converts the electrical signals of the M low-precision input bit slices into optical signals for output.

[0012] Furthermore, the weight bit slice reconstruction module includes an electrical input port and M×N electrical output ports, and the M×N electrical output ports output M copies of low-precision weight bit slices.

[0013] Furthermore, it also includes: The photonic computing core module has M optical input ports, M×N electrical input ports, M×N electrical output ports, M×N weight controllers and M×N photodetectors, which are used to perform dot product calculations on M low-precision input bit slices and M×N low-precision weight bit slices, and convert the calculation results into electrical signal outputs through photodetectors.

[0014] The output bit slice combination module has M×N electrical input ports and one electrical output port, and is used to multiply the M×N low-precision calculation results by corresponding weights and then sum them to obtain a high-precision calculation result, and output it through an electrical output port.

[0015] In a second aspect, the present invention further provides a high-precision analog optical computing system based on bit slicing, which is characterized by comprising: An input bit slice reconstruction module has an electrical input port, M optical output ports, and M electro-optical converters, where M is a positive integer ≥ 2; a high-precision input data stream with a word length of K bits is input through the electrical input port, and is split into M low-precision input bit slices with a word length of K / M bits by a reconstruction method, and the M electrical signals are converted into optical signals by the electro-optical converter and then output to the photonic computing core module, where K is a positive integer multiple of M; The weight bit slice reconstruction module has an electrical input port and M×N electrical output ports, where N is a positive integer ≥ 2; a high-precision weight with a word length of L bits is input through the electrical input port, split into N low-precision weight bit slices with a word length of L / N bits by reconstruction, and then copied M times and output to the photon computing core module, where L is a positive integer multiple of N; A photon computing core module, comprising M optical input ports connected to the M optical output ports of the input bit slice reconstruction module, M×N electrical input ports connected to the M×N electrical output ports of the weight bit slice reconstruction module, M×N electrical output ports, M×N weight controllers and M×N photodetectors; the M×N weight bit slices are loaded onto the high-parallel photon computing core via the M×N weight controllers, and the M×N output bit slice results are obtained by calculating with the M input bit slices, and are converted into electrical signals via the M×N photodetectors respectively, and transmitted to the output bit slice combination module as output bit slices; The output bit slice combination module has M×N electrical input ports connected to the M×N electrical output ports of the photon computing core module and one electrical output port; each input bit slice and each weight bit slice will calculate an output bit slice result, numbered (m, n), that is, the result of the calculation of the mth input bit slice and the nth weight bit slice, where 1≤m≤M, 1≤n≤N. The output bit slice numbered (m, n) is multiplied by the corresponding weight coefficient 2 (M -m)K / M+(N-n)L / N , then the M×N output bit slices are summed to obtain a high-precision calculation result, which is output through an electrical output port.

[0016] Furthermore, the electro-optical converter in the input bit slice reconstruction module is a coherent modulator, which works in a zero bias point mode and linearly modulates the input electrical signal to an optical amplitude quantity, and the value of the signal is represented by the amplitude value of the light.

[0017] Furthermore, the reconstruction methods in the input bit slice reconstruction module and the weight bit slice reconstruction module are implemented by a single chip microcomputer or a digital computer.

[0018] Furthermore, the photonic computing core module adopts an on-chip coherent optical computing architecture, an on-chip incoherent optical computing architecture, or a spatial diffraction optical computing architecture; or adopts different time, space, mode, and wavelength methods to realize operations on different signals and weights.

[0019] Furthermore, the photonic computing core module may use, including but not limited to, on-chip coherent optical computing architecture, on-chip incoherent optical computing architecture, and spatial diffraction optical computing architecture; it may also use, but not limited to, the same optical computing architecture at different time points, or different optical computing architectures in space.

[0020] Furthermore, the weight controller in the photonic computing core module may use photonic devices including but not limited to micro-rings, phase shifters, phase change materials, spatial light modulators, etc. that have a regulating effect on the amplitude of the light field.

[0021] Compared with the prior art, the technical advantages of the present invention are: (1) Compared with the traditional optical computing architecture, the present invention adopts a bit-sliced ​​optical computing method, which enables low-precision optical computing devices to perform high-precision optical computing tasks, greatly expanding the application scope of optical computing, providing new solutions for processing in complex scenarios, and breaking the limitations faced by traditional optical computing architectures in processing high-precision data.

[0022] (2) The present invention can make full use of existing low-precision computing system resources to perform high-precision computing, which reduces the requirements for optical computing system integration and packaging, making the present invention easier to integrate and upgrade with existing systems, thus facilitating the popularization and application of optical computing technology.

[0023] (3) The present invention realizes high-precision simulated optical computing through reconstruction of input and weights, efficient dot product calculation in the photon computing core, and shift-and-add combination of output results. It not only improves the accuracy and efficiency of the calculation, but also makes full use of the parallel processing advantages of photons, providing a promising solution for large-scale data processing.

[0024] (4) The present invention processes high-precision data by splitting it into multiple low-precision bit slices, and uses the high-speed parallel characteristics of photon computing to simultaneously process multiple input bit slices and weight bit slices, thereby significantly improving the computational efficiency of large-scale computing tasks. In addition, by adjusting the values ​​of M and N, it can flexibly adapt to computing tasks of different scales and precisions, and has strong adaptability and flexibility. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 FIG. 4 is a schematic diagram of an embodiment of a high-precision analog optical computing system based on bit slicing of the present invention.

[0026] Figure 2 It is a schematic diagram of the data flow of high-precision simulated light calculation of the present invention.

[0027] Figure 3 It is a schematic diagram of the principle of high-precision dot product calculation of the present invention.

[0028] Figure 4 It is a schematic diagram of the principle of high-precision input and weight reconstruction into bit slices of the present invention. DETAILED DESCRIPTION

[0029] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and embodiments, and detailed implementation methods and structures are given, but the protection scope of the present invention is not limited to the following embodiments.

[0030] See also Figure 1 , Figure 1 This is a schematic diagram of an embodiment of a high-precision analog optical computing system based on bit slicing of the present invention. As shown in the figure, the high-precision analog optical computing system based on bit slicing of this embodiment includes an input bit slice reconstruction module 100, a weight bit slice reconstruction module 200, a photon computing core module 300, and an output bit slice combination module 400.

[0031] The input bit slice reconstruction module 100 has one electrical input port, M optical output ports and M electro-optical converters. In this embodiment, M=3.

[0032] The three optical output ports of this module are connected to the optical input ports of the photon computing core module 300. The high-precision input data stream with a word length of K = 6 bits enters through the electrical input port and is split into three low-precision input bit slices with a word length of 2 bits using a reconstruction algorithm. These low-precision input bit slices are numbered 1, 2, and 3 in order from high to low. Then, the three electrical signals are converted into optical signals through an electro-optical converter and output to the photon computing core module 300.

[0033] The weight bit slice reconstruction module 200 has one electrical input port and M×N electrical output ports. In this embodiment, N=3.

[0034] The nine electrical output ports of this module are connected to the electrical input ports of the photon computing core module (300).

[0035] The high-precision weight with a word length of L=6 bits is input through the electrical input port, and is split into three low-precision weight bit slices with a word length of 2 bits using a reconstruction algorithm, from high to low bits as 1, 2, and 3. Then, these low-precision weight bit slices are copied three times and output to the photon computing core module 300.

[0036] The photon computing core module 300 has M optical input ports, M×N electrical input ports, M×N electrical output ports, M×N weight controllers and M×N photodetectors; that is, this embodiment has 3 optical input ports, 9 electrical input ports, 9 electrical output ports, 9 weight controllers and 9 photodetectors. The 9 electrical input ports of this module are connected to the electrical output ports of the weight bit slice reconstruction module 200, the 3 optical input ports of this module are connected to the optical output ports of the input bit slice reconstruction module 100, and the 9 electrical output ports of this module are connected to the electrical input ports of the output bit slice combination module 400.

[0037] The three input bit slices are input through the optical input port, the nine weight bit slices are input through the electrical input port, and are loaded onto the high-parallel photon computing core through the nine weight controllers. After the three input bit slices pass through the nine weight controllers, nine output bit slice results are obtained. Each input bit slice and each weight bit slice will calculate an output bit slice result, numbered (m, n), that is, the result of the calculation of the mth input bit slice and the nth weight bit slice, where 1≤m≤3, 1≤n≤3. Finally, it is converted into an electrical signal through a photodetector and output to the output bit slice combination module 400.

[0038] The output bit slice combination module 400 has 9 electrical input ports and one electrical output port. The 9 electrical input ports of this module are connected to the electrical output port of the photon computing core module (300). The output bit slice numbered (m, n) is multiplied by the corresponding weight coefficient and multiplied by 2. (M-m)K / M+(N-n)L / N , and then sum to get a high-precision calculation result. In this embodiment, the (m, n)th output bit slice is multiplied by 2 2(6-m-n) , the 9 output bit slices are weighted and summed to obtain a high-precision calculation result, which is output through an electrical output port, thus completing the final high-precision calculation result.

[0039] See also Figure 2 , Figure 2 It is a schematic diagram of the data flow of high-precision light simulation calculation of the present invention.

[0040] The input high-precision image data IN and the convolution kernel WEIGHT are reconstructed into high and low bits, namely IN1, IN2, WEIGHT1, WEIGHT2, respectively. Through the high-parallel photon computing core, the input bit slices and the convolution kernel bit slices are combined in pairs to obtain four output bit slices OUT. 1,1 ,OUT 1,2 ,OUT 2,1 ,OUT 2,2 Finally, the high-precision output result OUT is obtained by shift combination High , the result is significantly better than the low-precision calculation result OUT LOW .

[0041] See also Figure 3 , Figure 4 , Figure 3 Schematic diagram of the principle of high-precision dot product calculation of the present invention, Figure 4 The principle of high-precision input,weight reconstruction into bit slices is illustrated.

[0042] The 6-bit input and weight data can be divided into three 2-bit input and weight bit slices according to the number of bits, that is, three groups of 2-bit data streams. By taking the dot product of the two bit slices, we can get a 4-bit output bit slice. Then, by shifting, that is, multiplying by the coefficients during reconstruction, and finally adding them together, we can get a 12-bit high-precision output result. It can be seen that the calculation result is the same as the direct calculation with 6-bit precision.

[0043] The photon computing core module in the present invention utilizes the parallelism and high-speed transmission characteristics of photons to perform dot product calculations, that is, each weight bit slice and input bit slice are encoded as optical signals, and then interact through optical elements in the photon computing core to realize dot product operations. Due to the parallel processing capability of light, the photon computing core can process multiple data slices at the same time, thereby significantly improving the calculation speed. After calculating the dot product results of all low-precision bit slices, these results are combined through shift and addition operations to restore the original high-precision calculation results, which are completed through the output bit slice combination module to obtain the final high-precision output result. The present invention realizes high-precision analog optical computing through the reconstruction of input and weights, efficient dot product calculations in the photon computing core, and shift and add combination of output results, which not only improves the accuracy and efficiency of calculations, but also makes full use of the parallel processing advantages of photons, providing a promising solution for large-scale data processing.

Claims

1. A high-precision analog optical computing method based on bit slicing, characterized in that: The following steps are involved: S1. Decomposing the input signal and the weight signal by bit slicing according to the predetermined precision requirements, and generating a plurality of signal combinations with low bit precision, wherein each signal combination with low bit precision contains a part of the bits of the input signal or the weight signal; S2. Using analog optical computing technology, each low-bit precision signal combination generated in step S1 and the corresponding low-bit precision weight signal are independently calculated to obtain multiple low-precision calculation results; S3. Accumulate and combine the multiple low-precision calculation results obtained in step S2 according to predetermined weights to generate a high-precision matrix multiplication or convolution calculation result.

2. The high-precision analog optical computing method based on bit slicing according to claim 1, characterized in that: The bit slice decomposition in step S1 is to decompose the input signal and the weight signal into a plurality of binary bit segments respectively, and the number of bits of each binary bit segment is determined according to the required calculation accuracy.

3. The high-precision analog optical computing method based on bit slicing according to claim 2, characterized in that: The high-precision input data stream with a word length of K bits is split into M low-precision input bit slices with a word length of K / M bits through an input bit slice reconstruction module, where K is a positive integer multiple of M and M is a positive integer ≥ 2; The high-precision weight with a word length of L bits is split into N low-precision weight bit slices with a word length of L / N bits through the weight bit slice reconstruction module, and M copies of the output are copied, where L is a positive integer multiple of N and N is a positive integer ≥2.

4. The high-precision analog optical computing method based on bit slicing according to claim 3, characterized in that: The input bit slice reconstruction module includes an electrical input port, M optical output ports and M electro-optical converters. The electro-optical converter converts the electrical signals of the M low-precision input bit slices into optical signals and then outputs them.

5. The high-precision analog optical computing method based on bit slicing according to claim 3, characterized in that: The weight bit slice reconstruction module includes an electrical input port and M×N electrical output ports, and the M×N electrical output ports output M copies of low-precision weight bit slices.

6. The high-precision analog optical computing method based on bit slicing according to any one of claims 1 to 5, characterized in that: Also includes: A photonic computing core module, which has M optical input ports, M×N electrical input ports, M×N electrical output ports, M×N weight controllers and M×N photodetectors, and is used to perform dot product calculations on M low-precision input bit slices and M×N low-precision weight bit slices, and convert the calculation results into electrical signal outputs through the photodetectors; The output bit slice combination module has M×N electrical input ports and one electrical output port, and is used to multiply the M×N low-precision calculation results by corresponding weights and then sum them to obtain a high-precision calculation result, and output it through an electrical output port.

7. A high-precision analog optical computing system based on bit slicing, characterized in that: include: An input bit slice reconstruction module having an electrical input port, M optical output ports, and M electro-optical converters, where M is a positive integer ≥ 2; The high-precision input data stream with a word length of K bits is input through the electrical input port, and is split into M low-precision input bit slices with a word length of K / M bits by reconstruction. The M electrical signals are converted into optical signals by the electro-optical converter and then output to the photonic computing core module, where K is a positive integer multiple of M. The weight bit slice reconstruction module has an electrical input port and M×N electrical output ports, where N is a positive integer ≥ 2; a high-precision weight with a word length of L bits is input through the electrical input port, split into N low-precision weight bit slices with a word length of L / N bits by reconstruction, and then copied M times and output to the photon computing core module, where L is a positive integer multiple of N; A photon computing core module, comprising M optical input ports connected to the M optical output ports of the input bit slice reconstruction module, M×N electrical input ports connected to the M×N electrical output ports of the weight bit slice reconstruction module, M×N electrical output ports, M×N weight controllers and M×N photodetectors; the M×N weight bit slices are loaded onto the high-parallel photon computing core via the M×N weight controllers, and the M×N output bit slice results are obtained by calculating with the M input bit slices, and are converted into electrical signals via the M×N photodetectors respectively, and transmitted to the output bit slice combination module as output bit slices; The output bit slice combination module has M×N electrical input ports connected to the M×N electrical output ports of the photon computing core module and one electrical output port; each input bit slice and each weight bit slice calculate an output bit slice result, which is numbered (m, n), that is, the result of the calculation of the mth input bit slice and the nth weight bit slice, where 1≤m≤M, 1≤n≤N, and the output bit slice numbered (m, n) is multiplied by the corresponding weight coefficient (M-m)K / M+(N-n)L / N , then the M×N output bit slices are summed to obtain a high-precision calculation result, which is output through an electrical output port.

8. The high-precision analog optical computing system based on bit slicing according to claim 7, characterized in that: The electro-optical converter in the input bit slice reconstruction module is a coherent modulator, which works in a zero bias point mode and linearly modulates the input electrical signal to an optical amplitude quantity, and the value of the signal is represented by the amplitude value of the light.

9. The high-precision analog optical computing system based on bit slicing according to claim 7, characterized in that: The reconstruction methods in the input bit slice reconstruction module and the weight bit slice reconstruction module are implemented by a single chip microcomputer or a digital computer.

10. The high-precision analog optical computing system based on bit slicing according to claim 7, characterized in that: The photonic computing core module adopts an on-chip coherent optical computing architecture, an on-chip incoherent optical computing architecture, or a spatial diffraction optical computing architecture; or adopts different time, space, mode, and wavelength methods to realize operations on different signals and weights.

Images