709 results about "Recurrent neural nets" patented technology

A dialog device comprises a natural language interfacing device (chat interface or a telephonic device), and a natural language output device (the chat interface, a display device, or a speech synthesizer outputting to the telephonic device). A computer stores natural language dialog conducted via the interfacing device and constructs a current utterance word-by-word. Each word is chosen by applying a plurality of language models to a context comprising concatenation of the stored dialog and the current utterance thus far. Each language model outputs a distribution over the words of a vocabulary. A recurrent neural network (RNN) is applied to the distributions to generate a mixture distribution. The next word is chosen using the mixture distribution. The output device outputs the current natural language utterance after it has been constructed by the computer.

A method and system detects actions of an object in a scene by first acquiring a video of the scene as a sequence of images, wherein each image includes pixels, wherein the video is partitioned into chunks. The object in the video is tracked. For each object and each chunk of the video, trajectories of the pixels within a bounding box located over the object are tracked, and cropped trajectories and cropped images for one or more images in the chunk are produced using the bounding box. Then, the cropped trajectories and cropped images are passed to a recurrent neural network (RNN) that outputs a relative score for each action of interest.

There are disclosed an apparatus for calculating a posteriori probabilities of phoneme symbols and a speech recognition apparatus using the apparatus for calculating a posteriori probabilities of phoneme symbols. A feature extracting section extracts speech feature parameters from a speech signal of an uttered speech sentence composed of an inputted character series, and a calculating section calculates a a posteriori probability of a phoneme symbol of the speech signal, by using a bidirectional recurrent neural network. The bidirectional recurrent neural network includes (a) an input layer for receiving the speech feature parameters extracted by the feature extracting means and a plurality of hypothetical phoneme symbol series signals, (b) an intermediate layer of at least one layer having a plurality of units, and (c) an output layer for outputting a a posteriori probability of each phoneme symbol. The input layer includes (a) a first input neuron group having a plurality of units, for receiving a plurality of speech feature parameters and a plurality of phoneme symbol series signals, (b) a forward module, and (c) a backward module.

The invention discloses a video description method based on deep learning and text summarization. The video description method comprises the following steps: through a traditional image data set, training a convolutional neural network model according to an image classification task; extracting a video frame sequence of a video, utilizing the convolutional neural network model to extract convolutional neural network characteristics to form a <video frame sequence, text description sequence> pair which is used as the input of a recurrent neural network model, and training to obtain the recurrent neural network model; describing the video frame sequence of the video to be described through the recurrent neural network model obtained by training to obtain description sequences; and through a method that graph-based vocabulary centrality is used as the significance of the text summarization, sorting the description sequences, and outputting a final description result of the video. An event which happens in one video and object attributes associated with the event are described through a natural language so as to achieve a purpose that video contents are described and summarized.

The invention discloses a bidirectional long short-term memory unit-based behavior identification method for a video. The method comprises the steps of (1) inputting a video sequence and extracting an RGB (Red, Green and Blue) frame sequence and an optical flow image from the video sequence; (2) respectively training a deep convolutional network of an RGB image and a deep convolutional network of the optical flow image; (3) extracting multilayer characteristics of the network, wherein characteristics of a third convolutional layer, a fifth convolutional layer and a seventh fully connected layer are at least extracted, and the characteristics of the convolutional layers are pooled; (4) training a recurrent neural network constructed by use of a bidirectional long short-term memory unit to obtain a probability matrix of each frame of the video; and (5) averaging the probability matrixes, finally fusing the probability matrixes of an optical flow frame and an RGB frame, taking a category with a maximum probability as a final classification result, and thus realizing behavior identification. According to the method, the conventional artificial characteristics are replaced with multi-layer depth learning characteristics, the depth characteristics of different layers represent different pieces of information, and the combination of multi-layer characteristics can improve the accuracy rate of classification; and the time information is captured by use of the bidirectional long short-term memory, many pieces of time domain structural information are obtained and a behavior identification effect is improved.

Described herein are systems and methods for creating and using Convolutional Recurrent Neural Networks (CRNNs) for small-footprint keyword spotting (KWS) systems. Inspired by the large-scale state-of-the-art speech recognition systems, in embodiments, the strengths of convolutional layers to utilize the structure in the data in time and frequency domains are combined with recurrent layers to utilize context for the entire processed frame. The effect of architecture parameters were examined to determine preferred model embodiments given the performance versus model size tradeoff. Various training strategies are provided to improve performance. In embodiments, using only ˜230 k parameters and yielding acceptably low latency, a CRNN model embodiment demonstrated high accuracy and robust performance in a wide range of environments.

The present invention discloses a deep learning method-based short-term traffic flow prediction method. The influence of the traffic flow rate change of the neighbor points of a prediction point, the time characteristic of the prediction point and the influence of the periodic characteristic of the prediction point on the traffic flow rate of the prediction point are considered simultaneously. According to the deep learning method-based short-term traffic flow prediction method of the invention, a convolutional neural network and a long and short-term memory (LSTM) recurrent neural network are combined to construct a Conv-LSTM deep neural network model; a two-way LSTM model is used to analyze the traffic flow historical data of the point and extract the periodic characteristic of the point; and a traffic flow trend and a periodic characteristic which are obtained through analysis are fused, so that the prediction of traffic flow can be realized. With the method of the invention adopted, the defect of the incapability of an existing method to make full use of time and space characteristics can be eliminated, the time and space characteristics of the traffic flow are fully extracted, and the periodic characteristic of the data of the traffic flow is fused with the time and space characteristics, and therefore, the accuracy of short-term traffic flow prediction results can be improved.

The invention discloses a method and a system for enhancing features by the aid of bidirectional long-term and short-term memory recurrent neural networks. The method includes steps of training inputted features with noise and corresponding features without noise by the aid of bidirectional long-term and short-term memory recurrent neural network models; enhancing the features by the aid of the bidirectional long-term and short-term memory recurrent neural network models. The method and the system have the advantages that the long-term and short-term memory recurrent neural network models can be built in two directions by the aid of the inputted features, and accordingly models of contexts of current frames can be effectively built; long-term and short-term memory cells and certain control variables are introduced into the neural networks, and accordingly the models can be built in the dependence on the long contexts; excellent enhancement effects can be realized by the models owing to long-term dependence when the features are enhanced by the aid of the models, and accordingly the performance of voice recognition systems and audio event classification systems can be improved.

The invention discloses a method and a system for detecting voice spoofing attack of speakers on the basis of deep learning. The method includes constructing audio-frequency training sets, initializing deep feed-forward neural networks and deep recurrent neural networks and respectively training the deep feed-forward neural networks and the deep recurrent neural networks by the aid of multi-frame feature vectors and single-frame vector sequences of the training sets; respectively leading frame level and sequence level feature vectors of to-be-tested audio frequencies into two trained linear differential analysis models in test phases, weighting two obtained result grades to obtain scores and comparing the scores to predefined threshold values so as to discriminate voice spoofing. The method and the system have the advantages that local features can be captured, and global information can be grasped; the linear differential analysis models are used as classifiers in identification and verification phases, the voice spoofing attack can be judged by means of grade fusion, and accordingly the voice spoofing detection accuracy can be greatly improved.

Systems and processes for language identification using recurrent neural networks are provided. An example method includes, at an electronic device, receiving a first typed character of a character sequence and determining a character context of the first typed character based on the first typed character and a second typed character of the character sequence. The method further includes determining a confidence level that the character sequence is associated with a language of a plurality of languages based on the character context of the first typed character, and determining whether the confidence level exceeds a threshold, in accordance with a determination that the confidence level exceeds the threshold, providing the language as a candidate language, and in accordance with a determination that the confidence level does not exceed the threshold, forgoing providing the language as a candidate language.

Systems and processes for word encoding are provided. In accordance with one example, a method includes, at an electronic device with one or more processors and memory, receiving a user input, determining a first similarity between a representation of the user input and a first acoustic model of a plurality of acoustic models, and determining a second similarity between the representation of the user input and a second acoustic model of the plurality of acoustic models. The method further includes determining whether the first similarity is greater than the second similarity. In accordance with a determination that the first similarity is greater than the second similarity, the first acoustic model may be selected; and in accordance with a determination that the first similarity is not greater than the second similarity, the second acoustic model may be selected.

A method and apparatus for data communication in an oil well environment, wherein the method comprises detecting an acoustic signal transmitted along an acoustic channel, the acoustic signal being distorted from transmission through the acoustic channel, generating a transmitted data signal in response to the acoustic signal, inputting the transmitted data signal to an adaptive equalizer and adaptively equalizing the transmitted data signal to produce an equalized data signal related to the transmitted data signal by a mathematical function. The detecting step may include positioning an acoustic receiver in a communication unit along the acoustic channel. The communication unit may be positioned downhole and the adaptive equalizer may be positioned remotely relative to the communication unit or may be placed in the communication unit. The adaptive equalizer may be a frequency domain filter, a neural net adaptive equalizer or a nonlinear recurrent neural net equalizer. The acoustic signal may comprise a plurality of discrete transmissions which may be a training sequence for training the adaptive equalizer and may comprise a first discrete transmission transmitted repeatedly.The method of data communication in an oil well environment may comprise the steps of transmitting an acoustic signal from a first location along an acoustic channel, detecting the acoustic signal at a second location along the acoustic channel, generating a transmitted data signal in response to the acoustic signal, inputting the transmitted data signal to an adaptive equalizer and adaptively equalizing the transmitted data signal to produce an equalized data signal related to the transmitted data signal by a mathematical function. The transmitting step may further comprise positioning an acoustic transmitter in a first communication unit along the acoustic channel downhole or elsewhere. The method may further comprise acquiring data, generating an original data signal in response to the acquired data and inputting the original data signal to the acoustic transmitter. The acoustic signal may comprise a series of acoustic training signals for training the adaptive equalizer. The acoustic training signals may be transmitted at a predetermined time. A stored training signal may include a series of stored training data signals corresponding to the series of acoustic training signals. At least a portion of the stored training signals may be cross-correlated to the transmitted data signal. The acoustic signal may comprise a notification signal for notifying the adaptive equalizer of a training session.

Learning-type motion control is performed using a hierarchical recurrent neural network. A motion pattern provided through human teaching work is automatically time-serially segmented with the hierarchical recurrent neural network, and the motion control of a machine body is carried out with a combination of the segmented data, whereby various motion patterns can be produced. With the time-serial segmentation, local time-serial patterns and an overall pattern as a combination of the local time-serial patterns are produced. For those motion patterns, indices for static stability and dynamic stability of the machine body, e.g., ZMP stability criteria, are satisfied and hence control stability is ensured.

The invention discloses a behavior identification method based on a recurrent neural network and human skeleton movement sequences. The method comprises the following steps of normalizing node coordinates of extracted human skeleton posture sequences to eliminate influence of absolute space positions, where a human body is located, on an identification process; filtering the skeleton node coordinates through a simple smoothing filter to improve the signal to noise ratio; sending the smoothed data into the hierarchic bidirectional recurrent neural network for deep characteristic extraction and identification. Meanwhile, the invention provides a hierarchic unidirectional recurrent neural network model for coping with practical real-time online analysis requirements. The behavior identification method based on the recurrent neural network and the human skeleton movement sequences has the advantages of designing an end-to-end analyzing mode according to the structural characteristics and the motion relativity of human body, achieving high-precision identification and meanwhile avoiding complex computation, thereby being applicable to practical application. The behavior identification method based on the recurrent neural network and the human skeleton movement sequence is significant to the fields of intelligent video monitoring based on the depth camera technology, intelligent traffic management, smart city and the like.

The present disclosure generally relates to systems and processes for morpheme-based word prediction. An example method includes receiving a current word; determining a context of the current word based on the current word and a context of a previous word; determining, using a morpheme-based language model, a likelihood of a prefix based on the context of the current word; determining, using the morpheme-based language model, a likelihood of a stem based on the context of the current word; determining, using the morpheme-based language model, a likelihood of a suffix based on the context of the current word; determining a next word based on the likelihood of the prefix, the likelihood of the stem, and the likelihood of the suffix; and providing an output including the next word.

A dynamic bandwidth allocation method of an Ethernet passive optical network, comprises a predictor and a rule of QoS-promoted dynamic bandwidth allocation (PQ-DBA); the predictor predicts a client behavior and numbers of various kinds of packets by using a pipeline scheduling predictor consisted of a pipelined recurrent neural network (PRNN), and a learning rule of the extended recursive least squares (ERLS); the present invention establishes a better QoS traffic management for the OLT-allocated ONU bandwidth and client packets sent by priority.

Systems and processes for operating a digital assistant are provided. In accordance with one or more examples, a method includes, receiving training data for a data-driven learning network. The training data include a plurality of word sequences. The method further includes obtaining representations of an initial set of semantic categories associated with the words included in the training data; and training the data-driven learning network based on the plurality of word sequences included in the training data and based on the representations of the initial set of semantic categories. The training is performed using the word sequences in their entirety. The method further includes obtaining, based on the trained data-driven learning network, representations of a set of semantic embeddings of the words included in the training data; and providing the representations of the set of semantic embeddings to at least one of a plurality of different natural language processing tasks.

The present invention discloses a computer vision-based express parcel violent sorting identification method. The method comprises the following steps of: a depth camera-based pose estimation: a human pose estimation problem is converted into a problem of classifying depth image pixels captured by a depth camera, and human body pose estimation is realized by using a random forest method; human body three-dimensional pose relative spatial-temporal feature extraction: relative spatial-temporal positions of geometric elements of points, lines and surfaces formed by joints in three-dimensional poses and the measures of the change of the relative spatial-temporal positions are extracted and are adopted as the feature representations of the poses; and recurrent neural network-based violent sorting identification: modeling training is performed on the pose spatial-temporal relative features which are continuous in time and are extracted from the human body three-dimensional poses through a long and short memory model (LSTM), so that the identification of express parcel violent sorting behaviors can be realized.

The invention discloses a generation method of an image description from a structured text. The generation method comprises the steps of downloading pictures from the internet to form a picture training set; conducting morphological analysis on descriptions which correspond to the pictures in the picture training set to form the structured text; using an existing neural network model to extract convolution neural network characteristics of the pictures in the training set, and using <, picture characteristics and structured text < as inputs to form a multitasking recognition model; using the structured text extracted from the training set and a description which corresponds to the structured text as inputs of a recurrent neural network, and conducting training to obtain a parameter of a recurrent neural network model; inputting the convolution neural network characteristics of an image ready to be described, and obtaining a predicted structured text through the multitasking recognition model; inputting the predicted structured text, and obtaining the image description through the recurrent neural network model. Compared with the prior art, a better image description effect, accuracy and sentence variety can be generated through the method, and the generation method of the image description from the structured text can be effectively popularized in an application of image retrieval.

A spoken language understanding (SLU) system receives a sequence of words corresponding to one or more spoken utterances of a user, which is passed through a spoken language understanding module to produce a sequence of intentions. The sequence of words are passed through a first subnetwork of a multi-scale recurrent neural network (MSRNN), and the sequence of intentions are passed through a second subnetwork of the multi-scale recurrent neural network (MSRNN). Then, the outputs of the first subnetwork and the second subnetwork are combined to predict a goal of the user.

The invention provides a social emotion classification method based on multimodal fusion, which relates to information in the form of audio, visual and text. Most of affective computing research onlyextracts affective information by analyzing single-mode information, ignoring the relationship between information sources. The present invention proposes a 3D CNN ConvLSTM (3D CNN ConvLSTM) model forvideo information, and establishes spatio-temporal information for emotion recognition tasks through a cascade combination of a three-dimensional convolution neural network (C3D) and a convolution long-short-term memory recurrent neural network (ConvLSTM). For text messages, use CNN-RNN hybrid model is used to classify text emotion. Heterogeneous fusion of vision, audio and text is performed by decision-level fusion. The deep space-time feature learned by the invention effectively simulates the visual appearance and the motion information, and after fusing the text and the audio information,the accuracy of the emotion analysis is effectively improved.

Systems and methods for time series prediction are described. The systems and methods include encoding driving series into encoded hidden states, the encoding including adaptively prioritizing driving series at each timestamp using input attention, the driving series including data sequences collected from sensors. The systems and methods further includes decoding the encoded hidden states to generate a predicting model, the decoding including adaptively prioritizing encoded hidden states using temporal attention. The systems and methods further include generating predictions of future events using the predicting model based on the data sequences. The systems and methods further include generating signals for initiating an action to devices based on the predictions.

The invention discloses a hyper-spectral image classification method based on recurrent neural network with the object to solving the problems that in prior art, the input characteristic determination ability is weak and that the extraction of local spatial characteristics is not complete. The method comprises the following steps: 1) extracting the spatial texture characteristics and the sparse representation characteristics of a hyper-spectral image and piling and combining them as the low-level characteristics; 2) extracting from the low-level characteristics the sample local spatial sequence characteristics; 3) according to the local spatial sequence characteristics, creating a recurrent neural network model; and utilizing the training sample local spatial sequence characteristics to train the recurrent neural network model parameters; and 4) inputting the testing sample local spatial sequence characteristics into the well-trained recurrent neural network model; obtaining the highly abstract high-level semantic characteristics and obtaining the classification information of the testing sample. According to the deep learning method of the invention, the correct efficiency for hyper-spectral image classification is increased and the method can be used for vegetation investigation, disaster monitoring, map making and intelligence obtaining.