WO2022006734A1 - 一种数据稀疏投影算法、***、电子装置及存储介质 - Google Patents

一种数据稀疏投影算法、***、电子装置及存储介质 Download PDF

Info

Publication number
WO2022006734A1
WO2022006734A1 PCT/CN2020/100618 CN2020100618W WO2022006734A1 WO 2022006734 A1 WO2022006734 A1 WO 2022006734A1 CN 2020100618 W CN2020100618 W CN 2020100618W WO 2022006734 A1 WO2022006734 A1 WO 2022006734A1
Authority
WO
WIPO (PCT)
Prior art keywords
dimensional
sparse
dimensional matrix
output variable
matrix
Prior art date
Application number
PCT/CN2020/100618
Other languages
English (en)
French (fr)
Inventor
顾崇林
马畅翼
李文烨
崔曙光
Original Assignee
香港中文大学（深圳）
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 香港中文大学（深圳） filed Critical 香港中文大学（深圳）
Priority to PCT/CN2020/100618 priority Critical patent/WO2022006734A1/zh
Publication of WO2022006734A1 publication Critical patent/WO2022006734A1/zh
Priority to US18/133,523 priority patent/US11853387B2/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization

Definitions

  • the present invention relates to the technical field of data processing, and in particular, to a data sparse projection algorithm, a system, an electronic device and a storage medium.
  • Random projection technology is an important tool for data analysis. It uses the calculation method of matrix multiplication to re-express the original input sample data (matrix form), so as to obtain better performance while protecting the characteristics of the original data, such as: calculation speed , storage space, accuracy, etc.
  • the existing random projection techniques mainly include LSH (locality sensitive hashing, locality aware hashing) algorithm and FLY (fruit fly algorithm).
  • composition of LSH and FLY is relatively simple, which makes the generation of projection matrix too random, resulting in that the similarity features between samples before and after projection cannot be protected, and because the output results are constantly changing each time, it is unstable in practical applications.
  • the main purpose of the present invention is to provide a data sparse projection algorithm, system, electronic device and storage medium, aiming to solve the problem that the generation of projection matrix in the prior art is too random, resulting in that the similarity feature between samples before and after projection cannot be protected, And because the output results are constantly changing each time, it leads to unstable technical problems in practical applications.
  • a first aspect of the present invention provides a data sparse projection algorithm, including: randomly initializing a high-dimensional sparse two-dimensional matrix; fixing the high-dimensional sparse two-dimensional matrix, and using the high-dimensional sparse two-dimensional matrix Calculate the optimal output variable; fix the optimal output variable, and use the optimal output variable to calculate the optimal high-dimensional sparse two-dimensional matrix; cyclically fix the high-dimensional sparse two-dimensional matrix and the output variable until the fixed position is reached.
  • the optimal output variable is no longer increased, and the final high-dimensional sparse two-dimensional matrix is generated at this time.
  • fixing the high-dimensional sparse two-dimensional matrix and calculating the optimal output variable using the high-dimensional sparse two-dimensional matrix includes: training a set according to an existing sample, the sample including an input vector or an input vector. and the output variable, the input vector is a high-dimensional sparse two-dimensional matrix; the winner-take-all algorithm is used to establish the functional relationship between the input vector and the output vector in the set; the high-dimensional sparse two-dimensional matrix is fixed, according to the fixed The high-dimensional sparse two-dimensional matrix and the functional relationship calculate the optimal output variable.
  • the fixing the high-dimensional sparse two-dimensional matrix, and using the high-dimensional sparse two-dimensional matrix to calculate the optimal output variable also includes: according to a given constraint value, for the high-dimensional sparse two-dimensional matrix Make sparsity constraints; constrain the output variables according to the winner-take-all algorithm and the constrained high-dimensional sparse two-dimensional matrix.
  • a second aspect of the present invention provides a data sparse projection system, comprising: an initialization module for randomly initializing a high-dimensional sparse two-dimensional matrix; an optimal output variable calculation module for initializing the high-dimensional sparse two-dimensional matrix initialized by the initialization module
  • the dimensional matrix is fixed, and the high-dimensional sparse two-dimensional matrix is used to calculate the optimal output variable; the optimal high-dimensional sparse two-dimensional matrix calculation module is used to calculate the optimal output variable calculated by the optimal output variable calculation module.
  • Fix use the optimal output variable to calculate the optimal high-dimensional sparse two-dimensional matrix; the loop module is used to cyclically fix the high-dimensional sparse two-dimensional matrix and output variables until the optimal high-dimensional sparse two-dimensional matrix is fixed.
  • the optimal output variable no longer increases when a high-dimensional sparse two-dimensional matrix is obtained by the matrix calculation module.
  • the optimal output variable calculation module includes: a set training unit for training a set according to an existing sample, the sample includes an input vector or an input vector and an output variable, and the input vector is a high-dimensional sparse two-dimensional matrix; a functional relationship establishing unit for establishing a functional relationship between the input vector and the output vector in the set using the winner-take-all algorithm; a computing unit for fixing a high-dimensional sparse two-dimensional matrix, and according to the fixed The high-dimensional sparse two-dimensional matrix in the set trained by the set training unit and the functional relationship established by the functional relationship establishing unit calculate the optimal output variable.
  • the optimal output variable calculation module also includes: a high-dimensional sparse two-dimensional matrix constraint unit, used for making sparsity constraints on the high-dimensional sparse two-dimensional matrix according to a given constraint value; an output variable constraint unit, using The output variables are constrained according to the winner-take-all algorithm and the high-dimensional sparse two-dimensional matrix constrained by the high-dimensional sparse two-dimensional matrix constraining unit.
  • a third aspect of the present invention provides an electronic device, comprising: a memory, a processor, and a computer program stored on the memory and executable on the processor, when the processor executes the computer program, the above-mentioned any of the methods described in .
  • a fourth aspect of the present invention provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, implements any one of the methods described above.
  • the invention provides a data sparse projection algorithm, which has the beneficial effects of continuously optimizing the high-dimensional sparse two-dimensional matrix through random and iterative methods, which can make the high-dimensional sparse two-dimensional matrix more stable and better protect the similarity feature of the original data. , its sparsity and binary properties make it computationally faster in practical applications.
  • FIG. 1 is a schematic block diagram of the flow of a data sparse projection algorithm according to an embodiment of the present invention
  • FIG. 2 is a schematic block diagram of the structure of a data sparse projection system according to an embodiment of the present invention
  • FIG. 3 is a schematic block diagram of the structure of an electronic device according to an embodiment of the present invention.
  • Figure 1 is a data sparse projection algorithm, including: S1, randomly initialize a high-dimensional sparse two-dimensional matrix; S2, fix the high-dimensional sparse two-dimensional matrix, and use the high-dimensional sparse two-dimensional matrix to calculate the optimal output variable; S3, fix the optimal output variable, and use the optimal output variable to calculate the optimal high-dimensional sparse two-dimensional matrix; S4, loop to fix the high-dimensional sparse two-dimensional matrix and output variables, until the fixed high-dimensional sparse two-dimensional matrix is fixed. The optimal output variable no longer increases when the dimensional matrix is used.
  • W 1 is randomly initialized and denoted as W 1
  • W and Y are continuously optimized through interactive iteration
  • Y is the output variable.
  • the specific iterative optimization process is as follows:
  • Fix the high-dimensional sparse two-dimensional matrix, and use the high-dimensional sparse two-dimensional matrix to calculate the optimal output variables include: training sets based on existing samples, the samples include input vectors or input vectors and output variables, and the input vectors are high-dimensional sparse two-dimensional dimensional matrix; use the winner-take-all algorithm to establish the functional relationship between the input vector and the output vector in the set; fix the high-dimensional sparse two-dimensional matrix, and calculate the optimal sparse two-dimensional matrix according to the fixed high-dimensional sparse two-dimensional matrix and the functional relationship output variable.
  • Fixing the high-dimensional sparse two-dimensional matrix, and using the high-dimensional sparse two-dimensional matrix to calculate the optimal output variable also includes: according to the given constraint value, make sparsity constraints on the high-dimensional sparse two-dimensional matrix; according to the winner-take-all algorithm and The constrained high-dimensional sparse two-dimensional matrix performs binarization constraints on the output variables.
  • k represents the hash length of the projection. Due to the noise in the actual training data, in order to find the projection matrix W, the problem is transformed into a maximization problem about the variable W and the variable Y, namely:
  • our function value can obtain a local optimal solution, which converges when the function value L(W, Y) no longer rises.
  • a data sparse projection algorithm provided in this embodiment of the present application generates a high-quality projection matrix, and for a sample with given input and output, the optimal high-dimensional sparse projection matrix can be generated directly according to the above method.
  • dimensional matrix suitable for supervised learning.
  • the high-dimensional sparse two-dimensional matrix is continuously optimized by random high-dimensional sparse two-dimensional matrix and iteration, so that the generated projection matrix is more stable, and the original projection matrix is more stable.
  • the data similarity feature is better protected, and its sparsity and binary properties make it faster in practical applications and suitable for unsupervised learning. Therefore, the present invention is not only suitable for supervised learning, but also for unsupervised learning, and the projection matrix in the present invention is learned from the input data, which is more stable, better in feature protection, and faster in calculation than existing methods. beneficial effect.
  • an embodiment of the present application further provides a data sparse projection system, including: an initialization module 1, an optimal output variable calculation module 2, an optimal high-dimensional sparse two-dimensional matrix calculation module 3, and a loop module 4; an initialization module 1 is used to randomly initialize a high-dimensional sparse two-dimensional matrix; the optimal output variable calculation module 2 is used to fix the high-dimensional sparse two-dimensional matrix initialized by the initialization module 1, and use the high-dimensional sparse two-dimensional matrix to calculate the optimal output variable; The optimal high-dimensional sparse two-dimensional matrix calculation module 3 is used to fix the optimal output variable calculated by the optimal output variable calculation module 2, and use the optimal output variable to calculate the optimal high-dimensional sparse two-dimensional matrix; the loop module 4 is used to cyclically fix the high-dimensional sparse two-dimensional matrix and output variables, until the output variable obtained by the optimal output variable calculation module 2 is fixed, the high-dimensional sparse two-dimensional matrix will no longer increase.
  • an initialization module 1 is used to randomly initialize a high
  • the optimal output variable calculation module 2 includes: a set training unit, a functional relationship establishment unit and a calculation unit; the set training unit is used to train a set according to an existing sample, the sample includes an input vector or an input vector and an output variable, and the input vector is a high-dimensional Sparse two-dimensional matrix; the functional relationship establishment unit is used to use the winner-take-all algorithm to establish the functional relationship between the input vector and the output vector in the set; the calculation unit is used to fix the high-dimensional sparse two-dimensional matrix, and according to the fixed set The high-dimensional sparse two-dimensional matrix in the set trained by the training unit and the functional relationship established by the functional relationship establishment unit calculate the optimal output variable.
  • the optimal output variable calculation module 2 also includes: a high-dimensional sparse two-dimensional matrix constraint unit and an output variable constraint unit; the high-dimensional sparse two-dimensional matrix constraint unit is used to sparse the high-dimensional sparse two-dimensional matrix according to a given constraint value The output variable constraint unit is used to constrain the output variables according to the winner-take-all algorithm and the high-dimensional sparse two-dimensional matrix constrained by the high-dimensional sparse two-dimensional matrix constraint unit.
  • An embodiment of the present application provides an electronic device, please refer to 3.
  • the electronic device includes: a memory 601, a processor 602, and a computer program stored in the memory 601 and executable on the processor 602, and the processor 602 executes the computer program , the fault response method of the semiconductor processing module described in the foregoing is implemented.
  • the electronic device further includes: at least one input device 603 and at least one output device 604 .
  • the above-mentioned memory 601 , processor 602 , input device 603 and output device 604 are connected through a bus 605 .
  • the input device 603 may specifically be a camera, a touch panel, a physical button, a mouse, or the like.
  • the output device 604 may specifically be a display screen.
  • the memory 601 can be a high-speed random access memory (RAM, R and om Access Memory) memory, and can also be a non-volatile memory (non-volatile memory), such as a disk memory.
  • Memory 601 is used to store a set of executable program codes, and processor 602 is coupled to memory 601 .
  • an embodiment of the present application further provides a computer-readable storage medium.
  • the computer-readable storage medium may be provided in the electronic device in the above-mentioned embodiments, and the computer-readable storage medium may be one of the above-mentioned embodiments.
  • a computer program is stored on the computer-readable storage medium, and when the program is executed by the processor 602, the method for responding to the fault of the semiconductor processing module described in the foregoing method embodiments is implemented.
  • the computer-storable medium can also be a USB flash drive, a removable hard disk, a read-only memory 601 (ROM, Read-Only Memory), a RAM, a magnetic disk or an optical disk and other media that can store program codes.
  • ROM Read-Only Memory
  • RAM Random Access Memory
  • magnetic disk or an optical disk and other media that can store program codes.
  • the disclosed apparatus and method may be implemented in other manners.
  • the apparatus embodiments described above are only illustrative.
  • the division of the modules is only a logical function division. In actual implementation, there may be other division methods.
  • multiple modules or components may be combined or Can be integrated into another system, or some features can be ignored, or not implemented.
  • the shown or discussed mutual coupling or direct coupling or communication connection may be through some interfaces, indirect coupling or communication connection of devices or modules, and may be in electrical, mechanical or other forms.
  • modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical modules, that is, may be located in one place, or may be distributed to multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution in this embodiment.
  • each functional module in each embodiment of the present invention may be integrated into one processing module, or each module may exist physically alone, or two or more modules may be integrated into one module.
  • the above-mentioned integrated modules can be implemented in the form of hardware, and can also be implemented in the form of software function modules.
  • the integrated modules are implemented in the form of software functional modules and sold or used as independent products, they may be stored in a computer-readable storage medium.
  • the technical solution of the present invention is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium , including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Pure & Applied Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • Computational Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Computing Systems (AREA)
  • Algebra (AREA)
  • Databases & Information Systems (AREA)
  • Software Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Complex Calculations (AREA)
  • Image Analysis (AREA)

Abstract

一种数据稀疏投影方法，包括：随机初始化一个高维稀疏二维矩阵（S1）；将高维稀疏二维矩阵固定，使用高维稀疏二维矩阵计算最优的输出变量（S2）；将最优的输出变量固定，使用最优的输出变量计算最优的高维稀疏二维矩阵（S3）；循环固定所述高维稀疏二维矩阵及输出变量，直至在固定所述高维稀疏二维矩阵时最优的输出变量不再增大（S4）；通过随机和迭代的方式不断地优化高维稀疏二维矩阵，能够使得高维稀疏二维矩阵更加稳定，对原始数据相似度特征保护更好，并且该算法的稀疏性和二值性使得其在实际应用中计算速度更快。

Description

一种数据稀疏投影算法、***、电子装置及存储介质 技术领域
本发明涉及数据处理技术领域，尤其涉及一种数据稀疏投影算法、***、电子装置及存储介质。
背景技术
随机投影技术是数据分析的重要工具，其利用矩阵相乘的计算方式对原始输入的样本数据(矩阵形式)进行重新表达，从而在保护原始数据特征的同时获得更好的性能，比如：计算速度、存储空间、精确度等。
现有的随机投影技术主要有LSH(locality sensitive hashing，局部感知哈希)算法和FLY(果蝇算法)。
但是，LSH和FLY的构成较为简单，使得投影矩阵的生成过于随机，导致投影前后样本之间的相似度特征无法得到保护，且由于每次输出结果不断变化，导致在实际应用中不稳定。
发明内容
本发明的主要目的在于提供一种数据稀疏投影算法、***、电子装置及存储介质，旨在解决现有技术中投影矩阵的生成过于随机，导致投影前后样本之间的相似度特征无法得到保护，且由于每次输出结果不断变化，导致在实际应用中不稳定的技术问题。
为实现上述目的，本发明第一方面提供一种数据稀疏投影算法，包括：随机初始化一个高维稀疏二维矩阵；将所述高维稀疏二维矩阵固定，使用所述高维稀疏二维矩阵计算最优的输出变量；将最优的输出变量固定，使用最优的输出变量计算最优的高维稀疏二维矩阵；循环固定所述高维稀疏二维矩阵及输出变量，直至在固定所述高维稀疏二维矩阵时最优的输出变量不再增大，此时生 成最终的高维稀疏二维矩阵。
进一步地，所述将所述高维稀疏二维矩阵固定，使用所述高维稀疏二维矩阵计算最优的输出变量包括：根据已有的样本训练集合，所述样本包括输入向量或输入向量与输出变量，所述输入向量为高维稀疏二维矩阵；使用赢者通吃算法建立所述集合中的输入向量及输出向量之间的函数关系；将高维稀疏二维矩阵固定，根据固定的所述高维稀疏二维矩阵及所述函数关系计算最优的输出变量。
进一步地，所述将所述高维稀疏二维矩阵固定，使用所述高维稀疏二维矩阵计算最优的输出变量还包括：根据给定的约束值，对所述高维稀疏二维矩阵做稀疏性约束；根据赢者通吃算法及约束后的高维稀疏二维矩阵对输出变量进行约束。
本发明第二方面提供一种数据稀疏投影***，包括：初始化模块，用于随机初始化一个高维稀疏二维矩阵；最优输出变量计算模块，用于将所述初始化模块初始化的高维稀疏二维矩阵固定，使用所述高维稀疏二维矩阵计算最优的输出变量；最优高维稀疏二维矩阵计算模块，用于将所述最优输出变量计算模块计算出的最优的输出变量固定，使用最优的输出变量计算最优的高维稀疏二维矩阵；循环模块，用于循环固定所述高维稀疏二维矩阵及输出变量，直至在固定所述最优高维稀疏二维矩阵计算模块得到的高维稀疏二维矩阵时最优的输出变量不再增大。
进一步地，所述最优输出变量计算模块包括：集合训练单元，用于根据已有的样本训练集合，所述样本包括输入向量或输入向量与输出变量，所述输入向量为高维稀疏二维矩阵；函数关系建立单元，用于使用赢者通吃算法建立所述集合中的输入向量及输出向量之间的函数关系；计算单元，用于将高维稀疏二维矩阵固定，并根据固定的所述集合训练单元训练的集合中的高维稀疏二维矩阵及所述函数关系建立单元建立的函数关系计算最优的输出变量。
进一步地，所述最优输出变量计算模块还包括：高维稀疏二维矩阵约束单元，用于根据给定的约束值，对高维稀疏二维矩阵做稀疏性约束；输出变量约束单元，用于根据赢者通吃算法及所述高维稀疏二维矩阵约束单元约束后的高维稀疏二维矩阵对输出变量进行约束。
本发明第三方面提供一种电子装置，包括：存储器、处理器及存储在所述存储器上并可在所述处理器上运行的计算机程序，所述处理器执行所述计算机程序时，实现上述中的任意一项所述方法。
本发明第四方面提供一种计算机可读存储介质，其上存储有计算机程序，所述计算机程序被处理器执行时，实现上述中的任意一项所述方法。
本发明提供一种数据稀疏投影算法，有益效果在于：通过随机和迭代的方式不断地优化高维稀疏二维矩阵，能够使得高维稀疏二维矩阵更加稳定，对原始数据相似度特征保护更好、其稀疏性和二值性使得其在实际应用中计算速度更快。
附图说明
为了更清楚地说明本发明实施例或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本发明的一些实施例，对于本领域技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。
图1为本发明实施例数据稀疏投影算法的流程示意框图；
图2为本发明实施例数据稀疏投影***的结构示意框图；
图3为本发明实施例电子装置的结构示意框图。
具体实施方式
为使得本发明的发明目的、特征、优点能够更加的明显和易懂，下面将结合本发明实施例中的附图，对本发明实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例仅仅是本发明一部分实施例，而非全部实施例。基 于本发明中的实施例，本领域技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本发明保护的范围。
请参阅图1，为一种数据稀疏投影算法，包括：S1、随机初始化一个高维稀疏二维矩阵；S2、将高维稀疏二维矩阵固定，使用高维稀疏二维矩阵计算最优的输出变量；S3、将最优的输出变量固定，使用最优的输出变量计算最优的高维稀疏二维矩阵；S4、循环固定高维稀疏二维矩阵及输出变量，直至在固定高维稀疏二维矩阵时最优的输出变量不再增大。
具体地，在算法开始时，随机初始化一个W，并记做W 1，通过交互迭代对W和Y不断优化，Y为输出变量，具体迭代优化过程为：
对于第t次迭代，固定W记做W t同时最大化关于变量Y的函数L(W，Y)，此时获得最优的Y t；随后固定Y记做Y t同时最大化关于变量W的函数L(W，Y)，得到W t+1，重复以上步骤直至L不再增大，即函数值收敛。
将高维稀疏二维矩阵固定，使用高维稀疏二维矩阵计算最优的输出变量包括：根据已有的样本训练集合，样本包括输入向量或输入向量与输出变量，输入向量为高维稀疏二维矩阵；使用赢者通吃算法建立集合中的输入向量及输出向量之间的函数关系；将高维稀疏二维矩阵固定，并根据固定的高维稀疏二维矩阵及函数关系计算最优的输出变量。
将高维稀疏二维矩阵固定，使用高维稀疏二维矩阵计算最优的输出变量还包括：根据给定的约束值，对高维稀疏二矩阵做稀疏性约束；根据赢者通吃算法及约束后的高维稀疏二维矩阵对输出变量进行二值化约束。
在上述中，关于最优的Y t可以很容易通过
Figure PCTCN2020100618-appb-000001
求得，其中(1≤m≤n)。
其中，首先要根据给定的样本集合X＝{x .1，...，x .n}，其中每个样本x .m∈R d(1≤m≤n)。基于训练集X，我们希望获得一个高维稀疏二值矩阵W将d维输入向量{x .1，...，x .n}投影到d′维(d′远大于d)的稀疏二值向量{y .1，...，y .n}中， 并且每对输入和输出向量之间满足
Figure PCTCN2020100618-appb-000002
这望
Figure PCTCN2020100618-appb-000003
表示赢者通吃(Winner-Take-All)计算方法。即：对于1≤i≤d′，
Figure PCTCN2020100618-appb-000004
这里k表示投影的哈希长度。由于实际训练数据存在噪声，为了寻找投影矩阵W，该问题转化为关于变量W和变量Y的最大化问题，即：
Maximize：
Figure PCTCN2020100618-appb-000005
令l mij＝y im(1-y jm)(W i.x .m-w j.x .m)，当y im＝1并且y jm＝0，W i.x .m≥w j.x .m使得l mij增大，整体L(W，Y)朝着正向增加；否则，l mij＝0，对L(W，Y)没有影响，因此该问题最终就是找到理想的W和Y。对于稀疏二值投影，我们对W做稀疏性约束，对于给定的c，我们有：
w i.∈{0，1} d并且
Figure PCTCN2020100618-appb-000006
从WTA函数中，我们有关于Y的约束如下：
y .m∈{0，1} d′并且
Figure PCTCN2020100618-appb-000007
而关于最优的W t+1，有以下推导：
Figure PCTCN2020100618-appb-000008
因此最优的W t+1是
Figure PCTCN2020100618-appb-000009
对于所有的1≤i≤d，并且
Figure PCTCN2020100618-appb-000010
通过交互式迭代优化过程，我们的函数值可以获得局部最优解，当函数值L(W，Y)不再上升时收敛。
综上所述，本申请实施例提供的一种数据稀疏投影算法，生成了高质量投影矩阵，对于输入和输出都给定的样本，可直接根据上述中的方法生成最优的高维稀疏二维矩阵，适用于有监督学习。
在对于只给定输入的样本而不给定输出样本的情况下，对于通过随机高维稀疏二维矩阵和迭代的方式不断优化高维稀疏二维矩阵，使得生成的投影矩阵更稳定，对原始数据相似度特征保护更好、其稀疏性和二值性使得其在实际应用中计算速度更快，适合无监督学习。因此，本发明不仅适合于有监督学习，还适合无监督学习，而本发明中的投影矩阵是从输入数据中学习出来的，较已有方法具有更稳定、特征保护性更好、计算更快的有益效果。
请参阅图2，本申请实施例还提供一种数据稀疏投影***，包括：初始化模块1、最优输出变量计算模块2、最优高维稀疏二维矩阵计算模块3及循环模块4；初始化模块1用于随机初始化一个高维稀疏二维矩阵；最优输出变量计算模块2用于将初始化模块1初始化的高维稀疏二维矩阵固定，使用高维稀疏二维矩阵计算最优的输出变量；最优高维稀疏二维矩阵计算模块3用于将最优输出变量计算模块2计算出的最优的输出变量固定，使用最优的输出变量计算最优的高维稀疏二维矩阵；循环模块4用于循环固定高维稀疏二维矩阵及输出变量，直至固定最优输出变量计算模块2得到的输出变量时高维稀疏二维矩阵不再增大。
最优输出变量计算模块2包括：集合训练单元、函数关系建立单元及计算单元；集合训练单元用于根据已有的样本训练集合，样本包括输入向量或输入向量与输出变量，输入向量为高维稀疏二维矩阵；函数关系建立单元用于使用赢者通吃算法建立集合中的输入向量及输出向量之间的函数关系；计算单元用 于将高维稀疏二维矩阵固定，并根据固定的集合训练单元训练的集合中的高维稀疏二维矩阵及函数关系建立单元建立的函数关系计算最优的输出变量。
最优输出变量计算模块2还包括：高维稀疏二维矩阵约束单元及输出变量约束单元；高维稀疏二维矩阵约束单元用于根据给定的约束值，对高维稀疏二维矩阵做稀疏性约束；输出变量约束单元用于根据赢者通吃算法及高维稀疏二维矩阵约束单元约束后的高维稀疏二维矩阵对输出变量进行约束。
本申请实施例提供一种电子装置，请参阅3，该电子装置包括：存储器601、处理器602及存储在存储器601上并可在处理器602上运行的计算机程序，处理器602执行该计算机程序时，实现前述中描述的半导体加工模块的故障响应方法。
进一步的，该电子装置还包括：至少一个输入设备603以及至少一个输出设备604。
上述存储器601、处理器602、输入设备603以及输出设备604，通过总线605连接。
其中，输入设备603具体可为摄像头、触控面板、物理按键或者鼠标等等。输出设备604具体可为显示屏。
存储器601可以是高速随机存取记忆体(RAM，R且om Access Memory)存储器，也可为非不稳定的存储器(non-volatile memory)，例如磁盘存储器。存储器601用于存储一组可执行程序代码，处理器602与存储器601耦合。
进一步的，本申请实施例还提供了一种计算机可读存储介质，该计算机可读存储介质可以是设置于上述各实施例中的电子装置中，该计算机可读存储介质可以是前述实施例中的存储器601。该计算机可读存储介质上存储有计算机程序，该程序被处理器602执行时实现前述方法实施例中描述的半导体加工模块的故障响应方法。
进一步的，该计算机可存储介质还可以是U盘、移动硬盘、只读存储器601(ROM，Read-Only Memory)、RAM、磁碟或者光盘等各种可以存储程序代 码的介质。
在本申请所提供的几个实施例中，应该理解到，所揭露的装置和方法，可以通过其它的方式实现。例如，以上所描述的装置实施例仅仅是示意性的，例如，所述模块的划分，仅仅为一种逻辑功能划分，实际实现时可以有另外的划分方式，例如多个模块或组件可以结合或者可以集成到另一个***，或一些特征可以忽略，或不执行。另一点，所显示或讨论的相互之间的耦合或直接耦合或通信连接可以是通过一些接口，装置或模块的间接耦合或通信连接，可以是电性，机械或其它的形式。
所述作为分离部件说明的模块可以是或者也可以不是物理上分开的，作为模块显示的部件可以是或者也可以不是物理模块，即可以位于一个地方，或者也可以分布到多个网络模块上。可以根据实际的需要选择其中的部分或者全部模块来实现本实施例方案的目的。
另外，在本发明各个实施例中的各功能模块可以集成在一个处理模块中，也可以是各个模块单独物理存在，也可以两个或两个以上模块集成在一个模块中。上述集成的模块既可以采用硬件的形式实现，也可以采用软件功能模块的形式实现。
所述集成的模块如果以软件功能模块的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本发明的技术方案本质上或者说对现有技术做出贡献的部分或者该技术方案的全部或部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质中，包括若干指令用以使得一台计算机设备(可以是个人计算机，服务器，或者网络设备等)执行本发明各个实施例所述方法的全部或部分步骤。
需要说明的是，对于前述的各方法实施例，为了简便描述，故将其都表述为一系列的动作组合，但是本领域技术人员应该知悉，本发明并不受所描述的动作顺序的限制，因为依据本发明，某些步骤可以采用其它顺序或者同时进行。其次，本领域技术人员也应该知悉，说明书中所描述的实施例均属于优选实施 例，所涉及的动作和模块并不一定都是本发明所必须的。
在上述实施例中，对各个实施例的描述都各有侧重，某个实施例中没有详述的部分，可以参见其它实施例的相关描述。
以上为对本发明所提供的一种数据稀疏投影算法、***、电子装置及存储介质的描述，对于本领域的技术人员，依据本发明实施例的思想，在具体实施方式及应用范围上均会有改变之处，综上，本说明书内容不应理解为对本发明的限制。

Claims (8)

  1. 一种数据稀疏投影算法，其特征在于，包括：
    随机初始化一个高维稀疏二维矩阵；
    将所述高维稀疏二维矩阵固定，使用所述高维稀疏二维矩阵计算最优的输出变量；
    将最优的输出变量固定，使用最优的输出变量计算最优的高维稀疏二维矩阵；
    循环固定所述高维稀疏二维矩阵及输出变量，直至在固定所述高维稀疏二维矩阵时最优的输出变量不再增大，此时生成最终的高维稀疏二维矩阵。
  2. 根据权利要求1所述的数据稀疏投影算法，其特征在于，
    所述将所述高维稀疏二维矩阵固定，使用所述高维稀疏二维矩阵计算最优的输出变量包括：
    根据已有的样本训练集合，所述样本包括输入向量或输入向量与输出变量，所述输入向量为高维稀疏二维矩阵；
    使用赢者通吃算法建立所述集合中的输入向量及输出向量之间的函数关系；
    将高维稀疏二维矩阵固定，根据固定的所述高维稀疏二维矩阵及所述函数关系计算最优的输出变量。
  3. 根据权利要求2所述的数据稀疏投影算法，其特征在于，
    所述将所述高维稀疏二维矩阵固定，使用所述高维稀疏二维矩阵计算最优的输出变量还包括：
    根据给定的约束值，对所述高维稀疏二维矩阵做稀疏性约束；
    根据赢者通吃算法及约束后的高维稀疏二维矩阵对输出变量进行约束。
  4. 一种数据稀疏投影***，其特征在于，包括：
    初始化模块，用于随机初始化一个高维稀疏二维矩阵；
    最优输出变量计算模块，用于将所述初始化模块初始化的高维稀疏二维矩 阵固定，使用所述高维稀疏二维矩阵计算最优的输出变量；
    最优高维稀疏二维矩阵计算模块，用于将所述最优输出变量计算模块计算出的最优的输出变量固定，使用最优的输出变量计算最优的高维稀疏二维矩阵；
    循环模块，用于循环固定所述高维稀疏二维矩阵及输出变量，直至在固定所述最优高维稀疏二维矩阵计算模块得到的高维稀疏二维矩阵时最优的输出变量不再增大。
  5. 根据权利要求4所述的数据稀疏投影***，其特征在于，
    所述最优输出变量计算模块包括：
    集合训练单元，用于根据已有的样本训练集合，所述样本包括输入向量或输入向量与输出变量，所述输入向量为高维稀疏二维矩阵；
    函数关系建立单元，用于使用赢者通吃算法建立所述集合中的输入向量及输出向量之间的函数关系；
    计算单元，用于将高维稀疏二维矩阵固定，并根据固定的所述集合训练单元训练的集合中的高维稀疏二维矩阵及所述函数关系建立单元建立的函数关系计算最优的输出变量。
  6. 根据权利要求5所述的数据稀疏投影***，其特征在于，
    所述最优输出变量计算模块还包括：
    高维稀疏二维矩阵约束单元，用于根据给定的约束值，对高维稀疏二维矩阵做稀疏性约束；
    输出变量约束单元，用于根据赢者通吃算法及所述高维稀疏二维矩阵约束单元约束后的高维稀疏二维矩阵对输出变量进行二值化约束。
  7. 一种电子装置，包括：存储器、处理器及存储在所述存储器上并可在所述处理器上运行的计算机程序，其特征在于，所述处理器执行所述计算机程序时，实现权利要求1至3中的任意一项所述方法。
  8. 一种计算机可读存储介质，其上存储有计算机程序，其特征在于，所述计算机程序被处理器执行时，实现权利要求1至3中的任意一项所述方法。
PCT/CN2020/100618 2020-07-07 2020-07-07 一种数据稀疏投影算法、***、电子装置及存储介质 WO2022006734A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
PCT/CN2020/100618 WO2022006734A1 (zh) 2020-07-07 2020-07-07 一种数据稀疏投影算法、***、电子装置及存储介质
US18/133,523 US11853387B2 (en) 2020-07-07 2023-04-12 Method, electronic device and storage medium for data projection

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2020/100618 WO2022006734A1 (zh) 2020-07-07 2020-07-07 一种数据稀疏投影算法、***、电子装置及存储介质

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US18/133,523 Continuation US11853387B2 (en) 2020-07-07 2023-04-12 Method, electronic device and storage medium for data projection

Publications (1)

Publication Number Publication Date
WO2022006734A1 true WO2022006734A1 (zh) 2022-01-13

Family

ID=79553497

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/100618 WO2022006734A1 (zh) 2020-07-07 2020-07-07 一种数据稀疏投影算法、***、电子装置及存储介质

Country Status (2)

Country Link
US (1) US11853387B2 (zh)
WO (1) WO2022006734A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116386830A (zh) * 2023-04-10 2023-07-04 山东博鹏信息科技有限公司 一种基于大数据的医院管理***

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20150287223A1 (en) * 2014-04-04 2015-10-08 The Board Of Trustees Of The University Of Illinois Highly accelerated imaging and image reconstruction using adaptive sparsifying transforms
CN107145841A (zh) * 2017-07-20 2017-09-08 深圳大学 一种基于矩阵的低秩稀疏人脸识别方法及其***
CN109754008A (zh) * 2018-12-28 2019-05-14 上海理工大学 基于矩阵分解的高维对称稀疏网络缺失信息的估计方法
CN109948735A (zh) * 2019-04-02 2019-06-28 广东工业大学 一种多标签分类方法、***、装置及存储介质
CN110705343A (zh) * 2019-08-20 2020-01-17 西南科技大学 一种结构不相干投影字典对学习的人脸识别方法及***
CN110717519A (zh) * 2019-09-09 2020-01-21 深圳大学 训练、特征提取、分类方法、设备及存储介质

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109687843B (zh) * 2018-12-11 2022-10-18 天津工业大学 一种基于线性神经网络的稀疏二维fir陷波滤波器的设计方法
CN110796625B (zh) * 2019-10-30 2023-06-20 重庆邮电大学 一种基于组稀疏表示和加权全变分的图像压缩感知重构方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20150287223A1 (en) * 2014-04-04 2015-10-08 The Board Of Trustees Of The University Of Illinois Highly accelerated imaging and image reconstruction using adaptive sparsifying transforms
CN107145841A (zh) * 2017-07-20 2017-09-08 深圳大学 一种基于矩阵的低秩稀疏人脸识别方法及其***
CN109754008A (zh) * 2018-12-28 2019-05-14 上海理工大学 基于矩阵分解的高维对称稀疏网络缺失信息的估计方法
CN109948735A (zh) * 2019-04-02 2019-06-28 广东工业大学 一种多标签分类方法、***、装置及存储介质
CN110705343A (zh) * 2019-08-20 2020-01-17 西南科技大学 一种结构不相干投影字典对学习的人脸识别方法及***
CN110717519A (zh) * 2019-09-09 2020-01-21 深圳大学 训练、特征提取、分类方法、设备及存储介质

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116386830A (zh) * 2023-04-10 2023-07-04 山东博鹏信息科技有限公司 一种基于大数据的医院管理***
CN116386830B (zh) * 2023-04-10 2023-09-22 山东博鹏信息科技有限公司 一种基于大数据的医院管理***

Also Published As

Publication number Publication date
US11853387B2 (en) 2023-12-26
US20230259582A1 (en) 2023-08-17

Similar Documents

Publication Publication Date Title
US10372787B2 (en) Hardware accelerator pre-configured with coefficients for matrix-transform operations
Wang et al. Multiview spectral clustering via structured low-rank matrix factorization
US10482156B2 (en) Sparsity-aware hardware accelerators
WO2021218060A1 (zh) 基于深度学习的人脸识别方法及装置
US11816784B2 (en) Deep geometric model fitting
WO2019024808A1 (zh) 语义分割模型的训练方法和装置、电子设备、存储介质
US10474430B2 (en) Mixed-precision processing elements, systems, and methods for computational models
CN111144242B (zh) 一种三维目标检测方法、装置及终端
TW202030640A (zh) 一種跨模態訊息檢索方法、裝置和儲存介質
US20230401446A1 (en) Convolutional neural network pruning processing method, data processing method, and device
US11232273B2 (en) Multiplication-free approximation for neural networks and sparse coding
WO2021143267A1 (zh) 基于图像检测的细粒度分类模型处理方法、及其相关设备
CN111783033A (zh) 一种数据稀疏投影算法、***、电子装置及存储介质
TW202032423A (zh) 圖像處理方法及裝置
Ping et al. Marginal structured SVM with hidden variables
CN111435461B (zh) 使用降低精度的深度神经网络的对抗性输入识别
US20180114145A1 (en) Structured orthogonal random features for kernel-based machine learning
US10831818B2 (en) Digital image search training using aggregated digital images
Chang et al. Cross-domain kernel induction for transfer learning
EP3725084A1 (en) Deep learning on image frames to generate a summary
Huang et al. Double L2, p-norm based PCA for feature extraction
WO2022006734A1 (zh) 一种数据稀疏投影算法、***、电子装置及存储介质
WO2024060940A1 (zh) 图像处理方法、装置、电子设备和存储介质
Alom et al. State preserving extreme learning machine: A monotonically increasing learning approach
Luo et al. Adversarial nonnegative matrix factorization

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20944599

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20944599

Country of ref document: EP

Kind code of ref document: A1