公开(公告)号：EP3301617A1

公开(公告)日：2018-04-04

申请号：EP17306310.8

申请日：2017-10-02

申请人： Safran Identity & Security

发明人： CHABANNE, Hervé ,  MILGRAM, Jonathan ,  MOREL, Constance ,  PROUFF, Emmanuel

IPC分类号： G06N3/04

CPC分类号： G06N3/08 ,  G06F17/18 ,  G06N3/0454 ,  G06N3/0481 ,  H04L9/008

摘要： La présente invention concerne un procédé d'apprentissage sécurisé de paramètres d'un réseau de neurones à convolution, CNN, pour classification de données ;
le procédé comprenant la mise en oeuvre par des moyens de traitement de données (11a) d'un premier serveur (1a), d'étapes de :
(a0) Réception depuis un deuxième serveur (1b) d'une base de données d'apprentissage déjà classifiées, lesdites données d'apprentissage étant chiffrées de façon homomorphique ;
(a1) Apprentissage dans le domaine chiffré, à partir de ladite base de données d'apprentissage, des paramètres d'un CNN de référence comprenant au moins :
- une couche non-linéaire (POLYNOME) opérant une fonction polynomiale de degré au moins deux approximant une fonction d'activation ;
- une couche de normalisation en batch (BN) avant chaque couche non-linéaire (POLYNOME) ;

(a2) Transmission audit deuxième serveur (1b) des paramètres appris, pour déchiffrement et utilisation en classification
La présente invention concerne également des procédés de classification sécurisée d'une donnée d'entrée.

摘要翻译： 本发明涉及用于学习卷积神经元网络CNN的参数以用于数据分类的安全方法; 该方法包括：由数据处理装置（11A）的第一服务器（1a）中，实施步骤：（a0）的从第二服务器接收（1b）中的一个数据库的 学习已被分类，所述学习数据被同态加密; （A1）在学习的加密区域，从训练数据，包括至少一个参考CNN的参数的所述基： - 操作程度的至少两个的多项式函数的非线性层（POLYNOME） 近似激活函数; 在每个非线性层（POLYNOME）之前的批量标准化层（BN）; 本发明还涉及对输入数据进行安全分类的方法。

公开(公告)号：EP2994856A1

公开(公告)日：2016-03-16

申请号：EP13724774.8

申请日：2013-05-10

申请人： Prysmian S.p.A.

发明人： DI STEFANO, Antonio ,  CANDELA, Roberto ,  FISCELLI, Giuseppe ,  GIACONIA, Giuseppe, Costantino

IPC分类号： G06N3/10

CPC分类号： G01R31/02 ,  G01R31/1245 ,  G01R31/1272 ,  G01R31/14 ,  G06N3/04 ,  G06N3/0481 ,  G06N3/08 ,  G06N3/084 ,  G06N3/10

摘要： A partial discharge signals processing method includes: setting a first discrimination criterion among the following criteria: discharge signals acquisition, discharge signals noise filtering, and discharge signals classification; providing a plurality of pulse waveforms associated with detected partial discharge waveform signals; defining at least a first reference pulse waveform in accordance with the first criterion; performing a first training of a neural network module based on the at least a first reference pulse waveform to produce a similarity index adapted to selectively assume a first value and a second value representative of a similarity/non similarity of an input pulse waveform with the at least a first reference pulse waveform, respectively; comparing the plurality of pulse waveforms with the at least a first reference pulse waveform by means of the neural network module to obtain first similarity index values; and memorizing/rejecting each compared pulse waveform on the basis of the obtained first similarity index values and on the second discrimination criterion.

公开(公告)号：EP2533176A1

公开(公告)日：2012-12-12

申请号：EP12183557.3

申请日：2006-11-15

申请人： Garner, Bernadette

发明人： Garner, Bernadette

IPC分类号： G06N3/08

CPC分类号： G06N3/082 ,  G06N3/0481 ,  G06N3/088

摘要： The present invention provides a method for determining whether an input vector is known or unknown by a neuron. The method includes the steps of: constructing a constraint and its complement from the input vector; alternately adding the constraint and its complement to the constraints set of the neuron; and testing the constraints set to determine if there is a solution in either case. If there is a solution for either the constraint or its complement, but not both, it is determined that the input vector is known by the neuron, and if there is a solution when both the constraint and its complement are alternately added to the constraints set, it is determined that the input vector is not known by the neuron.

摘要翻译： 本发明提供了是否输入矢量是已知的或由神经元未知确定性采矿的方法。 该方法包括以下步骤：构造一约束，并从输入向量及其互补序列; 或者添加约束和补充设定神经元的限制; 如果有测试设置为确定性矿的约束在两种情况下的解决方案。 如果存在用于无论是约束或它的互补物，但不是两者的溶液中，它是确定性的开采并输入矢量是由神经元已知的，并且如果有一个解决方案时，这两个约束及其互补交替加入到设置的约束 时，确定性的开采并输入矢量不被神经元已知的。

公开(公告)号：EP1949313A4

公开(公告)日：2010-03-31

申请号：EP06804525

申请日：2006-11-15

申请人： GARNER BERNADETTE

发明人： GARNER BERNADETTE

IPC分类号： G06N3/08

CPC分类号： G06N3/082 ,  G06N3/0481 ,  G06N3/088

公开(公告)号：EP1444649A1

公开(公告)日：2004-08-11

申请号：EP02761440.3

申请日：2002-08-20

申请人： Exscientia, LLC

发明人： HAMPSHIRE, John, B., II

IPC分类号： G06N3/02

CPC分类号： G06K9/628 ,  G06K9/6262 ,  G06K9/6268 ,  G06N3/0481 ,  G06N3/08

摘要： An apparatus and method for training a neural network model (21) to classify patterns (26) or to assess the value of decisions associated with patterns by comprising the actual output of the network in response to an input pattern with the desired output for that pattern on the basis of a Risk Differential Learning (RDL) objective function (28), the results of the comparison governing adjustment of the neural network model's parameters by numerical optimization. The RDL objective function includes one or more terms, each being a risk/benefit/classification figure-of-merit (RBCFM) function, which is a synthetic, monotonically non-decreasing, anti-symmetric/asymmetric, piecewise-differentiable function of a risk differential (Fig. 6), which is the difference between outputs of the neural network model produced in response to a given input pattern. Each RBCFM function has mathematical attributes such that RDL can make universal guarantees of maximum correctness/profitability and minimum complexity. A strategy for profit-maximizing resource allocation utilizing RDL is also disclosed.

公开(公告)号：EP1014274A1

公开(公告)日：2000-06-28

申请号：EP98965909.9

申请日：1998-12-31

申请人： Joint-Stock Company Research Centre "Module"

发明人： CHERNIKOV, Vladimir Mikhailovich ,  VIKSNE, Pavel Evgenievich ,  FOMIN, Dmitry Viktorovich ,  SHEVCHENKO, Pavel Alexandrovich ,  YAFRAKOV, Mikhail Fedorovich

IPC分类号： G06F15/18 ,  G06F17/16 ,  G06T1/40 ,  G06F7/38 ,  G06F7/50 ,  G06F7/52 ,  G06F7/544

CPC分类号： G06N3/063 ,  G06N3/0481

摘要： The group of the inventions relates to the field of computer science and can be used for neural network emulation and digital signal processing. Increasing of the neural processor performance is achieved by means of the ability to change word lengths of results in program mode. The neural processor comprises six registers, a shift register, a AND gate, two FIFOs, a switch, a multiplexer, two saturation units, a calculation unit and a adder circuit to execute operations over vectors of programmable word length data. Increasing of the saturation unit performance is achieved by means of the ability to process vector of input operands with programmable word length at a time. Said unit comprises a carry look-ahead circuit and a carry propagation circuit, and also by two multiplexers, one EXCLUSIVE OR gate, one EQUIVALENCE gate, one NAND gate and one AND gate with inverted input in each bit. Functionality of the calculation unit is expanded. The calculation unit comprises a delay element N/2 AND gates with inverted input N/2 decoders of multiplier bits, a N-bit shift register, which each bit consists of a AND gate with inverted inputs, a multiplexer and a trigger, and a multiplier array, comprising N columns by N/2 cells, each of them consists of two triggers, a AND gate with inverted input, an one-bit partial product generation circuit, an one-bit adder and a multiplexer. Increasing of the adder circuit performance is achieved by means of ability to sum two vectors of input operands of programmable word lengths. The adder circuit comprises a carry look-ahead circuit, and also by two AND gates with inverted input, one half-adder and one EXCLUSIVE OR gate in each bit.

摘要翻译： 本发明涉及计算机科学领域，可用于神经网络仿真和数字信号处理。 通过在程序模式下改变结果字长的能力来实现神经处理器性能的提高。 神经处理器包括六个寄存器，移位寄存器，与门，两个FIFO，开关，多路复用器，两个饱和单元，计算单元和加法器电路，以对可编程字长数据的向量执行操作。 通过能够一次处理可编程字长的输入操作数向量来实现增加饱和度单位性能。 所述单元包括进位查询电路和进位传播电路，并且还通过两个复用器，每个位中具有反相输入的一个独占或门，一个等效门，一个与非门和一个与门。 计算单位的功能扩大。 计算单元包括具有乘法器位的反相输入N / 2解码器的延迟元件N / 2与门，每个位由具有反相输入的与门组成的N位移位寄存器，复用器和触发器，以及 乘法器阵列，由N / 2个单元组成的N列，它们分别由两个触发器，一个反相输入的与门，一位部分乘积生成电路，一位加法器和多路复用器组成。 通过对可编程字长的输入操作数的两个向量求和来实现加法器电路性能的增加。 加法器电路包括进位查找电路，并且还具有每个位中具有反相输入，一个半加法器和一个异或或门的两个与门。

公开(公告)号：EP0733982B1

公开(公告)日：1999-06-02

申请号：EP96104445.0

申请日：1996-03-20

申请人： CSELT Centro Studi e Laboratori Telecomunicazioni S.p.A.

发明人： Albesano, Dario ,  Gemello, Roberto ,  Mana, Franco

IPC分类号： G06F15/80

CPC分类号： G06N3/0481 ,  G06N3/063

公开(公告)号：EP0664516A3

公开(公告)日：1995-12-20

申请号：EP95100620.4

申请日：1995-01-18

申请人： NIPPON TELEGRAPH AND TELEPHONE CORPORATION

发明人： Aihara, Kimihisa ,  Uchimura, Kuniharu

IPC分类号： G06F15/80

CPC分类号： G06N3/063 ,  G06N3/0481

摘要： A neural network circuit and a processing scheme using the neural network circuit in which a synapse calculation for each input value and a corresponding synapse weight of each input value which are expressed by binary bit sequences is carried out by using a sequentially specified bit of the corresponding synapse weight, a summation calculation for sequentially summing synapse calculation results for the input values is carried out to obtain a summation value, a prescribed nonlinear processing is applied to the obtained summation value so as to determine the output value, whether the obtained summation value reached to a saturation region of a transfer characteristic of the prescribed nonlinear processing is judged, the synapse calculation and the summation calculation are controlled to sequentially carry out the synapse calculation from upper bits of the corresponding synapse weight, and to stop the synapse calculation and the summation calculation whenever it is judged that the obtained summation value reached to the saturation region.

公开(公告)号：EP0621549A3

公开(公告)日：1995-12-20

申请号：EP94104487.7

申请日：1994-03-22

申请人： FUJI XEROX CO., LTD.

发明人： Murai, Kazumasa, c/o Fuji Xerox Co., Ltd.

IPC分类号： G06F15/80 ,  G06G7/60

CPC分类号： G06N3/0481 ,  G06N3/063

摘要： A nonlinear operation unit includes nonlinear function operation means for receiving at least one input signal and performing the computing operation of the input signal by using a nonlinear function, multiplying means for multiplying the function value as the results of the computing operation by the nonlinear function operation means by a weight value, and adder means for adding together the results of the multiplying operations by the multiplying means and adding a threshold value to the sum.

公开(公告)号：EP0610805A3

公开(公告)日：1995-06-07

申请号：EP94101600.8

申请日：1994-02-03

申请人： E.I. DU PONT DE NEMOURS AND COMPANY

发明人： Stafford, Richard Gordon ,  Mickewich, Daniel James ,  Beutel, Jacob

IPC分类号： G06F15/42

CPC分类号： G06T7/0012 ,  G06F19/00 ,  G06F19/321 ,  G06K9/3233 ,  G06K9/6284 ,  G06K2209/05 ,  G06N3/0454 ,  G06N3/0481 ,  G06T2207/30068 ,  G16H15/00 ,  G16H50/20 ,  Y10S128/925 ,  Y10S706/924

摘要： While some of the characteristics of medical pathologies or industrial defects in images are subject to classification, many aspects of these abnormalities do not lend themselves to precise programmable characterizations. Neural networks have been useful in recognizing patterns in a number of applications involving multiple variables whose precise interactions are not well-understood or quantifiable. In order to aid the medical imaging personnel, industrial troubleshooters and others in locating potential anomalies in an image, this invention employs neural network analysis to help identify regions of suspected anomalies.