675 results about "Backpropagation" patented technology

The present invention extends to methods, systems, and computer program products for recognizing input gestures. A neural network is trained using example inputs and backpropagation to recognize specified input patterns. Input gesture data is representative of movements in contact on a multi-touch input display surface relative to one or more axes over time. Example inputs used for training the neural network to recognize a specified input pattern can be created from sampling input gesture data for example input gestures known to represent the specified input pattern. Trained neural networks can subsequently be used to recognize input gestures that are similar to known input gestures as the specified input pattern corresponding to the known input gestures.

The invention belongs to the technical field of image processing and computer vision, and relates to a novel biomedical image automatic segmentation method based on a U-net network structure, including dividing a biomedical data set into a training set and a test set, and normalizing the test set and augmented test set; inputting the images of the training set into the improved U-net network model, and generating a classification probability map by output image passing through a softmax layer; calculating the error between classification probability diagram and gold standard by a centralized loss function, and obtaining the weight parameters of network model by a gradient backpropagation method; entering the images in the test set into the improved U-net network model, and outputting the image to generate a classification probability map through the softmax layer; according to the class probability in the classification probability graph, obtaining the segmentation result graph of theimage. The invention solves the problems that simple samples in the image segmentation process contribute too much to the loss function to learn difficult samples well.

The invention relates to a control method of a bus arrival time prediction model, comprising the following steps: (1) firstly establishing a historical database of the bus operating state; (2) training historical data by a BP (backpropagation) neural network method to derive the optimal mean travel time of a bus from a departure point to an electronic stop board; and (3) introducing dynamic operating information of the bus to modify the optimal mean travel time derived by the BP neural network method. Compared with the prior art, the invention not only can accurately predict the bus arrival time under a normal operating condition of the bus, but also can accurately predict the bus arrival time under complex operating conditions of traffic jam, traffic accidents, bad weather conditions, crossing of more traffic light intersections, operation in rush hours or slack hours and the like, and also can automatically update parameters in the prediction model according to the latest historical data of bus operation.

A method and system for objectively evaluating the quality of speech in a voice communication system. A plurality of speech reference vectors is first obtained based on a plurality of clean speech samples. A corrupted speech signal is received and processed to determine a plurality of distortions derived from a plurality of distortion measures based on the plurality of speech reference vectors. The plurality of distortions are processed by a non-linear neural network model to generate a subjective score representing user acceptance of the corrupted speech signal. The non-linear neural network model is first trained on clean speech samples as well as corrupted speech samples through the use of backpropagation to obtain the weights and bias terms necessary to predict subjective scores from several objective measures.

The invention discloses a human face emotion recognition method in a mobile embedded complex environment. In the method, a human face is divided into main areas of a forehead, eyebrows and eyes, cheeks, a noise, a month and a chain, and 68 feature points are further divided. In view of the above feature points, in order to realize the human face emotion recognition rate, the accuracy and the reliability in various environments, a human face and expression feature classification method is used in a normal condition, a Faster R-CNN face area convolution neural network-based method is used in conditions of light, reflection and shadow, a method of combining a Bayes Network, a Markoff chain and variational reasoning is used in complex conditions of motion, jitter, shaking, and movement, and a method of combining a deep convolution neural network, a super-resolution generative adversarial network (SRGANs), reinforcement learning, a backpropagation algorithm and a dropout algorithm is used in conditions of incomplete human face display, a multi-human face environment and noisy background. Thus, the human face expression recognition effects, the accuracy and the reliability can be promoted effectively.

A learning method for acquiring a bounding box corresponding to an object in a training image from multi-scaled feature maps by using a CNN is provided. The learning method includes steps of: (a) allowing an N-way RPN to acquire at least two specific feature maps and allowing the N-way RPN to apply certain operations to the at least two specific feature maps; (b) allowing an N-way pooling layer to generate multiple pooled feature maps by applying pooling operations to respective areas on the at least two specific feature maps; and (c) (i) allowing a FC layer to acquire information on pixel data of the bounding box, and (ii) allowing a loss layer to acquire first comparative data, thereby adjusting at least one of parameters of the CNN by using the first comparative data during a backpropagation process.

The invention discloses an infrared target instance segmentation method based on feature fusion and a dense connection network, and the method comprises the steps: collecting and constructing an infrared image data set required for instance segmentation, and obtaining an original known infrared tag image; Performing image enhancement preprocessing on the infrared image data set; Processing the preprocessed training set to obtain a classification result, a frame regression result and an instance segmentation mask result graph; Performing back propagation in the convolutional neural network by using a random gradient descent method according to the prediction loss function, and updating parameter values of the convolutional neural network; Selecting a fixed number of infrared image data training sets each time and sending the infrared image data training sets to the network for processing, and repeatedly carrying out iterative updating on the convolutional network parameters until the convolutional network training is completed by the maximum number of iterations; And processing the test set image data to obtain average precision and required time of instance segmentation and a finalinstance segmentation result graph.

A system and method for semantic extraction using a neural network architecture includes indexing each word in an input sentence into a dictionary and using these indices to map each word to a d-dimensional vector (the features of which are learned). Together with this, position information for a word of interest (the word to labeled) and a verb of interest (the verb that the semantic role is being predicted for) with respect to a given word are also used. These positions are integrated by employing a linear layer that is adapted to the input sentence. Several linear transformations and squashing functions are then applied to output class probabilities for semantic role labels. All the weights for the whole architecture are trained by backpropagation.

The invention provides an image recognition and recommendation method based on neural network depth learning. The method obtains pictures and classification from an image database, inputs to a convolution neural network, trains the neural network through repeated forward and backward propagation, improves image recognition accuracy, and extracts a 20-layer neural network model. By using this model, the object recognition and classification is carried out by collecting static pictures. Results are recognized, and by combining with the personalized characteristics of the input, the input probability of interest is analyzed. By using the machine learning model based on the effective recognition and classification of the material cloud database, and using the recommendation system algorithm, the predicted content material is pushed to the image inputter for cognitive learning. The method of the invention has the advantages of high image recognition rate, multiple recognition types and accurate content recommendation, and can be applied to the electronic products of a computer with a digital camera, a mobile phone, a tablet and an embedded system, so that people can photograph and recognize the objects seen in the eyes and actively learn the knowledge of recognizing the objects.

The present invention is an automated optical inspection method using deep learning, comprising the steps of: providing a plurality of paired image combinations, wherein each said paired image combination includes at least one defect-free image and at least one defect-containing image corresponding to the defect-free image; providing a convolutional neural network to start a training mode of the convolutional neural network; inputting the plurality of paired image combinations into the convolutional neural network, and adjusting a weight of at least one fully connected layer of the convolutional neural network through backpropagation to complete the training mode of the convolutional neural network; and performing an optical inspection process using the trained convolutional neural network.

A system and method provides efficient parallel training of a neural network model on multiple graphics processing units. A training module reduces the time and communication overhead of gradient accumulation and parameter updating of the network model in a neural network by overlapping processes in an advantageous way. In a described embodiment, a training module overlaps backpropagation, gradient transfer and accumulation in a Synchronous Stochastic Gradient Decent algorithm on a convolution neural network. The training module collects gradients of multiple layers during backpropagation of training from a plurality of graphics processing units (GPUs), accumulates the gradients on at least one processor and then delivers the gradients of the layers to the plurality of GPUs during the backpropagation of the training. The whole model parameters can then be updated on the GPUs after receipt of the gradient of the last layer.

The invention belongs to the technical field of ingredient identification, and discloses a dirty picture cleaning model and an ingredient category identifying method during an ingredient picture preprocessing process. The method comprises the steps that based on a multi-task Auto-Clean convolution neural network model, a K class lexicon and a Yes / No clean tag are given, wherein two CNN models in the model carry out convolution operation on a class tag and a clean tag; after forward propagation, the softmax layer is optimized; and specific parameters are backpropagated. A cleaned image tagged with Yes / No and class is used for iteration in the whole network to acquire the model. The automatic picture cleaning and ingredient category identifying method is realized. According to the invention,on the basis of the prior art, disadvantages are changed in a targeted manner, and optimization is carried out; and for the characteristics of complex and ever-changing background and the like of ingredient images, ingredient images are efficiently, accurately and rapidly identified.

Systems and methods for optical communication that use a transmitter / receiver. The systems and methods include receiving a modulated, encoded input stream. Channel memory is reduced using coarse digital backpropagation and other channel impairments are removed using turbo equalization. Symbols are detected in the input stream that conform to a non-uniform, polar constellation having a Gaussian source distribution to produce a stream of encoded data. The stream of encoded data is decoded with one or more low density parity check (LDPC) decoders.

The invention discloses a device and method for suppressing subsynchronous oscillation of a power system. The method comprises the following steps of: firstly, filtering the rotation speed signal of a generator to obtain the subsynchronous rotation speed signal of each mode; processing the subsynchronous rotation speed signal of each mode respectively to obtain a change rate; then, generating an additional control signal through a Sugeno type fuzzy reasoning system; and finally, performing amplification, overlapping and amplitude limiting on the obtained additional control signal, and generating an exciting voltage additional control signal so as to change the exciting current, generate a subsynchronous frequency damping torque and suppress the subsynchronous oscillation. In the method provided by the invention, a training sample of a fuzzy controller is established according to the phase compensation principle, and the parameters of the fuzzy system are optimized and trained by use of a learning algorithm of an error backpropagation neural network. The method solves the problem that the expert experience is difficult to obtain by the fuzzy controller, and the additional exciting damping controller can effectively suppress the subsynchronous oscillation of the power system.

A system and method for evolving appropriate connections in feedforward topological networks. FIG. 1 illustrates an exemplary flow diagram of a method for evolving appropriate connections in a neural network device according to an aspect of the present invention. Initially, weight changes induced by each particular training sample or pattern are calculated using, for example, a conventional network training rule such as Hebb or backpropagation (step 101). Next, a ratio of the weight changes for existing connections to incipient connections ("K" ratio) is calculated (step 103). If this K ratio exceeds a specified threshold, weight changes are implemented (step 105), in which existing weights are increased by (1-E)x(total weight change of existing connections), where E is a network-wide parameter. The remaining amount (E)x(total weight change of existing connections) is added to form neighboring connections (step 107). It is to be noted that if neighboring connections are not yet in existence (i.e, they are incipient connections), they can be created by this mutation rule; however, whether such new connections are created depends on the size of the weight increase computed in step 101, together with the magnitude of E. After cycling through a training set, connections that are weak (e.g., weaker than a specified threshold) are deleted (step 109). Following step 109, the system returns to step 101. Advantageously, a system and method according to the present invention allows a combination of the advantages of fully connected networks and of sparse networks and reduces the number of calculations that must be done, since only the calculations corresponding to the existing connections and their neighbors need be determined.

The invention discloses a sketch face recognition method based on depth transfer learning, comprising the following steps: step 1, establishing a depth convolution neural network model; 2, preprocessing images in a CUFSF sketch image library; 3, preprocessing that LFW image library containing large-scale natural face photograph, taking the LFW image library as an initial training sample to train anetwork, and obtaining a training model; 4, migrating that training model to a network for sketch face photographs to obtain a pre-training model; 5, establishing a three-tuple image composed of a reference image, a positive sample image and a negative sample image; 6, taking that triple image group as the input of the pre-training model, minimizing the los function by using the back propagationalgorithm, and performing training to obtain the final target training model; 7, testing that target training model obtain in the step 6 with a test set to carry out face recognition of a sketch image. This method has the advantage of high recognition accuracy of sketch face image.

The invention discloses a traffic scene analysis method based on a multi-task network, comprising the following steps: a multi-task network is divided into an encoder, a partition decoder and a detection decoder, wherein the encoder extracts the features of the image and extracts the multi-scale information from the feature map; the segmentation decoder enlarges the size of the feature map and fuses it with the feature map; the detection decoder processes the input characteristic map and outputs a corresponding target detection result; using a deep learning framework Tensorflow to configure, train and test the above multitasking network. The multi-task network of the invention can extract abundant image features, make up for the loss of image detail information caused by downsampling in the encoder, and help to improve the segmentation and detection effect. The invention designs a multi-task network structure, which can realize semantic segmentation and target detection of traffic scene images through one-time back propagation, and has better real-time performance and higher accuracy rate.

The present invention extends to methods, systems, and computer program products for recognizing input gestures. A neural network is trained using example inputs and backpropagation to recognize specified input patterns. Input gesture data is representative of movements in contact on a multi-touch input display surface relative to one or more axes over time. Example inputs used for training the neural network to recognize a specified input pattern can be created from sampling input gesture data for example input gestures known to represent the specified input pattern. Trained neural networks can subsequently be used to recognize input gestures that are similar to known input gestures as the specified input pattern corresponding to the known input gestures.

A method for concealing original data to protect personal information is provided. The method includes steps of: a data obfuscation device (a) if the original data is acquired, inputting the original data or its modified data into a learning network, and allowing the learning network to (i) apply a network operation to the original data or the modified data using learned parameters of the learning network and thus to (ii) output characteristic information on the original data or the modified data; and (b) updating the original data or the modified data via backpropagation using part of (i) 1-st losses calculated by referring to the characteristic information and its corresponding 1-st ground truth, and (ii) 2-nd losses calculated by referring to (ii-1) a task specific output generated by using the characteristic information (ii-2) a 2-nd ground truth corresponding to the task specific output, to thereby generate obfuscated data.

The invention provides an open set domain adaptation method based on entropy minimization. The open set domain adaptation method comprises the steps of creating an open set domain adaptation network, and initializing network hyper-parameters; inputting a source domain picture and a target domain picture into an open set domain adaptive network, obtaining picture depth features through a networkfeature extraction layer, and then obtaining prediction probability vectors of various categories through a softmax layer according to a feature map; calculating an entropy loss function by using a known category prediction probability vector and a picture label so as to correctly identify a common category; on the basis, calculating binary cross entropy and a binary diversity loss function by using an unknown category prediction probability value so as to correctly classify unknown categories, updating network parameters by using a back propagation gradient until a loss function value is minimum, and stopping training the network; and storing the network model and the training result, and introducing the target domain data set into the network model to obtain a final target domain label.

The invention discloses an expressway traffic flow prediction method based on a DenseNet. On the basis of researching a deep two-dimensional convolution network DenseNet, one-dimensional traffic flowdata is imported, in addition, input and convolution ways in a network are revised, so that the network learns a hidden rule in one-dimensional time sequence data, and expressway traffic flow prediction of a next stage is effectively realized. The method comprises the following implementation steps that: (1) reading traffic flow data, and constructing a training test set; (2) carrying out data preprocessing, transforming a four-dimensional tensor, and carrying out normalization to adapt to network learning; (3) constructing the network; (4) training the network, inputting a training sample, updating network parameters through a forward direction prediction result and error backpropagation, and carrying out loop iteration until network convergence is realized; and (5) testing the network, and carrying out traffic flow prediction on the test set. The method can automatically learn a hidden characteristic relationship between flows in a flow dataset, has a better prediction effect and iswidely applied to traffic flow prediction.

A method for improving image segmentation using a learning device including steps of: (a) if a training image is obtained, acquiring (2-K)th to (2-1)th feature maps through an encoding layer and a decoding layer, and acquiring 1st to Hth losses from the 1st to the Hth loss layers respectively corresponding to H feature maps, obtained from the H filters, among the (2-K)th to the (2-1)th feature maps; and (b) upon performing a backpropagation process, performing processes of allowing the (2-M)th filter to apply a convolution operation to (M−1)2-th adjusted feature map relayed from the (2-(M−1))th filter to obtain M1-th temporary feature map; relaying, to the (2-(M+1))th filter, M2-th adjusted feature map obtained by computing the Mth loss with the M1-th temporary feature map; and adjusting at least part of parameters of the (1-1)th to the (1-K)th filters and the (2-K)th to the (2-1)th filters.

The invention provides a two-stage image retrieval method based on a convolutional neural network. The method comprises the following steps: adding a feature extraction layer between a convolutional layer and a dense connection layer of a VGG16 network to construct a convolutional neural network model; training the convolutional neural network model by using the training set and the verification set, and adjusting parameters of the convolutional neural network model by using back propagation; inputting the test set into a trained convolutional neural network model, mapping the feature vectorsby using hash function mapping to obtain binary hash codes, and classifying the vectors output by the dense connection layer by using a softmax classification function to construct a secondary index library; and inputting a to-be-retrieved image into the trained convolutional neural network model, and carrying out first-stage retrieval and second-stage retrieval. According to the method, further search is carried out under the corresponding image category, accurate classification and rapid retrieval of the images are achieved through classification optimization retrieval, the retrieval speed of similar features is increased, and the query efficiency is improved.

The invention relates to a gradient descent and generalized inverse-based complex-valued neural network training method. The method includes the following steps that: step 1, a single-hidden layer complex-valued neural network model is selected; step 2, the gradient descent and generalized inverse are utilized to calculate a weight matrix and a weight vector in the single-hidden layer complex-valued neural network model; step 3, the network parameters of the complex-valued neural network model are obtained according to the weight matrix and the weight vector, and mean square error is calculated, and 1 is added to the number of iterations, and the method returns to the step 2. According to the method of the invention, the input weight of a hidden layer is generated through the gradient descent, and the output weight of the hidden layer is always solved by the generalized inverse. The method of the invention has the advantages of small number of iterations, short corresponding training time, high convergence speed and high learning efficiency, and just needs few hidden layer nodes. Therefore, the method of the invention can reflect the performance of the complex-valued neural network more accurately compared with a BSCBP (Batch Split-Complex Backpropagation Algorithm) method and a CELM (Complex Extreme Learning Machine) method.

The invention provides a training method, a recognition method, a device and a processing device of a recurrent neural network, which relate to the technical field of motion recognition. The method comprises the following steps: a training sample is obtained, wherein the training sample comprises a multi-frame image sequence of a video and a motion identification corresponding to the video; feature extraction is carried out on multi-frame image sequences to obtain image sequence features, and the image sequence features include the features of each frame image; the feature of image sequence isinput into recurrent neural network for action classification, and the action classification probability of each image frame is obtained; among them, the action classification contains no action class; based on the action classification probability, the loss function is calculated according to the connection sequence classification method; the recurrent neural network is trained by back propagation of loss function. The embodiment of the invention can better learn the connection relationship between the actions and more accurately predict the actions of the time series, so that finer granularity and more accurate action recognition can be carried out on the actions.

Simulating uncertainty in an artificial neural network is provided. Aleatoric uncertainty is simulated to measure what the artificial neural network does not understand from sensor data received from an object operating in a real-world environment by adding random values to edge weights between nodes in the artificial neural network during backpropagation of output data of the artificial neural network and measuring impact on the output data by the added random values to the edge weights between the nodes. Epistemic uncertainty is simulated to measure what the artificial neural network does not know by dropping out a selected node from each respective layer of the artificial neural network during forward propagation of the sensor data and measuring impact of dropped out nodes on the output data of the artificial neural network. An action corresponding to the object is performed based on the impact of simulating the aleatoric and epistemic uncertainty.

A learning method for improving performance of a CNN by using Feature Up-sampling Networks is disclosed. The learning method includes steps of: (a) allowing the down-sampling block to acquire a down-sampling image; (b) allowing each of a (1-1)-th to a (1-k)-th filter blocks to respectively acquire each of a (1-1)-th to a (1-k)-th feature maps; (c) allowing a specific up-sampling block to (i) receive a particular feature map from its corresponding filter block, and (ii) receive another specific feature map from its previous up-sampling block, and then rescale a size of the specific feature map to be identical to that of the particular feature map and (iii) apply certain operations to the particular feature map and the resealed specific feature map to generate a feature map of the specific up-sampling block; and (d)(i) allowing an application block to acquire an application-specific output and (ii) performing a first backpropagation process.

A method for training a CNN by using a camera and a radar together, to thereby allow the CNN to perform properly even when an object depiction ratio of a photographed image acquired through the camera is low due to a bad condition of a photographing circumstance is provided. And the method includes steps of: (a) a learning device instructing a convolutional layer to apply a convolutional operation to a multichannel integrated image, to thereby generate a feature map; (b) the learning device instructing an output layer to apply an output operation to the feature map, to thereby generate estimated object information; and (c) the learning device instructing a loss layer to generate a loss by using the estimated object information and GT object information corresponding thereto, and to perform backpropagation by using the loss, to thereby learn at least part of parameters in the CNN.

The invention provides an optical diffraction neural network online training method and system based on an optical reciprocity and phase conjugation principle. For the optical diffraction neural network online training method, in the forward propagation step, input light reaches an imaging surface through a series of phase modulators, and light field distribution of each phase modulator surface and the imaging surface is recorded at the same time; in the loss field calculation step, the error between the intensity of the image plane light field and the standard value is calculated, and the image plane phase conjugation principle light field is modulated according to the error, and the loss light field is obtained through calculation; in the step of back propagation, a complex field generation module is used for generating a loss light field, and the loss light field is subjected to back propagation, and obtained accompanying light fields are recorded on conjugate surfaces of phase modulators one by one; and in the gradient calculation and updating step, the gradient of each pixel of the phase modulator is calculated according to the phase modulator surface light field recorded in the forward propagation step and the accompanying light field recorded in the reverse propagation step, and gradient descent is performed according to the gradient, and iteration is performed until convergence.

The invention discloses an Internet financial credit evaluation method based on a PSO-BP neural network, and the method comprises the steps: obtaining a result true value of information, carrying outthe normalization processing and principal component analysis dimensionality reduction of obtained data, dividing a test set and a training set, initializing the number of input nodes, the number of output nodes and the number of hidden layer nodes of the BP neural network; using a traditional gradient descent method and back propagation for continuously adjusting the weight and the threshold of anetwork to construct a BP neural network model, using a particle swarm algorithm for optimizing the connection weight and the threshold to obtain a PSOBP neural network model, and using a verification set for verification and optimization; and deploying the model to an application system to perform feature parameter extraction and prediction classification on data of a real-time application client. According to the invention, the convergence rate of the BP neural network is greatly improved, the obtained credit evaluation model of the PSO-BP neural network can accurately and quickly realize credit evaluation of an Internet financial applicant, the service timeliness of application approval is effectively improved, and the risk control cost and the application fraud risk are reduced.