WO2017191648A1 - Classificateur universel destiné à l'apprentissage et à la classification de données avec des utilisations dans l'apprentissage machine - Google Patents

Classificateur universel destiné à l'apprentissage et à la classification de données avec des utilisations dans l'apprentissage machine Download PDF

Info

Publication number
WO2017191648A1
WO2017191648A1 PCT/IN2017/000096 IN2017000096W WO2017191648A1 WO 2017191648 A1 WO2017191648 A1 WO 2017191648A1 IN 2017000096 W IN2017000096 W IN 2017000096W WO 2017191648 A1 WO2017191648 A1 WO 2017191648A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
points
point
classification
planes
Prior art date
Legal status (The legal status 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 status listed.)
Ceased
Application number
PCT/IN2017/000096
Other languages
English (en)
Inventor
Kumar ESWARAN
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Individual
Original Assignee
Individual
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 Individual filed Critical Individual
Publication of WO2017191648A1 publication Critical patent/WO2017191648A1/fr
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0499—Feedforward networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/09—Supervised learning

Definitions

  • the present invention discloses a new and novel methodology which can be used to classify specific, problems and describes a product such as the "classification engine” which is implementable in hardware for specific problems.
  • This invention is about An Universal Classifier for Learning and Classification of data with uses in Machine Learning.
  • a typical classification or pattern recognition problem involves a multi-dimensional feature space.
  • Features are variables: they can be, for example in a medical data, blood pressure, cholesterol, sugar content etc. of a patient, so each data point in feature space represents a patient.
  • Data points will be normally grouped into various clusters in feature space, each cluster will normally belong to a particular class (disease), the problem is further complicated by the fact that more than one cluster may belong to the same class (disease).
  • the problem in pattern recognition is how to make a computer recognize patterns and classify them.
  • the problem in the real world can be extremely complex as there are thousands of clusters and hundreds of classes in a space of a hundred or more dimensions.
  • Normally computers are used to detect patterns in such data and recognize classes for decision making by using Artificial Neural Network technique and an iterative algorithm called the Backpropagation technique to determine the architecture of a given problem.
  • Every classifier performs the task of partitioning of labelled points using various techniques ways e.g. statistical techniques, nearest neighbour methods, rule based methods and techniques used in "Deep Learning, all of which involve numerous computations and iterations.
  • various techniques ways e.g. statistical techniques, nearest neighbour methods, rule based methods and techniques used in "Deep Learning, all of which involve numerous computations and iterations.
  • one of the objective of the present invention is to develop the Machine Learning and Classification Engine (MLCE), which is an analytical engine that consists of a processor, devices for accepting data which may consist of image data, or, voice data or numerical data and a resident memory which contains machine coded algorithms which are executed by the CPU.
  • the MLCE learns patterns in the data so that (after the Training Phase) it may classify new incoming data, of similar nature in order to decide as to which pattern or class the present incoming data belongs to. While running the MLCE will be basically always be in one of two states viz. (i) The Learning State (Phase) and (ii) The Decision Making State (Phase).
  • Another objective is to see that, when the MLCE is in the Learning State it analyses all the input data and tries to detect patterns in the data which it associates with known data. The analysis is done using four main processes: (1) The Partition Process and (II) The Reduction of Dimension (of data) Process, (III) The Cluster Discovery Process and (IV) The Classification Process. In some cases (II) may not be required in such a case there are only three processes.
  • the analysis of incoming data is performed following the aforementioned processes and the incoming data has been "learnt" by the MLCE so that this knowledge can then be used to classify new incoming data in the Decision Making Phase.
  • this invention is an entirely new method of classification, which partitions data belonging to different categories.
  • every sample is treated as a point in feature space which has as many dimensions as the number of features (parameters).
  • the parameters belonging to a particular person such as: systolic pressure, diastolic pressure, white blood corpuscular count, red blood corpuscular count etc. are considered as distinct features and if there are say, 10 such parameters for each person then the data belonging to a particular person could be considered as a point in 10 dimensional space, whose 10 coordinates are nothing but the actual values of the said 10 features.
  • the analysis in the Learning Phase is then done in this 10 dimensional feature space.
  • the method is explained in detail in the section containing the detailed description of the invention. The method is not just confined to medical data but the MLCE can be trained to classify image and voice patterns too.
  • the system in the present innovated software-hardware system consists of a CPU (Arm Processor), which will be executing the processes indicated, viz: The Training Process which determines a Neural Architecture which then is used for The Classification and the Decision Making Process.
  • the input to the device could be data coming in various forms, either through camera feeds, audio feeds or as numerical data.
  • the incoming data is stored in memory locations during the Training stages.
  • a Neural Architecture is determined which classifies the data in the Classification Phase..
  • the MLCE can thus be used for classification of images eg. Face recognition, classification of voice data eg. Speech recognition, classification of numerical data eg. Disease diagnosis, etc. all done in real time.
  • the MLCE can be considered as powerful Intelligent Decision Making Analytical Engine which can learn as well as decide upon new data.
  • the computational complexity of the MLCE for the entire process is such that all the processes are completed in polynomial time.
  • given a data set of sufficient samples it is possible to deterministically classify the labelled data with near 100 percent accuracy (provided the data is consistent).
  • This invention is about An Universal Classifier for Learning and Classification of data with uses in Machine Learning.
  • a typical classification or pattern recognition problem involves a multi-dimensional feature space.
  • Features are variables', they can be, for example in a medical data, blood pressure, cholesterol, sugar content etc. of a patient, so each data point in feature space represents a patient.
  • Data points will be normally grouped into various clusters in feature space, each cluster will normally belong to a particular class (disease), the problem is further complicated by the feet that more than one cluster may belong to the same class (disease).
  • the problem in pattern recognition is how to make a computer recognise patterns and classify them.
  • the problem in the real world can be extremely complex as there are thousands of clusters and hundreds of classes in a space of a hundred or more dimensions.
  • Normally computers are used to detect patterns in such data and recognise classes for decision making by using Artificial Neural Network technique and an iterative algorithm called the Backpropagatipn technique to determine the architecture of a given problem.
  • This invention describes an entirely new methodology which can be used to classify specific problems and describes a product: the classification engine which is implementable in hardware for specific problems.
  • the specific problems could be: Face Recognition Systems, Disease diagnostic systems, Robotic Inspection systems for use in the factory shop floor etc.
  • the method given in this invention examines any form of data for the purposes of learning and classification so that any Machine such as a robot or machine will be able to handle the data and classify the data for automatic decision making. While analysing the data it performs , basically THREE main steps (I) The Partition Process and (II) The Reduction of Dimension (of data) Process, (III) The Cluster Discovery Process and (IV) The Classification Process
  • a medical history of a patient may consist of 50 to 100 parameters (ie 50 to 100 dimensions) and for images involving say a 30X 30 pixels involve 900 dimensions.
  • a d -dimension space will have 2 A d quadrants.
  • this Invention has provided a very ingenious method of dimension reduction, which rests on the partition algorithm described in the previous section, which will be presently described.
  • Fig. 3 shows the concept: Originally there are say N points in d-dimensional space, and suppose they are parted by using our partition by using q planes then then by taking the perpendicular distances si Si,S2 S3, s q , of any point P from the planes ( 1,2,3 q), we see that these distances serve as "coordinates" of the point P.
  • the Bubble algorithm In this algorithm examines every train point and sees its nearest neighbour, if it belongs to the same class it will include it in a sphere about the centroid. Then it searches for the next closest point, if the closest new point belongs to the same class as the other two it includes it in the same set, the centre is shifted to the centroid of 3 points. The process goes on till a closest point to the latest centroid does not belong to the same class then a sphere is drawn (about its previous centroid). By this process as shown in Fig. 4, all test data is partitioned into "pure bubbles", each bubble has points belonging to only one class. This takes approximately N 2 multiplications but is also non-iterative and will end successfully for every case.
  • step (b) The Cluster discovery algorithm. In this algorithm if two bubbles belonging to the same class are neighbours they can be combined provided they can be parted from others by a hyper plane. We search from the possible hyper planes if such a plane exists, if not we do not combine the bubbles. In the worst case each bubble is one cluster as in (b) above. Actually, step (c), is not fundamental, it only serves to reduce the number of clusters and this step (c) is not strictly necessary, so we will not talk anymore about it. For our purposes the Bubble Algorithm will serve as our "Cluster Discovery" Algorithm, though it may give many numbers of bubbles (spherical clusters).
  • the classification can be done by two alternative techniques viz. method I and method 2 which are described in the following:
  • Method 1 This method only uses the partition result (a) above, and is a kind of nearest k-neighbour algorithm but the search for the nearest neighbour is restricted to a very small subset of the train point. Whereas in the usual k-nearest neighbour method the search for the nearest point is exhaustive and is carried through the entire train data set involving many more computations of nearest Euclidean distance. This subset contains those train points whose Orientation Vectors differ from that of the test point by only 6-8 components. That is we search only those train points which are parted from the test point by only 6-8 planes. We then find that point which is closest (Euclidean distance) to the test point to classify the latter.
  • Method 2 This method finds those planes which part the bubbles from those found in 1(a) above and if needed add more planes. The process may require "disabling” some planes which pass through a particular candidate bubble. We finally end up with a smaller set of planes which can part each bubble and hence the
  • the procedure for software implementation as shown in Fig. 6A is as follows, when data is first given, it is analysed by several algorithms and three main process are executed: (I) The Partition Process and (II) The Reduction of Dimension (of data) Process, (III) The Cluster Discovery Process and (IV) The Classification Process.
  • MLCE Machine Learning and Classification Engine
  • MLCE Machine Learning and Classification Engine
  • the algorithm will then be used to detect the person (ie classify different people).
  • the output can come as a detected image or as an alarm.
  • the hardware for this will be an embedded program working on a suitable process (such as ARM).
  • the implementation of an MLCE, for this particular application will be as follows: A number of images of various personnel will be taken by cameras and stored as "Train Data", along with this collection a number of "Test Data”, would also be stored.
  • the 4 processes starting from the Partition Process and ending with the Classification Process will be implemented using the Train Data.
  • the Neural Architecture which will be tested using the Test Data.
  • the Neural Architecture will be implemented in Hardware using a suitable processor, say an Arm Processor.
  • This Hardware implementation will consist of a camera which takes images of people entering say a factory (or protected area) and these input images will be sent to the Arm processor which will use the a Neural Architecture to classify the input image and provide an output on a screen or and initiate an alarm system in case of un-authorised entry.
  • the input can also be voice data from microphones in case we wish to use the MLCE for voice identification or speech recognition.
  • the MLCE need to be trained separately for each of the different applications viz. face recognition or voice recognition.
  • the Fig. 6B shows a schematic of the Hardware implementation of the "Machine Learning and Classification Engine” (MLCE).
  • the MLE consists of a Central Processing Unit which could be an ARM processor, several input devices such as camera feeds through CCTV cameras, voice lines from microphones, data buses from other devices or computers, or from remote wireless lines etc, all these are not shown but may exist.
  • the MLCE function in two modes, that is it will be in two states (i) the Learning State and (ii) the decision making state. While it is in te learning state it acquires input from the input devices such as camera feeds or microphones.
  • the CPU will be running the classification process to learn the input images (say faces for recognition), all the inputs are learnt and classified by the classification process.
  • the output of this process is a Neural Network Architecture (the details for each layer of processing elements, their weights etc), which is stored and then used for the Decision making process to recognize the images that arrive in real time.
  • the input is analysed by the CPU which then uses the Neural Network Architecture to recognize each of the input data which arrive in real time (say faces of people) image, and then a decision is made either to sound an alarm/ accept entry of the said person or simply record the data.
  • Our task is to determine the equations of the planes that can part the points in such a way that every point is parted from another by at least one plane.
  • Orientation Vector Each component of the Orientation Vector of a point P in X space, is on the positive side or negative side of each plane.
  • the 'Orientation Vector' is defined in the next section, below.
  • Orientation Vector Suppose we have a point P in n-dimension space (also called X-space) and suppose we are given 3-planes. we define a Orientation Vector as a Hamming vector whose components precisely specify on which side the point P lies with respect to the 3 planes.
  • Q Space or Hamming space We will also call the space spanned by the Orientation Vector as Hamming Space or Q space. Since each point P in X-space has an Orientation (Hamming) Vector associated with it, we can imagine that all the points in X space are mapped to a point in Hamming space. Of course this mapping is many to one, but the fact to notice is the following: Let P be some point in X-Space, then all points, R, in X space which are not parted from P by planes will all have the same Orientation Vector as P i.e.
  • a Saturated Plane We consider a plane in n dimension space to be "saturated” if it has already been constrained to pass through n points and hence cannot be adjusted to pass through a new point, the coefficients of such a plane are completely determinable. Eg. A plane in 3 dimension gets saturated if it is made to pass througn three points.
  • a set Set S which will contain a list of N points and q planes.
  • S will contain the Orientation Vectors of all the points in S and the identification numbers (labels) for the planes, along with the coeDcients which define each plane.
  • S contains an array V for storing Orientation Vectors; and 'Counter': An integer number.
  • Another Set T which contains points and their Orientation Vectors and an array M :"List of Midpoints" which are the coordinates of the mid points of certain line segments.
  • this set T will contain pairs of points one not in S and the other in S. We are sure that each such point has only one such neighbor (in S), because the points in S are always part from one another and hence two points in the same quadrant cannot both belong to S.
  • Each such point stored in T will also store the coordinate of the point which is the midpoint of the line joining it to its neighbor and the line segment with its neighbor.
  • the points in T are the candidates waiting to be put in S but they first have to be parted from all the points in S.
  • Step 1 Initially collect a small number of initial points numbering No and choose a set of qo planes that part each one of these No points and put these planes in S.
  • T which will contain points which cannot become immediate members of S but are prospective members and will become members of S eventually.
  • Set T will contain points which are neighbors of a point which already is a member of S.
  • Step 2 If no more points in G go to step 7, else: Randomly choose a new point from G, which could be a candidate point to be put in S, from the remaining points not in S, go to Step 3.
  • Step 4 fYou will come here only if the current point has a neighbor in S. Notation and procedure for this Step: We keep count of the number of members of S which have first neighbors. Such points are called a;, its first neighbor will be called bi, if a has a second neighbour this will be called Cj and the third will be called dj .) First: Find the 'distance' of this new point from its neighbor, call
  • the coefficients of the plane containing the n midpoints can be found by using Gauss elimination, Gram-Schmidt evaluation or by using the QR algorithm, the last is more useful just in case the n mid points fall in a plane of n-1 dimension or less, then the rank of the coeDcient matrix will become less than n. This will happen rarely, and even if it does, it only means we can accommodate another oint (or points) in T with a neighbor in S, thus making the rank n. In such a case all the points in T (even though they are now more than n) can be parted by a single plane - the algorithm can proceed.
  • step 5 we are just collecting points which are separable and putting them S or in case they have a neighbor in S then we are putting the points in T. This can go on till there are n points in T and we arrive at Step 5.
  • step 5 and Step 6 we will have an enhanced set S with more points and one more plane and all of which are parted by these q + 1 planes. So we can call (q + 1) as q and then re-do steps 2 to 6 till all points N f in G are exhausted. Then S will contain all the points and all the planes which part them.
  • Step 7 will happen in the end when all the points are exhausted but there are less than n points in T and these must be parted.
  • the logic of Step 7 is similar to step 6.
  • Step 7 can be completely avoided.
  • All the Q's which are chosen all belong to a sequence which has an "accumulation" point. That is the remaining points in G which are left after all the others are parted are only those points which tend to be close to some unknown accumulation point (or limit point) close to P. This will happen since we are treating X-Space as real space and all the coordinates (xi; x 2 ; x 3 ; ::; x n ) are either rational numbers (or even real numbers which are uncountable as opposed to integers which are countable). And then we are trying to part an infinity of points or uncountable points in a sequence by a countable (or finite) set of planes an impossibility!
  • the input set G may be points generated by another computer program.
  • N f Total number of points
  • q f Total number of planes used
  • n Dimension of X-Space.
  • Training Set belongs to some class which is known.
  • the points in the Test Set are assumed to be belonging to some unknown class which has to be discovered by the classifier.
  • the Orientation Vector (OV) of each of the N points is found.
  • the OV is a Hamming vector, of dimension q, which is associated with each point P belonging to the Training Set. If P is a point in the Training Set than its orientation vector Ov(P ) indicates the "Orientation" of the point P w.r.t. each of the q planes See the Figure above.
  • each point N represents one d- dimensional sample in X-Space.
  • This sample, data point will have n numbers which may relate to a medical data of one person or alternatively it could be an image involving d pixels and represent a photograph of a person.
  • the idea is simple: after we had run the algorithm which separates each of the N data points (samples) by q number of planes, we will have the exact information of how each of the data points N reside in X-Space with respect to each other and the separation planes, because we " have the Orientation Vector for each point stored in S.
  • our JStorage Plan_ for each point L in is to use the Hamming Vector which is the Orientation Vector Ov(L) of a point L as its label (like a binary label).
  • This label is like a pointer to the information about L stored in a receptacle in the space next to this label of L for easy retrieval.
  • Method 2 Classification by using a Cluster Discovery algorithm in addition to the partitioning of points method.
  • the Bubble Discovery Algorithm This is a simple algorithm for obtaining "pure bubbles" from data
  • the meaning of a pure bubble is a spherical region containing only points belonging to one class, if the collection of input points are given then the algorithm finds the "bubbles" which separate points belonging to one class from points belonging to a different class.
  • Figure 8 shows a set of data points are given in G belonging to various classes.
  • Step 1 Find the distance from each point ' p ' from the set G to all other points in the set G. If set G has V points, then distance-matrix of size n x n is obtained.
  • Step 3 In each rowi, count the first ' m ' number of consecutive points, whose class is same as the class of point ' i ' .
  • rowi [di l ; di4; di7 ; di9; di2].
  • Points i, 1 , 4, and 7 belongs to class ' ⁇
  • point '9' belongs to class ' 3 '
  • point '2' belongs to class '2 ' . So make count of rowi as four.
  • Step 4 (i). Pick the rowi , whose count is highest
  • This algorithm finds the bubbles (spherical clusters) of various sizes each bubble having points belonging to the same class and a known radius.
  • the bubbles have been portioned y planes, and hence by following the methods of the previous invention (May 2015)we can immediately write down the Neural Architecture that classifies the above configuration, the neural Architecture is shown below and the weights of each processing elements are known because the equations of the portioning planes are known. Since there are five partitioning planes as shown in Fig. I I, the architecture contains five processing elements in the first layer, and since there are eight pure bubbles there are eight processing elements in the second layer and since there are only three classes there are three processing elements in the last layer. The above architecture completely solves the given classification problem. Note it has been obtained in a completely deterministic and non-iterative manner.
  • the first example (a) describes the application of the computer program that was developed for the portioning of points, the results are discussed briefly.
  • the second example (b), is the application of the entire portioning and classification process on the classical IRIS problem. We show how the computer program using the algorithms described in this invention automatically determines in a non-iterative manner the neural architecture to solve the IRIS problem.
  • This program first generated 2000 random points inside a 15 dimensioned unit cube and then applies the invention to show that only 22 hyper planes can separate each of the 2000 points.
  • the coefficients of the 22 hyperplanes are obtained by the program and the Orientation Vectors for each of the 2000 points are obtained and shown to be unique.
  • the 22 hyperplanes actually partition all the 2000 points in such a manner that every point is separated from another and if one chooses any pair of points there will be at least one hyper plane separating them. (The time taken to run the program was almost instantaneous in a Toshiba laptop).
  • the computer program was then successfully used for a larger sized problem involving 50,000 points which were similarly randomly generated this time in a 25 dimension space they could be separated by only 27 planes. The time taken was approximately 6 min 25 sees on a desktop.
  • the data was separated into two sets (a) the training set and (b) the test set
  • the test set was obtained from the IRIS data by choosing every 5th flower, so there were 120 flowers left in the training set and 29 in test set (it was discovered that two flowers had the same dimensions so one of them had to be removed, leaving 149 samples).
  • the classification system was built in the following steps:
  • Step 1 Use the separation algorithm to separate the 120 samples each of which can be represented by a point in 4-dimension space, by planes.
  • Step 2 Run the algorithm which transfers points from initial set G to set S such that each point is in its own 'quadrant' drawing planes whenever necessary. It was found that 24 planes separate all the 120 points.
  • Step 3 By the means of the coefficients of the 24 planes the classification system shown in the figure can be built, this system basically finds the OV's of any incoming 4-dimensional point P. In our case there will be 24 processing elements El ; E2 E24 in the first layer because there are 24 planes.
  • Bubble Discovery al gorithm (Method 2) was applied for the same IRIS Flower data set [14-15] (see Appendix for data). In this case, it was not found necessary to introduce more planes to separate the bubbles, 24 planes were sufficient. And the final number of bubbles was 25.
  • the given train data was tested on the neural network shown in the figure, by first applying it on 120 training points. All 120 points were correctly classified. Then the test data was passed through the neural network i.e, each of the 29 test points was given as input to the NN. Out of 29 points, two points (point nos. 85 and 120) were wrongly classified. All the other 27 points were correctly classified. 9. Additional Example Problems: Digit Recognition and Face Recognition
  • this algorithm was recently applied on the digit recognition data set MNIST and the Face recognition data set viz. the Extended Yale dataset In MNIST, the original dimensions of each sample was 28 X 28. It was reduced to 6X6 dimensions. It was found that it took 65 planes to separate the 60,000 training points and the time taken (on a Lap Top) was 4.8 minutes and the time taken to classify the 10,000 test points was 3.9 minutes. The classification accuracy was 92.2 %. In Yale dataset, the original dimensions of the each sample was 30 X 30. To separate approx.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Physics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Image Analysis (AREA)

Abstract

La présente invention concerne une nouvelle méthodologie qui peut être utilisée afin de classer des problèmes spécifiques et décrit un produit tel que le "moteur de classification" qui peut être mis en œuvre dans un matériel pour des problèmes spécifiques. Le problème spécifique pourrait être tel que des systèmes de reconnaissance de visage, des systèmes de diagnostic de maladie, des systèmes d'inspection robotiques pour une utilisation dans l'usine de fabrication, etc. Le système et le procédé décrits dans cette invention examinent une quelconque forme de données à des fins d'apprentissage et de classification de telle sorte qu'une quelconque machine telle qu'un robot ou une machine sera apte à gérer les données et à classer les données pour une prise de décision automatique. Pendant l'analyse des données, la présente invention réalise essentiellement TROIS étapes principales (I) le processus de division et (II) la réduction de la taille (des données), (III) le processus de découverte de grappe et (IV) le processus de classification.
PCT/IN2017/000096 2016-05-05 2017-04-28 Classificateur universel destiné à l'apprentissage et à la classification de données avec des utilisations dans l'apprentissage machine Ceased WO2017191648A1 (fr)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
IN201641015775 2016-05-05
IN201641015775 2016-05-05

Publications (1)

Publication Number Publication Date
WO2017191648A1 true WO2017191648A1 (fr) 2017-11-09

Family

ID=60202871

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/IN2017/000096 Ceased WO2017191648A1 (fr) 2016-05-05 2017-04-28 Classificateur universel destiné à l'apprentissage et à la classification de données avec des utilisations dans l'apprentissage machine

Country Status (1)

Country Link
WO (1) WO2017191648A1 (fr)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108490959A (zh) * 2018-05-22 2018-09-04 国网天津市电力公司 一种支持深度学习工作原理的人工智能机房巡检机器人
CN111401444A (zh) * 2020-03-16 2020-07-10 深圳海关食品检验检疫技术中心 红酒原产地的预测方法、装置、计算机设备及存储介质
WO2020247810A1 (fr) * 2019-06-06 2020-12-10 Home Depot International, Inc. Optimisation des données de formation destinée à une classification d'images
CN115620144A (zh) * 2022-10-27 2023-01-17 北京师范大学 一种特征变量优选的面向对象不透水面材质提取方法
CN117274727A (zh) * 2023-10-25 2023-12-22 荣耀终端有限公司 生物特征信息的处理方法、电子设备及可读存储介质
CN117532593A (zh) * 2023-12-20 2024-02-09 哈尔滨工业大学 一种基于元模块互助的模块化机器人自重构方法

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20060074824A1 (en) * 2002-08-22 2006-04-06 Jinyan Li Prediction by collective likelihood from emerging patterns
US7792770B1 (en) * 2007-08-24 2010-09-07 Louisiana Tech Research Foundation; A Division Of Louisiana Tech University Foundation, Inc. Method to indentify anomalous data using cascaded K-Means clustering and an ID3 decision tree

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20060074824A1 (en) * 2002-08-22 2006-04-06 Jinyan Li Prediction by collective likelihood from emerging patterns
US7792770B1 (en) * 2007-08-24 2010-09-07 Louisiana Tech Research Foundation; A Division Of Louisiana Tech University Foundation, Inc. Method to indentify anomalous data using cascaded K-Means clustering and an ID3 decision tree

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108490959A (zh) * 2018-05-22 2018-09-04 国网天津市电力公司 一种支持深度学习工作原理的人工智能机房巡检机器人
WO2020247810A1 (fr) * 2019-06-06 2020-12-10 Home Depot International, Inc. Optimisation des données de formation destinée à une classification d'images
US11687841B2 (en) 2019-06-06 2023-06-27 Home Depot Product Authority, Llc Optimizing training data for image classification
US11954572B2 (en) 2019-06-06 2024-04-09 Home Depot Product Authority, Llc Optimizing training data for image classification
CN111401444A (zh) * 2020-03-16 2020-07-10 深圳海关食品检验检疫技术中心 红酒原产地的预测方法、装置、计算机设备及存储介质
CN111401444B (zh) * 2020-03-16 2023-11-03 深圳海关食品检验检疫技术中心 红酒原产地的预测方法、装置、计算机设备及存储介质
CN115620144A (zh) * 2022-10-27 2023-01-17 北京师范大学 一种特征变量优选的面向对象不透水面材质提取方法
CN117274727A (zh) * 2023-10-25 2023-12-22 荣耀终端有限公司 生物特征信息的处理方法、电子设备及可读存储介质
CN117274727B (zh) * 2023-10-25 2024-04-12 荣耀终端有限公司 生物特征信息的处理方法、电子设备及可读存储介质
CN117532593A (zh) * 2023-12-20 2024-02-09 哈尔滨工业大学 一种基于元模块互助的模块化机器人自重构方法

Similar Documents

Publication Publication Date Title
WO2017191648A1 (fr) Classificateur universel destiné à l'apprentissage et à la classification de données avec des utilisations dans l'apprentissage machine
dos Santos et al. Improving Pantanal fish species recognition through taxonomic ranks in convolutional neural networks
KR102092263B1 (ko) 일정한 처리 시간 내에 k개의 극값을 찾는 방법
CN108595585B (zh) 样本数据分类方法、模型训练方法、电子设备及存储介质
CN106358444B (zh) 用于面部验证的方法和系统
Filonenko et al. Comparative study of modern convolutional neural networks for smoke detection on image data
Bakr et al. DenseNet based model for plant diseases diagnosis
JP5992073B2 (ja) 情報絞り込み検出機能を備えたメモリ、このメモリを用いた情報検出方法、このメモリを含む装置、情報の検出方法、メモリの使用方法、およびメモリアドレス比較回路
JP5944406B2 (ja) 最近傍を見つけ出すための方法および装置
CN110765860A (zh) 摔倒判定方法、装置、计算机设备及存储介质
Qiu et al. A parallel neuromorphic text recognition system and its implementation on a heterogeneous high-performance computing cluster
CN113569081B (zh) 图像识别方法、装置、设备及存储介质
WO2017149559A1 (fr) Un système de traitement d'information, un procédé de traitement d'information et un support de stockage lisible par ordinateur
Shoohi et al. DCGAN for Handling Imbalanced Malaria Dataset based on Over-Sampling Technique and using CNN.
Termritthikun et al. NU-LiteNet: Mobile landmark recognition using convolutional neural networks
CN115222061A (zh) 基于持续学习的联邦学习方法以及相关设备
CN118608823A (zh) 基于改进yolov8的无序零件检测方法及系统
CN117011577A (zh) 图像分类方法、装置、计算机设备和存储介质
CN120726421A (zh) 基于噪声标签的多模态识别方法、装置、设备、存储介质及程序产品
Atique et al. Identification of plant species through leaf vein morphometric and deep learning
Laouarem et al. A deep learning model for CXR-based COVID-19 detection
CN118035478B (zh) 跨设备采集图像的检索排序方法、装置、设备及介质
Lavanya et al. Classification of underwater fish species using custom-built deep learning architectures
CN118733806A (zh) 一种相似图像检索方法及装置
WO2022003981A1 (fr) Dispositif, procédé et programme de spécification d'action

Legal Events

Date Code Title Description
NENP Non-entry into the national phase

Ref country code: DE

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

Ref document number: 17792619

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 17792619

Country of ref document: EP

Kind code of ref document: A1