Nuclear thread pipeline electron transfer computing architecture

Abstract

The invention discloses a kernel thread pipeline electron transfer computing architecture, which adopts the following five-layer pyramid architecture: a first layer is a maximum real-time dirty data cleaning layer for realizing acquisition, cleaning, extraction, conversion and loading of original time sequence data and giving data integrity; the second layer is a global hash sorting index layer, each field is sorted and indexed independently, an address is established for each data element, sorting and indexing are combined together, and sorting and index creation logic is generated at the same time; the third layer provides the capability of solving an ultra-large and multi-dimensional manifold space in an ultra-fast and ultra-stable manner; the fourth layer is a decision inference layer, automatically makes a decision according to the modeling and analysis result of the third layer, and sends an inference result to the fifth layer to judge whether the inference result is good or bad; and the fifth layer is used for judging whether the result is good or not and completing cyclic feedback of adaptive deep learning. According to the method, the linear speed can be achieved when more than hundreds of millions of records are analyzed and hundreds of variables are modeled.

Description

Technical Field

[0001] The present invention relates to a distributed parallel time-series big data system, and in particular to a core-thread pipeline electron transfer computing architecture. Background Art

[0002] In today's world, time series big data is everywhere, presenting a variety of forms. After the emergence of the quantum era, geometry has entered the Riemannian quantized space that cannot be observed by humans from the Euclidean continuous space. The existing space-time system is to explain the problems in the Euclidean space (points, lines, surfaces, continuity) and the Riemannian space (higher dimensions, quantization), so it is necessary to find new solutions in the field of quantum electrodynamics. Quantum has the "trinity" characteristics, namely: particle nature (position, momentum, mass), wave nature (frequency, phase, wave density) and commutativity (energy exchange, interaction, quantum superposition, quantum entanglement).

[0003] After the advent of the computer age, computing has evolved from the traditional single-core CPU to the parallel multi-core GPU model, and the computing power in the traditional sense has increased significantly. In the past, people used GPUs to transform points into lines, converting single-point calculations into line calculations. From a physical perspective, it is to convert time into frequency and particles into waves. Later, hardware and algorithms were used to convert lines into surfaces, but they still belonged to Euclidean continuous space and were difficult to extend to high-dimensional data space.

[0004] The larger the amount of data, the closer it is to the concept of field theory. Quantum must move in the field. What is currently being studied is the use of quantum superposition and quantum entanglement to solve the E level (10 18 ) is a big data problem. The amount of information accumulated by human civilization up to 2000 is about 10 18 Byte, expressed in units of computer storage capacity, is equivalent to one E-byte. At the same time, the demand for fast analytical algorithms and high-performance computing architectures to process massive amounts of data has grown rapidly. The huge growth has led to the application of artificial intelligence technology in many industries, including banking, healthcare, energy, technology, consumer and manufacturing.

[0005] In the process of digital transformation, "zero-delay" industry means real-time data collection system instead of manual data collection. -9 In order to achieve the "nanosecond (10 -9 "Nanosecond (10 THz)-level" reactions require 6G of terahertz (THz), with a bandwidth from 100 GHz to 10 THz. For example, the spectrum of steel mills, gas turbine temperature, liquid power and aerodynamic parameters, X-ray structure images of rare earth metallurgical lattices, facial features, gene sequences, science and technology finance, etc. all require "nanosecond (10 -9 Seconds) level "reaction.

[0006] As can be seen above, in the era of intelligent interconnection of all things, the data source and data generation method have undergone fundamental changes, and connected devices have replaced human behavior as the new core data production source. The explosive growth of time series data has brought new requirements for the storage and processing of massive time series data, and also created opportunities for changes in the time series data processing and analysis architecture. Summary of the invention

[0007] The technical problem to be solved by the present invention is to provide a core-thread pipeline electron transfer computing architecture, which provides the ability to solve ultra-large, multi-dimensional manifold topological spaces in an ultra-fast and ultra-stable manner, and can process scalar, vector, matrix, and tensor big data existing in topological spaces including Euclidean space, Riemann space, Hilbert space, etc.

[0008] The technical solution adopted by the present invention to solve the above technical problems is to provide a core thread pipeline electron transfer computing architecture, which adopts the following five-layer pyramid architecture: the first layer is a large real-time dirty data cleaning layer, which realizes the collection, cleaning, refining, conversion, loading of original time series data, and gives data integrity; the second layer is a global hash sorting index layer, which sorts and indexes each field separately, establishes an address for each data element, combines sorting and indexing, and generates sorting and indexing creation logic at the same time; the third layer provides the ability to solve ultra-large, multi-dimensional manifold space in an ultra-fast and ultra-stable manner, and helps automatic modeling and intelligent optimization; the fourth layer is a decision inference layer, which automatically makes decisions based on the modeling and analysis results of the third layer, and sends the inference results to the fifth layer for judgment; the fifth layer is used to judge the quality of the results and complete the loop feedback of adaptive deep learning. This layer will detect the quality of the fourth layer's decisions through real-time profit monitoring; if the decision is correct, it returns to the third layer for repeated optimization, and if the decision is wrong, it automatically starts to re-learn, analyze, and construct data from the first layer until the correct decision is made.

[0009] Furthermore, the second layer automatically performs sorting and index pipeline processing on the time series data cleaned by the first layer. The sorting logic occurs within the time required to access each data element once. The previously accessed elements are downstream in the pipeline. After the sorting order is established, the indexing part will use a similar pipeline process to dynamically generate indexes.

[0010] Furthermore, the third layer adopts a standard maximum likelihood estimation method to derive parameters of the conditional probability distribution, and utilizes a path weighted gradient search method to iteratively obtain maximum likelihood estimates of the model parameters.

[0011] Furthermore, the restrictions on the parameter space in the third layer include consideration of the signal-to-noise ratio, special treatment of zero and full data space, correction of high-order expansion deviations of model parameters, invariance of constant translation between local parameters and global parameters, and treatment of local parameter collinearity.

[0012] Furthermore, the judgment process of whether the model in the third layer meets the standard is as follows: (1) Resolution: the degree of distinguishing good from bad rankings is judged by KS, Gini or AUC of cross-time samples. The specific resolution index to be used depends on the application scenario; the three statistical indicators of model resolution are written as: KS = max{F B -F g}; Gini = ∫F B dF G ,AUC=∫F G dF B , where F G and F B are the good and bad cumulative sample curves along the scoring path, respectively. KS is the maximum difference at a certain point. Gini and AUC are both average differences in a certain statistical sense. (2) Accuracy: The statistical error between the data and the model, expressed as the cumulative average chi-square χ on the model variable dimension. 2 (3) Algorithm efficiency: the computing speed of model training and reasoning, judged by the length of time required; (4) Storage efficiency: the storage space required for model training and reasoning, judged by the size of the memory; (5) Energy efficiency: the power consumption required for model operation, judged by the amount of energy consumed per unit time.

[0013] Furthermore, the third layer adopts the following Fermi-Dirac probability distribution function as a probability model;

[0014] remember Let p be the probability corresponding to y, for The corresponding probabilities are:

[0015]

[0016] where f(x|θ,α)=θ μ δ μ (x-α δ )+o(…) (3)

[0017] The superscript θ of the Riemannian geometry vector in equation (3) μ Row vector and subscript δ μ The column vectors are repeated, representing the Einstein summation norm:

[0018]

[0019] Equation (11) is used to solve the extreme value, and the corresponding objective function is:

[0020]

[0021] It is equivalent to adding an L2 regularization term to the original loss function, which realizes the process of removing noise and finding signal, or the process of separating the true from the false in big data artificial intelligence. Among them, the first item requires the model to approximate the observed data at discrete points, and the result is to eliminate statistical noise and minimize the noise likelihood. The second item requires the model to approximate the observed structure on the Riemannian geometry space-time projection set, and the result is to eliminate the Riemannian geometry space-time structure noise and minimize the physical entropy.

[0022] Further, the third layer adopts the following Bose-Einstein probability distribution function as a probability model;

[0023] remember Let p be the probability corresponding to y, for The corresponding probabilities are:

[0024]

[0025] where f(x|θ,α)=θ μ δ μ (x-α δ )+o(…) (26)

[0026] The superscript θ of the Riemannian geometry vector in equation (26) μ Row vector and subscript δ μ The column vectors are repeated, representing the Einstein summation norm:

[0027]

[0028] Equation (34) is used to solve the extreme value, and the corresponding objective function is:

[0029]

[0030] It is equivalent to adding an L2 regularization term to the original loss function, which realizes the process of removing noise and finding signal, or the process of separating the true from the false in big data artificial intelligence. Among them, the first item requires the model to approximate the observed data at discrete points, and the result is to eliminate statistical noise and minimize the noise likelihood. The second item requires the model to approximate the observed structure on the Riemannian geometry space-time projection set, and the result is to eliminate the Riemannian geometry space-time structure noise and minimize the physical entropy.

[0031] Furthermore, the probability model includes multiple code segments that need to traverse all data rows for processing; each code segment that needs to process all data rows is implemented in the form of a pipeline core thread pipeline structure for reading and processing data in an FPGA structure controlled by the Risc-V standard instruction set.

[0032] Furthermore, the decision inference of the fourth layer and the loop feedback stage of the fifth layer also include making various decisions in different scenarios, and conducting a large number of backtests and out-of-sample tests through multi-dimensional and multi-time period data analysis for each decision in each scenario.

[0033] Furthermore, it also includes access functions for single-column sum, filtered sum, grouped sum, table connection grouped sum, dynamic grouped sum, dynamic ranking sum, and grouped percentile sum, and after obtaining the query results, provides visualization options to generate trend charts, bar charts, histograms, and heat maps that intuitively reflect the relationship between data.

[0034] Compared with the prior art, the present invention has the following beneficial effects: the core thread pipeline electron transfer computing architecture provided by the present invention is a big data artificial intelligence processing engine with "calculation", "search" and "storage" as the core. It is composed of a series of pipeline algorithms, starting from the import of time series data, through cleaning, sorting, indexing, grouping, optimization, to achieving prediction, making decisions, and completing adaptive loop feedback.

[0035] The innovations of the system engineering of the present invention include ultra-fast global distributed parallel sorting, space-efficient global indexing, dynamic grouping, on-demand optimization, decision inference and loop feedback. The solution operation for over 100 million data can be processed by single array loop processing or multiple arrays in parallel. The combination of multiple pipelines enables the present invention to achieve linear speed when analyzing over 100 million records and modeling hundreds of variables, and is suitable for completing supercomputing functions in ultra-large-scale digital integrated circuits through core thread pipeline control such as data registers, logic gates, cross-connection networks and other basic building blocks.

[0036] The present invention is not only an advanced comprehensive big data artificial intelligence system engineering, but also an innovative big data artificial intelligence processing chip implemented on the FPGA space-time structure controlled by the reduced instruction set computer RISC-V. Its purpose is to provide fast, high-performance and low-power services to extract intelligent information from today's real-time big data. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a schematic diagram of the core thread pipeline electron transfer computing architecture of the present invention;

[0038] Figure 2 This is a trend chart of capacity demand for big data AI;

[0039] Figure 3 It is a schematic diagram of the fusion of the transactional time series database and the analytical time series data matrix field of the present invention;

[0040] Figure 4 RISC-V FPGA structure diagram of the core thread pipeline algorithm of the present invention;

[0041] Figure 5 It is a schematic diagram of the expansion of the core thread pipeline electron transfer computing architecture of the present invention;

[0042] Figure 6 A schematic diagram of the present invention showing that heterogeneous devices can access main memory through an open cache coherent bus;

[0043] Figure 7 A table showing the number of transistors required to implement the core-thread pipeline algorithm of the present invention on a single chip. DETAILED DESCRIPTION

[0044] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0045] See also Figure 1 The core thread pipeline electron transfer computing architecture provided by the present invention is a five-layer pyramid architecture. When the time series data enters the system engineering of the present invention in real time, the processing of the first to fifth layers is also carried out in real time, and debugging and verification are continuously realized. The system engineering of the present invention can be applied to various different data. The larger the data volume, the more and more complex the participants, and the greater the frequency of the data flow, the more effective the architecture.

[0046] The first layer is "huge real-time dirty data cleaning", which basically fully automates the collection, cleaning, refinement, conversion, and loading of original time series data to ensure data integrity.

[0047] The second layer, "global hash sort index", uses the most advanced sort index technology. Each field is sorted and indexed separately, and an address is established for each data element, which exponentially improves the data processing speed. The present invention combines sorting and indexing together, and generates sorting and index creation logic at the same time. The sorting logic occurs within the time required to access each data element once. By sorting and indexing the entire amount of data, no information will be lost. At the same time, in the model analysis, each piece of information will be quickly located, speeding up the data analysis time. After the analysis data is captured, the analysis and optimization process in the model is carried out.

[0048] The third layer provides the ability to solve ultra-large, multi-dimensional manifold spaces in an ultra-fast and ultra-stable manner, helping with automatic modeling and intelligent optimization. The architecture can therefore thoroughly process massive amounts of time series data and automatically find the best variables. From another perspective, the algorithm can automatically group, package, model, and then intelligently optimize. When these two capabilities are combined to meet multiple different application requirements with different constraints and conditions, and to automatically find the best variables, the algorithm of the present invention provides a powerful real-time analysis application foundation for general-purpose artificial intelligence based on the 5G and 6G eras.

[0049] The fourth layer is "decision inference". Based on the modeling and analysis results of the third layer, decisions are made automatically, and the inference results are sent to the fifth layer for judgment. The core task of artificial intelligence is to make predictions based on massive time series data, and further complete intelligent decision-making and feedback. This layer gives the architecture the ability to provide more immediate predictions, decisions, and evaluations in big data.

[0050] The fifth layer "judges the results and completes the feedback loop of adaptive deep learning". This layer will detect the quality of the fourth layer's decisions through real-time profit monitoring, thereby further optimizing the model. Correct decisions will be returned to the third layer for repetition, amplification, and deepening, bringing more benefits; wrong decisions will automatically start from the first layer to relearn, analyze, and construct data until the correct decision is made.

[0051] The first two layers of the system engineering of the present invention provide digital integration capabilities, organically integrating all big data resources covering the entire value chain of the business on a standardized, unified, and efficient system platform. The third layer provides digital insight capabilities, mining the laws and values ​​hidden behind the data in all dimensions, and has the ability of digital management and intelligent decision-making. The last two layers provide digital innovation capabilities, and intelligent analysis and decision-making enable enterprises to develop differentiated competitive innovations.

[0052] The fifth layer is the core of the entire system engineering of the present invention. In fact, the first four layers of the architecture are the foundation of the fifth layer. Because in the future where billions of terminals and sensors are connected to each other, and in the future where Industry 4.0 emphasizes extremely accurate industrial control, intelligent decision-making, feedback and adaptive mechanisms are essential. Even if a big data processing system that lacks this mechanism can be used now, its computing efficiency will become low or even impossible to calculate in the era of data explosion of the Internet of Things and artificial intelligence.

[0053] The system engineering of the present invention can execute all steps in parallel between multiple computers (such as computing nodes). In order to achieve the overall performance goal, the number of computing nodes is initially set to meet both the database size requirements and the necessary user requirements. The present invention can increase the number of computing nodes at any time to meet new performance requirements or the number of record sources to be classified. The present invention keeps all nodes synchronized, aggregates intermediate results, and processes user tasks in a concurrent manner. The architecture can also be expanded by adding a higher-level main control core.

[0054] This ability to scale system performance in a cluster-parallel manner at the node level enables the overall throughput of the system engineering of the present invention to maintain a linear growth level far exceeding that achieved by other traditional methods. The present invention can scale 1,000PB (10 18Bytes) range of heterogeneous databases to sort, index, retrieve, manage, stream, buffer and deliver. This invention is not only advanced big data system engineering, but also an innovative artificial intelligence chip implemented on a RISC-V controlled FPGA, which aims to provide fast, high performance and low power services to extract intelligent information from today's real-time big data. When used in real-time mode, this invention can also discover implicit structures in historical data and, most importantly, monitor the overall time series alignment of real-time data.

[0055] 1. Sorting and Indexing Pipeline

[0056] After cleaning, the time series data enters the second layer of the system engineering of the present invention and will be automatically processed by the sorting and indexing pipeline. Sorting is an algorithm that arranges data elements in a certain order. The order is usually numeric, alphabetic, symbolic, or specified by preference. Sorting and indexing are the basis of many data warehouse operations and data analysis predictions. They fundamentally provide a descriptive basis for understanding the order of things happening in the data.

[0057] In the system engineering of the present invention, the second-layer fast sorting and efficient indexing are the prerequisites for the third-layer automatic grouping and on-demand optimization to achieve pipeline efficiency.

[0058] First, in any high-performance production environment, there needs to be fast access to raw, un-cleaned data and on-demand derived data. This is an integral part of creating an iterative machine learning knowledge infrastructure for any adaptive decision workflow process.

[0059] Secondly, when performing big data slice and dice queries, many existing analysis tools use a combination of indexes on some commonly used fields and linear table scans on non-indexed fields. Traditional index generation and use time is relatively long. Usually the index of a field is larger than the actual data field being indexed, so due to time and space constraints, only a few key fields of the data record can be indexed.

[0060] The system engineering of the present invention combines sorting and indexing together, generating sorting and index creation logic simultaneously. The sorting logic occurs in the time required to access each data element once. Previously accessed elements are downstream in the pipeline. After the sort order is established, the indexing portion will use a similar pipeline process to dynamically generate an index. The index also needs to be used quickly. The index implantation of the present invention is very suitable for handling non-unique key situations, such as the original field "state of address" or the derived field "credit score", where different data records may have the same value.

[0061] 2. Grouping and Optimizing Pipelines

[0062] After the sorting and indexing pipeline is complete, the data will enter the third-level grouping and optimization pipeline. This pipeline puts the data into a certain scale and scope to provide a diagnostic basis for better detection of signals, discovery of useful structures, and definition of logical and significant characteristics. The scale in the grouping scheme of the present invention is inherent to the structure of the data and is important for both signal-to-noise ratio control and discovery of identifying features. The scope is more or less a choice of human preference, business practices or industry practices, and government compliance. The final optimization pipeline transforms data into value, extracts valuable and responsible intelligence from the data, and builds predictive models from which it can be inferred what will happen in the future.

[0063] There are many different ways to group one-dimensional variables. One common automatic grouping method is to group by percentiles of the sort order. In this case, the hash table sets derived from the sort and index pipeline are particularly useful and efficient, because the start and end positions of the sort index are already associated with the data variables. Simple percentile groupings of all individual data attributes are essentially contained in the set of hash table indexes.

[0064] In this simple grouping scheme, the size of the groups can be large or small. For the convenience and simplicity of analyzing the data, a large scale can be chosen. A small scale can also be chosen to show greater granularity in the data space. For discrete events, there is always a scale threshold below which even the grouped data is essentially noise. The grouping scheme of the present invention is combined with a dynamic evaluation based on the signal-to-noise ratio. In addition, the present invention introduces the concept of multiple scales to obtain important global interaction terms between these variables.

[0065] The grouping scheme for multidimensional variables is more complex. The present invention uses a combination of subspace clustering and local neural networks to reveal effective data features, especially at the local scale. The clustered class is a group of data that is clustered in a way that makes the data in the same group more similar than the data in other groups. A neural network is a group of nodes or derivatives of data that identifies potential or hidden relationships by simulating the processes of the human brain. This additional grouping scheme is intended to discover small-scale interaction terms in the subspace of interest.

[0066] Before optimization, it is necessary to set up a pre-optimization pipeline or preprocessor that can detect and correct any potential errors in the input data, which is especially important for real-time data feeds. The solution of the present invention is to reallocate any erroneous or out-of-range input data into special values ​​according to the data dimension during the grouping pipeline stage. If the percentage of erroneous data is too large, special processing will be called. Otherwise, these errors will be treated as special groups in the dimensions of their respective data variables. Another use of the preprocessor is to pre-screen the usefulness of data variables when the number of variables exceeds the memory size of the chip design.

[0067] The final optimization pipeline of the third layer of the system engineering of the present invention involves the process of finding the most likely mathematical representation of the data. The present invention uses a standard maximum likelihood estimation method to derive the parameters of the conditional probability distribution. The restrictions on the parameter space include consideration of the signal-to-noise ratio, special treatment of zero and full data space, correction of model parameter high-order expansion deviations, invariance of constant translation between local parameters and global parameters, treatment of local parameter collinearity, etc. The maximum likelihood estimate of the model parameters is iteratively obtained using the path weighted gradient search method.

[0068] The kernel pipeline algorithm of the present invention will judge whether the model meets the standard based on five criteria, namely (1) resolution: the degree of distinguishing good from bad sorting, using KS, Gini or AUC of cross-time samples for judgment. The specific resolution indicator to be used depends on the application scenario. The three statistical indicators of model resolution can be written as: KS = max{F B -F G}; Gini = ∫F B dF G ,AUC=∫F G dF B , where F G and F B are the good and bad cumulative sample curves along the scoring path. It is easy to see from the mathematical formula that KS is the maximum difference at a certain point, and Gini and AUC are both average differences in a certain statistical sense; (2) Accuracy: The statistical error between the data and the model, expressed as the cumulative average chi-square (χ 2 ) value judgment; (3) Algorithm efficiency: the computing speed of model training and reasoning, judged by the length of time required; (4) Storage efficiency: the storage space required for model training and reasoning, judged by the size of the memory; (5) Energy efficiency: the power consumption required for model operation, judged by the amount of energy consumed per unit time.

[0069] 3. Decision Inference and Loop Feedback

[0070] After the model is established, how should the model be used to predict the results? This requires entering the fourth level of the system engineering of the present invention, the "decision inference" stage, making various decisions under different scenarios (conditions), and making a large number of comparisons for each decision under each scenario, which involves multi-dimensional influencing factors and multi-time period variable analysis.

[0071] In order to call the model-related data (including original time series data, target variables, model variables, model prediction values) for query and analysis, the zero-code access platform of the system engineering of the present invention is used, and you can access the data according to your needs by just clicking the mouse. At present, the zero-code access platform can provide a series of complex access functions such as single-column summation, screening summation, grouping summation, table connection grouping summation, dynamic grouping summation, dynamic ranking summation, grouping percentile, etc., and after obtaining the query results, it provides visualization options to generate tools such as trend charts, bar charts, histograms, heat maps, etc. that intuitively reflect the relationship between data, so as to better serve decision-making, backtesting, screening, and execution.

[0072] After the decision is executed, the fifth layer of the system engineering of the present invention, the "loop feedback" stage, will be entered, and the quality of the decision execution will be judged through real-time profit monitoring. If the decision is correct (profitable), it will return to the third layer to repeat, amplify, and deepen the existing model to bring more benefits; if the decision is wrong (loss), it will automatically return to the first layer, re-learn, analyze, build models, screen decisions, and judge profits and losses from the original data until the correct (profitable) decision is made, realizing an adaptive loop feedback closed loop.

[0073] 3.1 Core Thread Pipeline Algorithm for Fermi-Dirac Distribution

[0074] Discrete events are the most common type of events we encounter in our daily lives. These events have discrete results and belong to the category of mathematical classification problems. We use the Fermi-Dirac probability distribution to build a model for the residual. For example, a doctor diagnoses whether a tumor in a patient's body is malignant, which is a binary event; college students choose to study abroad, take the postgraduate entrance examination, or work after graduation, which is a ternary event; WeChat or Douyin users will like, comment, collect, forward, or recommend a short video after viewing it, which is a multivariate event.

[0075] From simple binary events to complex multivariate events, there are many different ways to analyze and explain the events of daily life that people observe, so as to achieve intelligent inference. In the simplest binary case, you can come up with a rule to "estimate" the conditional probability of each binary event from some explanatory factors (technically called factor variables). This is the use of probability in statistical mathematics to explain binary events. Because multivariate events can be decomposed into multiple binary events, this section will use binary events as an example to illustrate.

[0076] In the quantum world, all observable data values ​​are discrete, rather than continuous results of classical physics. The probability analysis method of binary events also has a profound impact on understanding the physical world. Below, the present invention first derives the mathematical formula of the nuclear thread pipeline algorithm of the Fermi-Dirac probability distribution from the extreme value problem of statistical mathematics, and finally introduces the physical meaning of the nuclear thread pipeline algorithm of the Fermi-Dirac distribution from the physical principle.

[0077] 3.1.1 Extreme Value Problem

[0078] Consider a binary classification problem, given a data set:

[0079] D = {x i ,y i},y i ∈{0,1} (1)

[0080] remember Let p be the probability corresponding to y, for The corresponding probabilities are:

[0081]

[0082] where f(x|θ,α)=θ μ δ μ (x-α δ )+o(…) (3)

[0083] The superscript (θ) of the Riemannian geometry vector in equation (3) μ row vector) and subscript (δ μ Column vectors) appear repeatedly, representing the Einstein summation norm:

[0084]

[0085] The likelihood function for a binary event is:

[0086]

[0087] Taking the logarithm of both sides of the equation, we can write it as the log-likelihood function:

[0088]

[0089] In the classification problem, the present invention has the concept of loss function, which measures the degree of model prediction error.

[0090] Take the inverse of the average log-likelihood loss over all discrete datasets:

[0091]

[0092] That is, in classification problems, maximizing the likelihood function and minimizing the loss function are actually equivalent.

[0093] 3.1.2 Constraints

[0094] In order to avoid overfitting of extreme value solutions, regularization constraints can be added. That is, on the basis of minimizing the training error, a model with a stable structure is adopted as much as possible to improve the generalization of the model. λ is the normalization coefficient, see formula (8), and the definition of ρ is shown in formula (9)

[0095] Using L2 regularization, the likelihood function can be rewritten as:

[0096]

[0097] Where ρ is a function defined on the data structure set δ, note:

[0098]

[0099] Taking the logarithm of L(θ,α), we get:

[0100]

[0101] The corresponding objective function (loss function) is:

[0102]

[0103] It is equivalent to adding the L2 regularization term to the original loss function.

[0104] 3.1.3 Extreme value solution

[0105] The main goal of solving the classification problem is to find the parameter θ μ and α δ Minimize the loss function J(θ,α). In the first part, use Newton's iteration method to solve θ μ As follows. Perform multivariate Taylor expansion of J(θ) to the second-order derivative:

[0106]

[0107]

[0108] Let J′(θ μ ) = 0 gives:

[0109]

[0110] Parameter θ μ The iterative solution can be written as:

[0111]

[0112] The first-order partial derivative (gradient vector) of the loss function:

[0113]

[0114]

[0115] The second-order partial derivative of the loss function (Hessian matrix):

[0116]

[0117] When the iteration converges, ρ<<1, and the second-order derivative can be simplified to:

[0118]

[0119] The parameter α is derived below δ The second part uses the machine learning gradient method to solve α δ As follows. Perform multivariate Taylor expansion of J(α) to the first-order derivative:

[0120]

[0121]

[0122] Using a derivation process similar to equation (16), we can obtain:

[0123]

[0124] When the iteration converges, ρ<<1, and the first-order derivative is simplified to:

[0125]

[0126] Parameter α δ The iterative solution can be taken as follows:

[0127]

[0128] Among them, the learning step size is proportional to The constant.

[0129] 3.1.4 Physical meaning

[0130] In the quantum world, the probability model of equation (2) describes the energy distribution of fermions, referred to as the Fermi-Dirac probability distribution function, where the parameter f is the relative energy difference between free energy and chemical energy. In the core thread pipeline algorithm of the present invention, the extreme value solution of equation (11) is actually the process of removing noise and seeking signals in the physical world, or the process of removing the false and retaining the true in big data artificial intelligence. Among them, the first item requires the model to approximate the observed data at discrete points, and the result is to eliminate statistical noise and minimize the noise likelihood; the second item requires the model to approximate the observed structure on the Riemannian geometry space-time projection set, and the result is to eliminate the Riemannian geometry space-time structure noise and minimize the physical entropy.

[0131] 3.2 Core Thread Pipeline Algorithm for Bose-Einstein Distribution

[0132] The value prediction of continuous variables is also a type of problem we often encounter in our daily lives. These problems all want to predict the results of continuous values, which belong to the category of mathematical regression problems. The present invention uses the Bose-Einstein probability distribution to establish a model for the residual. For example, predicting the income of platform users (4,000, 20,000, 100,000...), predicting the commuting distance of employees (500m, 1km, 20,000 miles...), predicting the rate of return of stock transactions (1%, 50%, 99%...), etc. These problems all want to predict the results of continuous values, which belong to the category of mathematical regression problems.

[0133] Similar to the classification problem mentioned above, the present invention also needs to come up with a rule to "estimate" the value of the target variable from some explanatory factors (factor variables), that is, to establish a mathematical model for the relationship between one or more independent variables. Next, the mathematical formula of the core thread pipeline algorithm of the Bose-Einstein probability distribution will be derived from the extreme value problem of statistical mathematics.

[0134] 3.2.1 Extreme Value Problem

[0135] Consider the problem of predicting the value of a continuous variable, given a data set:

[0136]

[0137] remember Let p be the model prediction value corresponding to y, for The corresponding model prediction values ​​are:

[0138]

[0139] where f(x|θ,α)=θ μ δ μ (x-α δ )+o(…) (26)

[0140] The superscript (θ) of the Riemannian geometry vector in equation (26) μ row vector) and subscript (δ μ Column vectors) appear repeatedly, representing the Einstein summation norm:

[0141]

[0142] The likelihood function for a continuous variable is:

[0143]

[0144] Taking the logarithm of both sides of the equation, we can write it as the log-likelihood function:

[0145]

[0146] There is also the concept of loss function in the problem of predicting the value of continuous variables, which measures the degree of model prediction error.

[0147] Take the inverse of the average log-likelihood loss over all discrete datasets:

[0148]

[0149] That is, in the problem of predicting the value of a continuous variable, maximizing the likelihood function and minimizing the loss function are also equivalent.

[0150] 3.2.2 Constraints

[0151] In order to avoid overfitting of extreme value solutions, regularization constraints can be added. That is, on the basis of minimizing the training error, a model with a stable structure is adopted as much as possible to improve the generalization of the model.

[0152] Using L2 regularization, the likelihood function can be rewritten as:

[0153]

[0154] Where ρ is a function defined on the data structure set δ, note:

[0155]

[0156] Taking the logarithm of L(θ,α), we get:

[0157]

[0158] The corresponding objective function (loss function) is:

[0159]

[0160] It is equivalent to adding the L2 regularization term to the original loss function.

[0161] 3.2.3 Extreme value solution

[0162] The main goal of solving the problem of predicting the value of a continuous variable is to find the parameter θ μ and α δ Minimize the loss function J(θ,α). In the first part, use Newton's iteration method to solve θ μ As follows. Perform multivariate Taylor expansion of J(θ) to the second-order derivative:

[0163]

[0164]

[0165] Let J′(θ μ ) = 0 gives:

[0166]

[0167] Parameter θ μ The iterative solution can be written as:

[0168]

[0169] The first-order partial derivative (gradient vector) of the loss function:

[0170]

[0171] The second-order partial derivative of the loss function (Hessian matrix):

[0172]

[0173] When the iteration converges, ρ<<1 and The second-order derivative can be simplified to:

[0174]

[0175] The parameter α is derived below δ The second part uses the machine learning gradient method to solve α δ As follows. Perform multivariate Taylor expansion of J(α) to the first-order derivative:

[0176]

[0177]

[0178] Using a derivation process similar to equation (39), we can obtain:

[0179]

[0180] When the iteration converges, ρ<<1, and the first-order derivative is simplified to:

[0181]

[0182] Parameter α δ The iterative solution can be taken as follows:

[0183]

[0184] Among them, the learning step size is proportional to The constant.

[0185] 3.2.4 Physical meaning

[0186] In the quantum world, the probability model of equation (25) describes the energy distribution of bosons, referred to as the Bose-Einstein probability distribution function, where the parameter f is the relative energy difference between free energy and chemical energy. In the core thread pipeline algorithm of the present invention, the extreme value solution of equation (34) is actually the process of removing noise and seeking signals in the physical world, or the process of removing the false and retaining the true in big data artificial intelligence. Among them, the first item requires the model to approximate the observed data at discrete points, and the result is to eliminate statistical noise and minimize the noise likelihood; the second item requires the model to approximate the observed structure on the Riemannian geometry space-time projection set, and the result is to eliminate the Riemannian geometry space-time structure noise and minimize the physical entropy.

[0187] 4. Computing power scalability

[0188] Big data and artificial intelligence, an important technology, have been in great demand since the technological breakthrough in 2012. They have become very popular, causing the demand for AI computing power to explode by 300,000 times in just six years. It can be said that it doubles every 100 days on average. Figure 2 shown.

[0189] The computing power scalability of the system engineering of the present invention is huge, because its core core thread pipeline algorithm does not use random sampling in the "training stage", but needs to perform model training on the entire training data set, that is, each iteration needs to traverse all data. In addition, in the fourth layer "decision inference" and the fifth layer "loop feedback" stages, it is necessary to make various different decisions under different scenarios (conditions), and do a large number of backtests and out-of-sample tests for each decision in each scenario, which involves multi-dimensional and multi-time period data analysis, and the demand for computing power is greater than the training stage.

[0190] The electron transfer model of the core thread pipeline algorithm of the present invention includes the Fermi-Dirac space-time energy distribution model, the Bose-Einstein space-time energy distribution model and other test models and working models. Each distribution model has multiple electron transfers that need to traverse all data rows for processing.

[0191] Each logical calculation that needs to process all the data rows for the electronic transfer is suitable for the FPGA to read and process the data in the form of a pipeline structure, because these electronic transfers will be called frequently, and the benefits of accelerating such processes are the greatest. Some loop iterations can run multiple copies in parallel in an FPGA, while other loop iterations can only run a single copy. Because multiple logical operations occur at a time, running pipeline operations in FPGA logic can greatly speed up calculations, such as Figure 4 shown.

[0192] For highly pipelined electronic transfer logic, data can be processed at a rate close to the maximum I / O rate of the main memory. If more FPGA logic space is needed, multiple chips can be combined on a circuit integration substrate so that all chips can access the memory, and the system expansion can be as large as Figure 5 In the above design, the present invention will use TileLink and OmniXtend of the RISC-V standard instruction set to achieve cache consistency, coherence and scalability.

[0193] TileLink is an open standard chip interconnect designed for RISC-V processors. Its purpose is to ensure cache-consistent memory reads and writes. It can be used in system-on-chip (SoC) to connect accelerators, FPGA-based processing blocks, general-purpose multiprocessors, coprocessors, and DMA engines to achieve complex designs. With the increasing application of RISC-V in the microarchitectures of many processors, the RISC-V ecosystem urgently needs to provide a common extension protocol for the cache consistency bus, such as OmniXtend. It is the first cross-device cache consistency memory technology that uses the programmability of modern Ethernet switches to support the processor's cache and memory controllers and directly exchange consistency messages through Ethernet-compatible structures. It provides an open standard interface for memory access and data sharing for various processors, FPGAs, GPUs, machine learning accelerators, and other components, allowing heterogeneous RISC-V devices to access main memory. For details, see Figure 6 .

[0194] For any distributed model of the core-thread pipeline algorithm, the present invention can implement 33 code blocks and about 600,000 programmable logic blocks (CLBs), which requires about 672,000 CMOS transistors. In addition to other resources (such as interfaces) and FPGA resources (such as BRAM), it is estimated that the total number of transistors required to implement the optimization algorithm on a single chip is at least 100 million. For details, see Figure 7 .

[0195] Although the present invention has been disclosed as above with preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art may make some modifications and improvements without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be based on the definition of the claims.

Claims

1. A core thread pipeline electron transfer computing architecture, characterized in that: Adopt the following tightly integrated five-layer pyramid architecture: The first layer is a large real-time dirty data cleaning layer, which realizes the collection, cleaning, refinement, conversion, and loading of original time series data, and provides data integrity; The second layer is the global hash sorting index layer, which sorts and indexes each field separately, establishes an address for each data element, combines sorting and indexing together, and generates sorting and index creation logic at the same time; The third layer provides the ability to solve ultra-large, multi-dimensional manifold spaces in an ultra-fast and ultra-stable manner, helping with automatic modeling and intelligent optimization; The fourth layer is the decision-making inference layer, which automatically makes decisions based on the modeling and analysis results of the third layer, and sends the inference results to the fifth layer for judgment; The fifth layer is used to judge whether the results are good or bad and complete the loop feedback of adaptive deep learning. This layer will detect the quality of the fourth layer's decisions through real-time profit monitoring. If the decision is correct, it will return to the third layer for repeated optimization. If the decision is wrong, it will automatically start from the first layer to re-learn, analyze, and construct data until the correct decision is made.

2. The core-thread pipeline electron transfer computing architecture of claim 1, characterized in that: The second layer automatically sorts and indexes the time series data cleaned by the first layer through pipeline processing. The sorting logic occurs within the time required to access each data element once. The previously accessed elements are downstream in the pipeline. After the sorting order is established, the indexing part will dynamically generate indexes using a similar pipeline process.

3. The core-thread pipeline electron transfer computing architecture of claim 1, wherein: The third layer adopts a standard maximum likelihood estimation method to derive the parameters of the conditional probability distribution, and uses a path weighted gradient search method to iteratively obtain the maximum likelihood estimation of the model parameters.

4. The core-thread pipeline electron transfer computing architecture of claim 3, characterized in that: The restrictions on the parameter space in the third layer include consideration of the signal-to-noise ratio, special treatment of zero and full data space, correction of high-order expansion bias of model parameters, invariance of constant translation between local and global parameters, and treatment of local parameter collinearity.

5. The core-thread pipeline electron transfer computing architecture of claim 3, characterized in that: The process of judging whether the model in the third layer meets the standard is as follows: (1) Resolution: The degree of distinguishing good from bad rankings, which is judged by KS, Gini or AUC of cross-time samples. The specific resolution index to be used depends on the application scenario. The three statistical indicators of model resolution are written as: KS = max{F B -F G }; Gini = ∫F B dF G ,AUC=∫F G dF B , where F G and F B They are respectively the good and bad cumulative sample curves along the scoring path, KS is the maximum difference at a certain point, and Gini and AUC are both average differences in a certain statistical sense; (2) Accuracy: The statistical error between the data and the model, expressed as the cumulative average chi-squared value on the model variable dimension. 2 Value judgment; (3) Algorithm efficiency: the computational speed of model training and inference, measured by the time required; (4) Storage efficiency: The storage space required for model training and inference is determined by the size of the memory; (5) Energy efficiency: The power consumption required for model operation, which is determined by the amount of energy consumed per unit time.

6. The core-thread pipeline electron transfer computing architecture of claim 3, characterized in that: The third layer adopts the following Fermi-Dirac probability distribution function as a probability model; remember Let p be the probability corresponding to y, for The corresponding probabilities are: where f(x|θ,α)=θ μ d μ (x-a δ )+o(…) (3) The superscript θ of the Riemannian geometry vector in equation (3) μ Row vector and subscript δ μ The column vectors are repeated, representing the Einstein summation norm: Equation (11) is used to solve the extreme value, and the corresponding objective function is: It is equivalent to adding an L2 regularization term to the original loss function, which realizes the process of removing noise and finding signal, or the process of separating the true from the false in big data artificial intelligence. Among them, the first item requires the model to approximate the observed data at discrete points, and the result is to eliminate statistical noise and minimize the noise likelihood. The second item requires the model to approximate the observed structure on the Riemannian geometry space-time projection set, and the result is to eliminate the Riemannian geometry space-time structure noise and minimize the physical entropy.

7. The core-thread pipeline electron transfer computing architecture of claim 3, characterized in that: The third layer adopts the following Bose-Einstein probability distribution function as the probability model; remember Let p be the probability corresponding to y, for The corresponding probabilities are: Where f(x|θ,α)=θ μ d μ (x-a δ )+o(…) (26) The superscript θ of the Riemannian geometry vector in equation (26) μ Row vector and subscript δ μ The column vectors are repeated, representing the Einstein summation norm: Equation (34) is used to solve the extreme value, and the corresponding objective function is: It is equivalent to adding an L2 regularization term to the original loss function, which realizes the process of removing noise and finding signal, or the process of separating the true from the false in big data artificial intelligence. Among them, the first item requires the model to approximate the observed data at discrete points, and the result is to eliminate statistical noise and minimize the noise likelihood. The second item requires the model to approximate the observed structure on the Riemannian geometry space-time projection set, and the result is to eliminate the Riemannian geometry space-time structure noise and minimize the physical entropy.

8. The core-thread pipeline electron transfer computing architecture according to claim 6 or 7, characterized in that: The probability model includes multiple code segments that need to traverse all data rows for processing; each code segment that needs to process all data rows is implemented in the form of a pipeline core thread pipeline structure in an FPGA structure controlled by the Risc-V standard instruction set for reading and processing data.

9. The core-thread pipeline electron transfer computing architecture of claim 1, wherein: The fourth-level decision inference and the fifth-level loop feedback stage also include making various decisions in different scenarios, and conducting a large number of backtests and out-of-sample tests through multi-dimensional and multi-time period data analysis for each decision in each scenario.

10. The core-thread pipeline electron transfer computing architecture of claim 1, wherein: It also provides access to single-column sum, filtered sum, grouped sum, table join grouped sum, dynamic grouped sum, dynamic ranking sum, and grouped percentile sum. After obtaining the query results, it provides visualization options to generate trend charts, bar charts, histograms, and heat maps that intuitively reflect the relationship between data.

Images