WO2020151108A1 - 基于情境信号类前额叶网络的信息处理方法、系统、装置 - Google Patents

基于情境信号类前额叶网络的信息处理方法、系统、装置 Download PDF

Info

Publication number
WO2020151108A1
WO2020151108A1 PCT/CN2019/083356 CN2019083356W WO2020151108A1 WO 2020151108 A1 WO2020151108 A1 WO 2020151108A1 CN 2019083356 W CN2019083356 W CN 2019083356W WO 2020151108 A1 WO2020151108 A1 WO 2020151108A1
Authority
WO
WIPO (PCT)
Prior art keywords
feature vector
context
information
prefrontal
network
Prior art date
Application number
PCT/CN2019/083356
Other languages
English (en)
French (fr)
Inventor
曾冠雄
陈阳
余山
Original Assignee
中国科学院自动化研究所
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 中国科学院自动化研究所 filed Critical 中国科学院自动化研究所
Priority to US16/971,691 priority Critical patent/US10915815B1/en
Publication of WO2020151108A1 publication Critical patent/WO2020151108A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/211—Selection of the most significant subset of features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G06F18/24133—Distances to prototypes
    • G06F18/24143—Distances to neighbourhood prototypes, e.g. restricted Coulomb energy networks [RCEN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
    • G06V10/449—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
    • G06V10/451—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters with interaction between the filter responses, e.g. cortical complex cells
    • G06V10/454—Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/768—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using context analysis, e.g. recognition aided by known co-occurring patterns
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/172—Classification, e.g. identification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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
    • G06N3/045—Combinations of networks

Definitions

  • the invention belongs to the field of pattern recognition and brain-like machine learning, and specifically relates to an information processing method, system and device based on a situational signal-like prefrontal network.
  • the prefrontal lobe is highly refined in primates and plays a key role in achieving this ability.
  • the prefrontal lobe can quickly learn the "rules of the game” and dynamically apply them to map sensory inputs to context-sensitive tasks based on different actions. This process is called cognitive control, which allows primates to Appropriate behavior is possible under unlimited circumstances.
  • the current artificial neural networks are very powerful in extracting advanced features from raw data for pattern classification and learning complex mapping rules. However, their response is mainly determined by the network input and presents a rigid input and output mapping. In addition, once the network training is completed, the mapping of the network is usually fixed.
  • the current artificial neural network lacks the necessary flexibility in complex situations, mainly because the mapping rules may change according to the context, and these rules need to be learned "at any time” from a small number of training samples. It can be seen from this that there is a huge capability gap between artificial neural networks and the human brain.
  • the information processing method of Ye Network includes:
  • Step S10 selecting a feature vector extractor based on the acquired information to perform feature extraction to obtain an information feature vector
  • Step S20 input the information feature vector into the prefrontal lobe-like network, and perform dimensional matching with each context signal in the input context signal set to obtain context feature vectors to form a context feature vector set;
  • Step S30 Classify each feature vector in the context feature vector set through a pre-built feature vector classifier to obtain classification information of each feature vector to form a classification information set; the feature vector classifier is a mapping network of the context feature vector and the classification information.
  • step S10 the method of "selecting feature vector extractor" in step S10 is:
  • the corresponding feature vector extractor is selected according to the category of the acquired information.
  • the construction method of the feature vector extractor is:
  • the Adam algorithm is used to iteratively update the weights of parameters in the feature vector extraction network;
  • the feature vector extraction network is constructed based on a deep neural network;
  • the network after the removal of the final classification layer of the trained feature extraction network is used as the feature vector extractor.
  • step S20 "input the information feature vector into the prefrontal lobe-like network, and perform dimensional matching with each context signal in the input context signal set to obtain a context feature vector", the steps are:
  • Step S201 construct a weight matrix based on the context signal and the prefrontal lobe-like network, and perform modular normalization on each column of the weight matrix;
  • W in is the weight matrix, for The normalized modulus of, i is the dimension index of the input feature, k is the dimension of the input feature, and m is the dimension of the hidden layer;
  • Step S202 Dimensionally match the context signal and the information feature vector based on the weight matrix to obtain the context feature vector;
  • Y out is the context feature vector
  • F is the information feature vector
  • C is the situational signal
  • ⁇ represents the multiplication of the corresponding elements of the vector
  • W in is the weight matrix
  • Step S203 the context feature vector obtained after dimension matching of each context signal in the context signal set with the information feature vector constitutes a context feature vector set.
  • the feature vector classifier in step S30 is constructed based on the following formula:
  • Y Lable is the classification information
  • W out is the classification weight of the classifier
  • Y out is the output feature of the class prefrontal network
  • n is the dimension of the output weight of the class prefrontal network
  • F is the information feature vector.
  • the parameters of the Adam algorithm are configured as follows:
  • the learning rate of the Adam algorithm is 0.1, the weight decay rate is 0.0001, and the number of training samples in each batch is 256.
  • the context signal is a multi-dimensional word vector corresponding to the classification attribute; the dimension of the word vector is 200 dimensions.
  • the weight matrix W in is:
  • a matrix of the dimension of the word vector ⁇ the prefrontal quasi-frontal lobe dimension is constructed.
  • an information processing system based on the contextual signal type prefrontal network which includes an acquisition module, a feature extraction module, a dimension matching module, a classification module, and an output module;
  • the acquisition module is configured to acquire and input input information and a set of situational signals
  • the feature extraction module is configured to extract features of the input information by using a feature vector extractor corresponding to the input information to obtain an information feature vector;
  • the dimensional matching module is configured to input the information feature vector into the prefrontal lobe-like network, and perform dimensional matching with each context signal in the input context signal set to obtain the context feature vector to form a context feature vector set;
  • the classification module is configured to classify each feature vector in the context feature vector set through a pre-built feature vector classifier to obtain classification information of each feature vector to form a classification information set;
  • the output module is configured to output the acquired classification information set.
  • a storage device in which a plurality of programs are stored, and the programs are adapted to be loaded and executed by a processor to implement the above-mentioned information processing method based on the context signal type prefrontal network.
  • a processing device including a processor and a storage device; the processor is suitable for executing each program; the storage device is suitable for storing multiple programs; the program is suitable for It is loaded and executed by the processor to realize the above-mentioned information processing method based on the context signal type prefrontal network.
  • the multi-task information processing method based on the context signal of the present invention uses a module similar to the prefrontal lobe to realize context-oriented multi-task learning.
  • the mapping that depends on the context information can be gradually learned.
  • the data processed by the method of the present invention can be applied to multi-task learning or higher-required continuous multi-task learning, and can simplify the network structure, reduce the difficulty of multi-task learning, and increase system flexibility.
  • the present invention uses a deep neural network as a feature extractor, and then uses the designed optimization method in the linear layer. This not only gives full play to the role of the deep neural network, but also reduces the design difficulty.
  • the situational signal is designed in the method of the present invention. This situational signal can be changed according to the current working environment, which solves the problem that the neural network cannot respond differently to the same stimulus according to different goals, environments, and internal conditions. Defects.
  • FIG. 1 is a schematic flow chart of the information processing method based on the contextual signal-like prefrontal network of the present invention
  • FIG. 2 is a schematic diagram of the network structure of the information processing method based on the contextual signal-like prefrontal network of the present invention
  • FIG. 3 is a three-dimensional schematic diagram of an embodiment of the information processing method based on the contextual signal-like prefrontal network of the present invention
  • FIG. 4 is a schematic diagram of the network architecture of the traditional multi-task learning network and the continuous multi-task learning network based on the information processing method of the context signal-like prefrontal lobe network of the present invention
  • Fig. 5 is a schematic diagram of the accuracy of face recognition tasks of multi-task training and continuous multi-scene training of the information processing method based on the context signal-like prefrontal network of the present invention.
  • An information processing method based on the contextual signal type prefrontal network of the present invention includes:
  • Step S10 selecting a feature vector extractor based on the acquired information to perform feature extraction to obtain an information feature vector
  • Step S20 input the picture feature vector into the prefrontal lobe-like network, and perform dimension matching with each context signal in the input context signal set to obtain context feature vectors to form a context feature vector set;
  • Step S30 Classify each feature vector in the context feature vector set through a pre-built feature vector classifier to obtain classification information of each feature vector to form a classification information set; the feature vector classifier is a mapping network of the context feature vector and the classification information.
  • the information processing method based on the context signal type prefrontal network includes steps S10 to S30, and each step is described in detail as follows:
  • step S10 a feature vector extractor is selected based on the acquired information to perform feature extraction to obtain an information feature vector.
  • the corresponding feature vector extractor is selected according to the category of the acquired information.
  • the feature vector extractor library includes one or more of image feature vector extractor, speech feature vector extractor, and text feature vector extractor; it can also include feature vector extractors of other common information categories, which will not be listed here. .
  • the feature vector extractor in the present invention can be constructed based on a deep neural network. For example, for image input information, deep neural networks such as ResNet can be selected; for voice input information, deep neural networks such as CNN, LSTM, GRU, etc. can be selected; for text input information , You can use deep neural networks such as fastText, TextCNN and TextRNN.
  • information is generally multi-modal, and multiple feature vector processors can be used to extract features at the same time, which can enrich the expression of information and greatly reduce the dimensionality of the original information, making downstream information easier to process.
  • Each feature vector extractor in the feature vector extractor library its construction method is:
  • the Adam algorithm is used to iteratively update the weights of parameters in the feature vector extraction network; the feature vector extraction network is constructed based on a deep neural network.
  • the network after the removal of the final classification layer of the trained feature extraction network is used as the feature vector extractor.
  • the learning rate of the Adam algorithm is 0.1
  • the weight decay rate is 0.0001
  • the number of training samples (Batch Size) of each batch is 256.
  • the training data set obtained by the present invention is the face data set CelebA (CelebA is the open data of the Chinese University of Hong Kong, containing 202,599 pictures of 10,177 celebrity identities, and each face has 40 attributes corresponding to each attribute as a scene. Different scenes correspond to different situational signals).
  • the Adam algorithm is used to train the deep neural network ResNet50, the final classification layer of the ResNet50 network is removed, and the output of the penultimate layer is used as the feature of the face data.
  • the output of the penultimate layer has 2048 dimensions.
  • Step S20 Input the information feature vector into the prefrontal lobe-like network, and perform dimensional matching with each context signal in the input context signal set to obtain context feature vectors to form a context feature vector set.
  • Fig. 2 is a schematic diagram of a prefrontal lobe-like network structure based on contextual signals according to an embodiment of the present invention.
  • the English word vector trained with the default parameters of the gensim toolkit is used as the context signal, and the context signals of 40 attribute classification tasks are the 200-dimensional word vectors of the corresponding attribute tags.
  • Step S201 Construct a weight matrix based on the context signal and the prefrontal lobe-like network, and perform modular normalization on each column of the weight matrix.
  • the dimension of the situational signal is 200 dimensions, and the frontal layer of the forehead is 5000 dimensions, and a weight matrix with a size of 200 ⁇ 5000 is constructed.
  • the three-dimensional schematic diagram of the preferred embodiment of the present invention is shown in FIG. 3.
  • the constructed weight matrix W in is shown in formula (1):
  • k is the dimension of the input feature
  • m is the dimension of the hidden layer.
  • i is the dimension index of the input feature.
  • Step S202 Dimensionally match the context signal and the information feature vector based on the weight matrix to obtain the context feature vector, as shown in formula (3):
  • Y out is the context feature vector
  • F is the information feature vector
  • C is the situational signal
  • ⁇ represents the multiplication of the corresponding elements of the vector
  • W in is the weight matrix
  • Step S203 the context feature vector obtained after the dimension matching of each context signal in the context signal set with the input information feature vector constitutes a context feature vector set.
  • Step S30 Classify each feature vector in the context feature vector set through a pre-built feature vector classifier to obtain classification information of each feature vector to form a classification information set; the feature vector classifier is a mapping network of the context feature vector and the classification information.
  • the feature vector classifier is constructed based on formula (4) and formula (5):
  • Y Lable is the classification information
  • W out is the classification weight of the classifier
  • Y out is the output feature of the class prefrontal network
  • n is the dimension of the output weight of the class prefrontal network
  • F is the information feature vector.
  • FIG. 4 it is a schematic diagram of the network architecture of the traditional multi-task learning network and the continuous multi-task learning network based on the information processing method of the context signal prefrontal network of the present invention, and C represents the classifier.
  • C represents the classifier.
  • a switch module and n classifiers are needed in multi-task training, where n is the number of context signals.
  • FIG. 5 it is a schematic diagram of the face recognition task accuracy of the multi-task training and continuous multi-scene training of the information processing method based on the context signal prefrontal network of the present invention.
  • Each point represents a face attribute, totaling 40.
  • Each attribute is associated with a context signal, so that continuous multi-task learning based on the context signal can be performed to achieve the results obtained by multi-task training.
  • the information processing system based on the context signal type prefrontal network includes an acquisition module, a feature extraction module, a dimension matching module, a classification module, and an output module;
  • the acquisition module is configured to acquire and input input information and a set of situational signals
  • the feature extraction module is configured to extract features of the input information by using a feature vector extractor corresponding to the input information to obtain an information feature vector;
  • the dimensional matching module is configured to input the information feature vector into the prefrontal lobe-like network, and perform dimensional matching with each context signal in the input context signal set to obtain the context feature vector to form a context feature vector set;
  • the classification module is configured to classify each feature vector in the context feature vector set through a pre-built feature vector classifier to obtain classification information of each feature vector to form a classification information set;
  • the output module is configured to output the acquired classification information set.
  • the information processing system based on the contextual signal type prefrontal network provided in the above embodiments is only illustrated by the division of the above functional modules.
  • the above functions can be allocated to different Functional modules are implemented, that is, the modules or steps in the embodiments of the present invention are further decomposed or combined.
  • the modules of the above embodiments can be combined into one module, or further divided into multiple sub-modules to complete all or Part of the function.
  • the names of the modules and steps involved in the embodiments of the present invention are only for distinguishing each module or step, and are not regarded as improper limitations on the present invention.
  • a plurality of programs are stored therein, and the programs are adapted to be loaded and executed by a processor to implement the above-mentioned information processing method based on the context signal type prefrontal network.
  • a processing device includes a processor and a storage device; the processor is suitable for executing each program; the storage device is suitable for storing multiple programs; the program is suitable for being loaded and executed by the processor In order to realize the above-mentioned information processing method based on the situational signal type prefrontal network.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Data Mining & Analysis (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Multimedia (AREA)
  • General Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Molecular Biology (AREA)
  • Biomedical Technology (AREA)
  • Medical Informatics (AREA)
  • Databases & Information Systems (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Mathematical Physics (AREA)
  • Human Computer Interaction (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • Biodiversity & Conservation Biology (AREA)
  • Image Analysis (AREA)
  • Character Discrimination (AREA)

Abstract

本发明属于模式识别及类脑机器学习领域，具体涉及了一种基于情境信号类前额叶网络的信息处理方法、系统、装置，旨在解决复杂情况下即复杂多任务情况下系统结构复杂、灵活性差、训练样本需求量大的问题。本发明方法包括：选择对应特征向量提取器进行特征提取；将信息特征向量与情境信号集中每一个情境信号进行维度匹配；维度匹配后的情境特征向量输入特征向量分类器，获得分类信息。本发明方法利用类似于前额叶的模块，实现面向情境信息的多任务学习，在上下文情境信息不能事先确定的情况下，可以逐步学习依赖于上下文情境信息的映射，处理后的数据可应用于多任务学习或更高要求的连续多任务学习。

Description

基于情境信号类前额叶网络的信息处理方法、系统、装置 技术领域
本发明属于模式识别及类脑机器学习领域，具体涉及了一种基于情境信号类前额叶网络的信息处理方法、系统、装置。
背景技术
高级智能的特点之一就是具有灵活性。人类可以根据不同的目标，环境和内部状态等不同情境对同一刺激做出不同的反应。前额叶在灵长类中高度精细化，在实现这种能力中起着关键的作用。前额叶可以快速学习“游戏规则”，并动态地将它们应用于将感官输入映射到以不同的动作为依据的上下文相关的任务中,这个过程被称为认知控制，它让灵长类在无限多种情况下都可以有适当的行为。
当前的人工神经网络在从原始数据中提取高级特征去做模式分类，学习复杂的映射规则方面功能非常强大，然而，它们的响应主要由网络输入决定的，并呈现出刻板的输入输出映射。另外，一旦网络训练完成，那么网络的映射通常是固定的。
因此，目前的人工神经网络在复杂情况下缺乏必要的灵活性，主要是因为映射规则可能根据上下文而变化，而且这些规则需要从少量训练样本中“随时”学习。从这可以看出人工神经网络和人类大脑之间具有巨大的能力差距。
发明内容
为了解决现有技术中的上述问题，即复杂多任务情况下系统结构复杂，灵活性差，以及训练样本需求量大的问题，本发明提供了一种受前额叶功能启发的，基于情境信号类前额叶网络的信息处理方法，包括：
步骤S10，基于获取信息选择特征向量提取器进行特征提取，得到信息特征向量；
步骤S20，将所述信息特征向量输入类前额叶网络，与所输入的情境信号集中每一个情境信号进行维度匹配，获得情境特征向量，构成情境特征向量集；
步骤S30，将情境特征向量集中各特征向量通过预先构建的特征向量分类器分类获得各特征向量分类信息，构成分类信息集；所述特征向量分类器为情境特征向量和分类信息的映射网络。
在一些优选的实施例中步骤S10中“选择特征向量提取器”，其方法为：
基于预设的特征向量提取器库，依据获取信息的类别选择对应的特征向量提取器。
在一些优选的实施例中，所述特征向量提取器，其构建方法为：
基于训练数据集，采用Adam算法迭代地更新特征向量提取网络中参数的权重；所述特征向量提取网络基于深度神经网络构建；
将训练后的特征提取网络最后的分类层去除后的网络作为特征向量提取器。
在一些优选的实施例中，步骤S20中“将所述信息特征向量输入类前额叶网络，与所输入的情境信号集中每一个情境信号进行维度匹配，获得情境特征向量”，其步骤为：
步骤S201，基于情境信号及类前额叶网络构建权重矩阵，并对该权重矩阵每一列进行模归一化；
Figure PCTCN2019083356-appb-000001
Figure PCTCN2019083356-appb-000002
其中，W in为权重矩阵，
Figure PCTCN2019083356-appb-000003
为
Figure PCTCN2019083356-appb-000004
的归一化的模，i为输入特征的维度索引，k为输入特征的维度，m为隐藏层的维度；
步骤S202，基于权重矩阵，将情境信号与信息特征向量进行维度匹配，获得情境特征向量；
Y out＝g([c 1cosθ 1,c 2cosθ 2,…,c mcosθ m] T)||F||
其中，Y out为情境特征向量，
Figure PCTCN2019083356-appb-000005
F为信息特征向量，
Figure PCTCN2019083356-appb-000006
C为情境信号，
Figure PCTCN2019083356-appb-000007
⊙代表向量对应元素相乘；θ m是
Figure PCTCN2019083356-appb-000008
和F之间的角度，g＝max(0,x)；W in为权重矩阵，
Figure PCTCN2019083356-appb-000009
步骤S203，所述情境信号集中每一个情境信号与信息特征向量进行维度匹配后获得的情境特征向量构成情境特征向量集。
在一些优选的实施例中，步骤S30中所述特征向量分类器基于下式构建：
Y Lable＝(W out) TY out＝||W out||||F||cosφ
Figure PCTCN2019083356-appb-000010
其中，Y Lable为分类信息，W out为分类器的分类权重，Y out为类前额叶网络输出的特征，n为类前额叶网络输出权重的维度，F为信息特征向量。
在一些优选的实施例中，所述Adam算法，其参数配置为：
Adam算法的学习率为0.1，权重衰减率为0.0001，每个批次的训练样本数为256。
在一些优选的实施例中，所述情境信号为对应分类属性的多维词向量；所述词向量维度为200维。
在一些优选的实施例中，所述权重矩阵W in为：
基于词向量的维度以及类前额前层的维度，构建的词向量维度×类前额叶维度大小的矩阵。
本发明的另一方面，提出了一种基于情境信号类前额叶网络的信息处理系统，包括获取模块、特征提取模块、维度匹配模块、分类模块、输出模块；
所述获取模块，配置为获取输入信息以及情境信号集并输入；
所述特征提取模块，配置为采用对应输入信息的特征向量提取器提取输入信息的特征，获得信息特征向量；
所述维度匹配模块，配置为将信息特征向量输入类前额叶网络，与所输入的情境信号集中每一个情境信号进行维度匹配，获得情境特征向量，构成情境特征向量集；
所述分类模块，配置为将情境特征向量集中各特征向量通过预先构建的特征向量分类器分类获得各特征向量分类信息，构成分类信息集；
所述输出模块，配置为将获取的分类信息集输出。
本发明的第三方面，提出了一种存储装置，其中存储有多条程序，所述程序适于由处理器加载并执行以实现上述的基于情境信号类前额叶网络的信息处理方法。
本发明的第四方面，提出了一种处理装置，包括处理器、存储装置；所述处理器，适于执行各条程序；所述存储装置，适于存储多 条程序；所述程序适于由处理器加载并执行以实现上述的基于情境信号类前额叶网络的信息处理方法。
本发明的有益效果：
(1)本发明基于情境信号的多任务信息处理方法利用类似于前额叶的模块，实现面向情境信息的多任务学习。在上下文情境信息不能事先确定的情况下，可以逐步学习依赖于上下文情境信息的映射。经过本发明方法处理后的数据可以应用于多任务学习或更高要求的连续多任务学习，并且可以简化网络结构，减少多任务学习难度，增加系统灵活性。
(2)本发明使用深度神经网络作为特征提取器，然后在线性层使用设计的优化方法。这样既充分发挥了深度神经网络的作用，又降低了设计难度。
(3)本发明方法中设计了情境信号，这个情境信号可以根据当前工作的环境变化而变化，解决了神经网络不能根据不同的目标，环境和内部状态等不同情境对同一刺激做出不同的反应的缺陷。
附图说明
通过阅读参照以下附图所作的对非限制性实施例所作的详细描述，本申请的其它特征、目的和优点将会变得更明显：
图1是本发明基于情境信号类前额叶网络的信息处理方法的流程示意图；
图2是本发明基于情境信号类前额叶网络的信息处理方法的网络结构示意图；
图3是本发明基于情境信号类前额叶网络的信息处理方法的实施例的三维空间示意图；
图4是本发明基于情境信号类前额叶网络的信息处理方法的传统多任务学习网络和连续多任务学习网络的网络架构示意图；
图5是本发明基于情境信号类前额叶网络的信息处理方法的多任务训练和连续多场景训练的人脸识别任务精度示意图。
具体实施方式
下面结合附图和实施例对本申请作进一步的详细说明。可以理解的是，此处所描述的具体实施例仅用于解释相关发明，而非对该发明的限定。另外还需要说明的是，为了便于描述，附图中仅示出了与有关发明相关的部分。
需要说明的是，在不冲突的情况下，本申请中的实施例及实施例中的特征可以相互组合。下面将参考附图并结合实施例来详细说明本申请。
本发明的一种基于情境信号类前额叶网络的信息处理方法，包括：
步骤S10，基于获取信息选择特征向量提取器进行特征提取，得到信息特征向量；
步骤S20，将所述图片特征向量输入类前额叶网络，与所输入的情境信号集中每一个情境信号进行维度匹配，获得情境特征向量，构成情境特征向量集；
步骤S30，将情境特征向量集中各特征向量通过预先构建的特征向量分类器分类获得各特征向量分类信息，构成分类信息集；所述特征向量分类器为情境特征向量和分类信息的映射网络。
为了更清晰地对本发明基于情境信号类前额叶网络的信息处理方法进行说明，下面结合图1对本发明方法实施例中各步骤展开详述。
本发明一种实施例的基于情境信号类前额叶网络的信息处理方法，包括步骤S10-步骤S30，各步骤详细描述如下：
步骤S10，基于获取信息选择特征向量提取器进行特征提取，得到信息特征向量。
基于预设的特征向量提取器库，依据获取信息的类别选择对应的特征向量提取器。
特征向量提取器库包含图片特征向量提取器、语音特征向量提取器、文本特征向量提取器中的一个或多个；还可以包括其他常见信息类别的特征向量提取器，此处不再一一列举。本发明中特征向量提取器可以基于深度神经网络构建，例如，对于图片输入信息，可选择ResNet等深度神经网络；对于语音输入信息，可选择CNN、LSTM、GRU等深度神经网络；对于文本输入信息，可以使用fastText、TextCNN和TextRNN等深度神经网络。
在现实环境中，信息一般是多模态的，可同时结合使用多种特征向量处理器提取特征，能够丰富信息的表达，还可以大大降低原始信息的维度，使得下游信息更容易处理。
特征向量提取器库中的每一个特征向量提取器，其构建方法为：
基于训练数据集，采用Adam算法迭代地更新特征向量提取网络中参数的权重；所述特征向量提取网络基于深度神经网络构建。
将训练后的特征提取网络最后的分类层去除后的网络作为特征向量提取器。
本发明优选的实施例中，Adam算法的学习率为0.1，权重衰减率为0.0001，每个批次的训练样本数(Batch Size)为256。本发明获取的训练数据集为人脸数据集CelebA(CelebA是香港中文大学的开放数据， 包含10177个名人身份的202599张图片，每个人脸有对应的40个属性，每个属性作为一种场景，不同的场景对应不同的情境信号)。
采用Adam算法训练深度神经网络ResNet50，去除ResNet50网络最后的分类层，使用倒数第二层的输出作为人脸数据的特征，倒数第二层的输出有2048维度。
步骤S20，将所述信息特征向量输入类前额叶网络，与所输入的情境信号集中每一个情境信号进行维度匹配，获得情境特征向量，构成情境特征向量集。如图2所示为本发明实施例的基于情境信号的类前额叶网络结构示意图。
本发明优选的实施例中，采用gensim工具包的默认参数训练的英文词向量作为情境信号，40个属性分类任务的情境信号为其对应属性标签的200维词向量。
步骤S201，基于情境信号及类前额叶网络构建权重矩阵，并对该权重矩阵每一列进行模归一化。
情境信号维度为200维，类前额前层为5000维，构建大小为200×5000的权重矩阵。本发明优选实施例的三维空间示意图如图3所示。
构建的权重矩阵W in如式(1)所示：
Figure PCTCN2019083356-appb-000011
其中，k为输入特征的维度，m为隐藏层的维度。
对权重矩阵每一列归一化，如式(2)所示：
Figure PCTCN2019083356-appb-000012
其中，i为输入特征的维度索引。
步骤S202，基于权重矩阵，将情境信号与信息特征向量进行维度匹配，获得情境特征向量，如式(3)所示：
Figure PCTCN2019083356-appb-000013
Figure PCTCN2019083356-appb-000014
其中，Y out为情境特征向量，
Figure PCTCN2019083356-appb-000015
F为信息特征向量，
Figure PCTCN2019083356-appb-000016
C为情境信号，
Figure PCTCN2019083356-appb-000017
⊙代表向量对应元素相乘；θ m是
Figure PCTCN2019083356-appb-000018
和F之间的角度，g＝max(0,x)；W in为权重矩阵，
Figure PCTCN2019083356-appb-000019
步骤S203，所述情境信号集中每一个情境信号与输入信息特征向量进行维度匹配后获得的情境特征向量构成情境特征向量集。
步骤S30，将情境特征向量集中各特征向量通过预先构建的特征向量分类器分类获得各特征向量分类信息，构成分类信息集；所述特征向量分类器为情境特征向量和分类信息的映射网络。
特征向量分类器基于式(4)和式(5)构建：
Y Lable＝(W out) TY out＝||W out||||F||cosφ     式(4)
Figure PCTCN2019083356-appb-000020
其中，Y Lable为分类信息，W out为分类器的分类权重，Y out为类前额叶网络输出的特征，n为类前额叶网络输出权重的维度，F为信息特征向量。
如图4所示，为本发明基于情境信号类前额叶网络的信息处理方法的传统多任务学习网络和连续多任务学习网络的网络架构示意图， C代表分类器。为了实现依赖于上下文的处理，在多任务训练中需要开关模块和n个分类器，其中n是情境信号的个数。
如图5所示，为本发明基于情境信号类前额叶网络的信息处理方法的多任务训练和连续多场景训练的人脸识别任务精度示意图，每个点代表一个人脸属性，总共40个。每个属性与一个情境信号相关联，从而可以做基于情境信号的连续多任务学习，实现多任务训练获得的结果。
本发明第二实施例的基于情境信号类前额叶网络的信息处理系统，包括获取模块、特征提取模块、维度匹配模块、分类模块、输出模块；
所述获取模块，配置为获取输入信息以及情境信号集并输入；
所述特征提取模块，配置为采用对应输入信息的特征向量提取器提取输入信息的特征，获得信息特征向量；
所述维度匹配模块，配置为将信息特征向量输入类前额叶网络，与所输入的情境信号集中每一个情境信号进行维度匹配，获得情境特征向量，构成情境特征向量集；
所述分类模块，配置为将情境特征向量集中各特征向量通过预先构建的特征向量分类器分类获得各特征向量分类信息，构成分类信息集；
所述输出模块，配置为将获取的分类信息集输出。
所属技术领域的技术人员可以清楚地了解到，为描述的方便和简洁，上述描述的系统的具体工作过程及有关说明，可以参考前述方法实施例中的对应过程，在此不再赘述。
需要说明的是，上述实施例提供的基于情境信号类前额叶网络的信息处理系统，仅以上述各功能模块的划分进行举例说明，在实际应用中，可以根据需要而将上述功能分配由不同的功能模块来完成，即 将本发明实施例中的模块或者步骤再分解或者组合，例如，上述实施例的模块可以合并为一个模块，也可以进一步拆分成多个子模块，以完成以上描述的全部或者部分功能。对于本发明实施例中涉及的模块、步骤的名称，仅仅是为了区分各个模块或者步骤，不视为对本发明的不当限定。
本发明第三实施例的一种存储装置，其中存储有多条程序，所述程序适于由处理器加载并执行以实现上述的基于情境信号类前额叶网络的信息处理方法。
本发明第四实施例的一种处理装置，包括处理器、存储装置；处理器，适于执行各条程序；存储装置，适于存储多条程序；所述程序适于由处理器加载并执行以实现上述的基于情境信号类前额叶网络的信息处理方法。
所属技术领域的技术人员可以清楚地了解到，为描述的方便和简洁，上述描述的存储装置、处理装置的具体工作过程及有关说明，可以参考前述方法实施例中的对应过程，在此不再赘述。
本领域技术人员应该能够意识到，结合本文中所公开的实施例描述的各示例的模块、方法步骤，能够以电子硬件、计算机软件或者二者的结合来实现，软件模块、方法步骤对应的程序可以置于随机存储器(RAM)、内存、只读存储器(ROM)、电可编程ROM、电可擦除可编程ROM、寄存器、硬盘、可移动磁盘、CD-ROM、或技术领域内所公知的任意其它形式的存储介质中。为了清楚地说明电子硬件和软件的可互换性，在上述说明中已经按照功能一般性地描述了各示例的组成及步骤。这些功能究竟以电子硬件还是软件方式来执行，取决于技术方案的特定应用和设计约束条件。本领域技术人员可以对每个特定的应用来使用不同方法来实现所描述的功能，但是这种实现不应认为超出本发明的范围。
术语“包括”或者任何其它类似用语旨在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、物品或者设备/装置不仅包括那些要素，而且还包括没有明确列出的其它要素，或者还包括这些过程、方法、物品或者设备/装置所固有的要素。
至此，已经结合附图所示的优选实施方式描述了本发明的技术方案，但是，本领域技术人员容易理解的是，本发明的保护范围显然不局限于这些具体实施方式。在不偏离本发明的原理的前提下，本领域技术人员可以对相关技术特征作出等同的更改或替换，这些更改或替换之后的技术方案都将落入本发明的保护范围之内。

Claims (11)

  1. 一种基于情境信号类前额叶网络的信息处理方法，其特征在于，包括：
    步骤S10，基于获取信息选择特征向量提取器进行特征提取，得到信息特征向量；
    步骤S20，将所述信息特征向量输入类前额叶网络，与所输入的情境信号集中每一个情境信号进行维度匹配，获得情境特征向量，构成情境特征向量集；
    步骤S30，将情境特征向量集中各特征向量通过预先构建的特征向量分类器分类获得各特征向量分类信息，构成分类信息集；所述特征向量分类器为情境特征向量和分类信息的映射网络。
  2. 根据权利要求1所述的基于情境信号类前额叶网络的信息处理方法，其特征在于，步骤S10中“选择特征向量提取器”，其方法为：
    基于预设的特征向量提取器库，依据获取信息的类别选择对应的特征向量提取器。
  3. 根据权利要求2所述的基于情境信号类前额叶网络的信息处理方法，其特征在于，所述特征向量提取器，其构建方法为：
    基于训练数据集，采用Adam算法迭代地更新特征向量提取网络中参数的权重；所述特征向量提取网络基于深度神经网络构建；
    将训练后的特征提取网络最后的分类层去除后的网络作为特征向量提取器。
  4. 根据权利要求1所述的基于情境信号类前额叶网络的信息处理方 法，其特征在于，步骤S20中“将所述信息特征向量输入类前额叶网络，与所输入的情境信号集中每一个情境信号进行维度匹配，获得情境特征向量”，其步骤为：
    步骤S201，基于情境信号及类前额叶网络构建权重矩阵，并对该权重矩阵每一列进行模归一化；
    Figure PCTCN2019083356-appb-100001
    Figure PCTCN2019083356-appb-100002
    其中，Win为权重矩阵，
    Figure PCTCN2019083356-appb-100003
    为
    Figure PCTCN2019083356-appb-100004
    的归一化的模，i为隐藏层的维度索引，k为输入特征的维度，m为隐藏层的维度；
    步骤S202，基于权重矩阵，将情境信号与信息特征向量进行维度匹配，获得情境特征向量；
    Y out＝g([c 1 cosθ 1,c 2 cosθ 2,…,c m cosθ m] T)||F||
    其中，Y out为情境特征向量，
    Figure PCTCN2019083356-appb-100005
    F为信息特征向量，
    Figure PCTCN2019083356-appb-100006
    C为情境信号，
    Figure PCTCN2019083356-appb-100007
    ⊙代表向量对应元素相乘；θ m是
    Figure PCTCN2019083356-appb-100008
    和F之间的角度，g＝max(0,x)；W in为权重矩阵，
    Figure PCTCN2019083356-appb-100009
    步骤S203，所述情境信号集中每一个情境信号与输入信息特征向量进行维度匹配后获得的情境特征向量构成情境特征向量集。
  5. 根据权利要求1所述的基于情境信号类前额叶网络的信息处理方法，其特征在于，步骤S30中所述特征向量分类器基于下式构建：
    Y Lable＝(W out) TY out＝||W out||||F||cosφ
    Figure PCTCN2019083356-appb-100010
    其中，Y Lable为分类信息，W out为分类器的分类权重，Y out为类前额叶网络输出的特征，n为类前额叶网络输出权重的维度，F为信息特征向量。
  6. 根据权利要求2所述的基于情境信号类前额叶网络的信息处理方法，其特征在于，所述Adam算法，其参数配置为：
    Adam算法的学习率为0.1，权重衰减率为0.0001，每个批次的训练样本数为256。
  7. 根据权利要求3所述的基于情境信号类前额叶网络的信息处理方法，其特征在于，所述情境信号为对应分类属性的多维词向量；所述词向量维度为200维。
  8. 根据权利要求6所述的基于情境信号类前额叶网络的信息处理方法，其特征在于，所述权重矩阵W in为：
    基于词向量的维度以及类前额前层的维度，构建的词向量维度×类前额叶维度大小的矩阵。
  9. 一种基于情境信号类前额叶网络的信息处理系统，其特征在于，包括获取模块、特征提取模块、维度匹配模块、分类模块、输出模块；
    所述获取模块，配置为获取输入信息以及情境信号集并输入；
    所述特征提取模块，配置为采用对应输入信息的特征向量提取器提取输入信息的特征，获得信息特征向量；
    所述维度匹配模块，配置为将信息特征向量输入类前额叶网络，与所输入的情境信号集中每一个情境信号进行维度匹配，获得情境特征向量，构成情境特征向量集；
    所述分类模块，配置为将情境特征向量集中各特征向量通过预先构建的特征向量分类器分类获得各特征向量分类信息，构成分类信息集；
    所述输出模块，配置为将获取的分类信息集输出。
  10. 一种存储装置，其中存储有多条程序，其特征在于，所述程序适于由处理器加载并执行以实现权利要求1-8任一项所述的基于情境信号类前额叶网络的信息处理方法。
  11. 一种处理装置，包括
    处理器，适于执行各条程序；以及
    存储装置，适于存储多条程序；
    其特征在于，所述程序适于由处理器加载并执行以实现：
    权利要求1-8任一项所述的基于情境信号类前额叶网络的信息处理方法。
PCT/CN2019/083356 2019-01-22 2019-04-19 基于情境信号类前额叶网络的信息处理方法、系统、装置 WO2020151108A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US16/971,691 US10915815B1 (en) 2019-01-22 2019-04-19 Information processing method, system and device based on contextual signals and prefrontal cortex-like network

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910058284.2 2019-01-22
CN201910058284.2A CN109784287A (zh) 2019-01-22 2019-01-22 基于情景信号类前额叶网络的信息处理方法、系统、装置

Publications (1)

Publication Number Publication Date
WO2020151108A1 true WO2020151108A1 (zh) 2020-07-30

Family

ID=66501994

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/083356 WO2020151108A1 (zh) 2019-01-22 2019-04-19 基于情境信号类前额叶网络的信息处理方法、系统、装置

Country Status (3)

Country Link
US (1) US10915815B1 (zh)
CN (1) CN109784287A (zh)
WO (1) WO2020151108A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116244517B (zh) * 2023-03-03 2023-11-28 北京航空航天大学 基于层次化信息抽取网络的多场景多任务的模型训练方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20050036676A1 (en) * 2003-06-30 2005-02-17 Bernd Heisele Systems and methods for training component-based object identification systems
US8442321B1 (en) * 2011-09-14 2013-05-14 Google Inc. Object recognition in images
CN106650796A (zh) * 2016-12-06 2017-05-10 国家纳米科学中心 一种基于人工智能的细胞荧光图像分类方法和系统
CN107690034A (zh) * 2017-10-27 2018-02-13 中国科学技术大学苏州研究院 基于环境背景声音的智能情景模式切换系统及方法
CN108937968A (zh) * 2018-06-04 2018-12-07 安徽大学 基于独立分量分析的情感脑电信号的导联选择方法

Family Cites Families (22)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN100422999C (zh) * 2006-09-14 2008-10-01 浙江大学 基于内容相关性的跨媒体检索方法
US20090157481A1 (en) * 2007-12-13 2009-06-18 Searete Llc, A Limited Liability Corporation Of The State Of Delaware Methods and systems for specifying a cohort-linked avatar attribute
US20090157625A1 (en) * 2007-12-13 2009-06-18 Searete Llc, A Limited Liability Corporation Of The State Of Delaware Methods and systems for identifying an avatar-linked population cohort
US20090157660A1 (en) * 2007-12-13 2009-06-18 Searete Llc, A Limited Liability Corporation Of The State Of Delaware Methods and systems employing a cohort-linked avatar
US20190159712A1 (en) * 2009-03-30 2019-05-30 Donald H. Marks Brain function decoding process and system
US20100249573A1 (en) * 2009-03-30 2010-09-30 Marks Donald H Brain function decoding process and system
CN101853426A (zh) * 2010-05-21 2010-10-06 南京邮电大学 普适计算环境中基于神经网络的上下文融合方法
WO2011160309A1 (zh) * 2010-06-25 2011-12-29 中国科学院自动化研究所 基于鲁棒统计信息传播的多模态三维磁共振图像脑肿瘤分割方法
WO2012135068A1 (en) * 2011-03-25 2012-10-04 Drexel University Functional near infrared spectrocopy based brain computer interface
CN107847194B (zh) * 2014-06-30 2020-11-24 塞罗拉公司 使有操作延迟的pc与有实时时钟的微控制器同步的系统
CN105678150A (zh) * 2016-01-11 2016-06-15 成都布林特信息技术有限公司 一种用户权限管理方法
US11138503B2 (en) * 2017-03-22 2021-10-05 Larsx Continuously learning and optimizing artificial intelligence (AI) adaptive neural network (ANN) computer modeling methods and systems
CN106934068A (zh) * 2017-04-10 2017-07-07 江苏东方金钰智能机器人有限公司 机器人基于环境上下文的语义理解的方法
US10939833B2 (en) * 2017-05-01 2021-03-09 Samsung Electronics Company, Ltd. Determining artery location using camera-based sensing
EP3684463A4 (en) * 2017-09-19 2021-06-23 Neuroenhancement Lab, LLC NEURO-ACTIVATION PROCESS AND APPARATUS
CN107766324B (zh) * 2017-09-25 2020-09-01 浙江大学 一种基于深度神经网络的文本一致性分析方法
US11199904B2 (en) * 2017-10-06 2021-12-14 Holland Bloorview Kids Rehabilitation Hospital Brain-computer interface platform and process for classification of covert speech
CN109171769A (zh) * 2018-07-12 2019-01-11 西北师范大学 一种应用于抑郁症检测的语音、面部特征提取方法及系统
US20200060566A1 (en) * 2018-08-24 2020-02-27 Newton Howard Automated detection of brain disorders
WO2020084574A1 (en) * 2018-10-24 2020-04-30 Translational Research Institute Pty Ltd As Trustee For Translational Research Institute Trust Functional analysis of human brain using functional magnetic resonance imaging (fmri) for acute stress and post traumatic stress disorder (ptsd) monitoring neuroplasticity
CN109726389B (zh) * 2018-11-13 2020-10-13 北京邮电大学 一种基于常识和推理的中文缺失代词补全方法
US20200387603A1 (en) * 2019-06-04 2020-12-10 International Business Machines Corporation Device protection based on prediction and contextual analysis

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20050036676A1 (en) * 2003-06-30 2005-02-17 Bernd Heisele Systems and methods for training component-based object identification systems
US8442321B1 (en) * 2011-09-14 2013-05-14 Google Inc. Object recognition in images
CN106650796A (zh) * 2016-12-06 2017-05-10 国家纳米科学中心 一种基于人工智能的细胞荧光图像分类方法和系统
CN107690034A (zh) * 2017-10-27 2018-02-13 中国科学技术大学苏州研究院 基于环境背景声音的智能情景模式切换系统及方法
CN108937968A (zh) * 2018-06-04 2018-12-07 安徽大学 基于独立分量分析的情感脑电信号的导联选择方法

Also Published As

Publication number Publication date
CN109784287A (zh) 2019-05-21
US20210056415A1 (en) 2021-02-25
US10915815B1 (en) 2021-02-09

Similar Documents

Publication Publication Date Title
CN108491765B (zh) 一种蔬菜图像的分类识别方法及系统
Pandey et al. Deep learning techniques for speech emotion recognition: A review
Chu et al. Image style classification based on learnt deep correlation features
Gao et al. Discriminative multiple canonical correlation analysis for information fusion
Shao et al. Feature learning for image classification via multiobjective genetic programming
JP6159489B2 (ja) 顔認証方法およびシステム
WO2020103700A1 (zh) 一种基于微表情的图像识别方法、装置以及相关设备
Kim et al. Deep temporal models using identity skip-connections for speech emotion recognition
CN109284749A (zh) 精细化图像识别
Santhalingam et al. Sign language recognition analysis using multimodal data
Teney et al. Visual question answering: A tutorial
Zhang et al. Facial smile detection based on deep learning features
Paul et al. Rethinking generalization in american sign language prediction for edge devices with extremely low memory footprint
Lu et al. Enhance deep learning performance in face recognition
Rasheed et al. Handwritten Urdu characters and digits recognition using transfer learning and augmentation with AlexNet
WO2020151108A1 (zh) 基于情境信号类前额叶网络的信息处理方法、系统、装置
Chen et al. Cross-situational noun and adjective learning in an interactive scenario
Travieso et al. Using a Discrete Hidden Markov Model Kernel for lip-based biometric identification
Hebda et al. A compact deep convolutional neural network architecture for video based age and gender estimation
CN109961152B (zh) 虚拟偶像的个性化互动方法、系统、终端设备及存储介质
Aljaafari Ichthyoplankton classification tool using Generative Adversarial Networks and transfer learning
CN109685146A (zh) 一种基于双卷积和主题模型的场景识别方法
CN111354364B (zh) 一种基于rnn聚合方式的声纹识别方法与系统
Shukla et al. Deep Learning Model to Identify Hide Images using CNN Algorithm
Mauricio et al. High-resolution generative adversarial neural networks applied to histological images generation

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: 19911506

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: 19911506

Country of ref document: EP

Kind code of ref document: A1