A computer-implement method comprises: training a classifier with labeled data from a dataset; classifying, by the trained classifier, unlabeled data from the dataset; providing, by the classifier to a policy gradient, a reward signal for each data/query pair; transferring, by the classifier to a ranker, learning; training, by the policy gradient, the ranker; ranking data from the dataset based on a query; and retrieving data from the ranked data in response to the query.

A model learning device provided with: an error-added movement locus generation unit for adding an error to movement locus data for action learning that represents the movement locus of a subject and to which is assigned an action label that is information representing the action of the subject, and thereby generating error-added movement locus data; and an action recognition model learning unit for learning a model, using at least the error-added movement locus data and learning data created on the basis of the action label, by which model the action of some subject can be recognized from the movement locus of the subject. Thus, it is possible to provide a model by which the action of a subject can be recognized with high accuracy on the basis of the movement locus of the subject estimated using a camera image.

A method of continuous decoding of motion for a direct neural interface. The method of decoding estimates a motion variable from an observation variable obtained by a time-frequency transformation of the neural signals. The observation variable is modelled using a HMM model whose hidden states include at least an active state and an idle state. The motion variable is estimated using a Markov mixture of experts where each expert is associated with a state of the model. For a sequence of observation vectors, the probability that the model is in a given state is estimated, and from this a weighting coefficient is deduced for the prediction generated by the expert associated with this state. The motion variable is then estimated by combination of the estimates of the different experts with these weighting coefficients.

According to an embodiment, a pattern recognition device recognizes a pattern of an input signal by converting the input signal to a feature vector and matching the feature vector with a recognition dictionary. The recognition dictionary includes a dictionary subspace basis vector for expressing a dictionary subspace which is a subspace of a space of the feature vector, and a plurality of probability parameters for converting similarity calculated from the feature vector and the dictionary subspace into likelihood. The device includes a recognition unit configured to calculate the similarity using a quadratic polynomial of a value of an inner product of the feature vector and the dictionary subspace basis vector, and calculate the likelihood using the similarity and an exponential function of a linear sum of the probability parameters. The recognition dictionary is trained by using an expectation maximization method using a constraint condition between the probability parameters.

The invention discloses a non-invasive load decomposition method, which includes: step 1, obtaining the power fingerprint information of each load; step 2, clustering the operating state of loads through the clustering algorithm, calculate statistical values of each cluster, and encoding the operating state of electrical appliances; step 3, establishing a hidden Markov model with multiple-parameters and calculating the model parameters; step 4, performing state recognition based on Viterbi algorithm and obtaining predicted state sequence; step 5, according to the predicted state sequence and the statistical values of each cluster, decomposing the load power based on the maximum likelihood estimation principle; step 6, outputting the state sequence and power decomposition results. The invention solves the conventional load identification algorithm problems, such as complex model, insufficient use of electrical features and low accuracy of unknown information.

A method for recognizing handwritten text is disclosed. The method comprises receiving data comprising a sequence of ink points; applying the received data to a neural network-based sequence classifier trained with a Connectionist Temporal Classification (CTC) output layer using forced alignment to generate an output; generating a character hypothesis as a portion of the sequence of ink points; applying the character hypothesis to a character classifier to obtain a first probability corresponding to the probability that the character hypothesis includes the given character; processing the output of the CTC output layer to determine a second probability corresponding to the probability that the given character is observed within the character hypothesis; and combining the first probability and the second probability to obtain a combined probability corresponding to the probability that the character hypothesis includes the given character.

Systems and methods are disclosed to objectively identify sub-second behavioral modules in the three-dimensional (3D) video data that represents the motion of a subject. Defining behavioral modules based upon structure in the 3D video data itself—rather than using a priori definitions for what should constitute a measurable unit of action—identifies a previously-unexplored sub-second regularity that defines a timescale upon which behavior is organized, yields important information about the components and structure of behavior, offers insight into the nature of behavioral change in the subject, and enables objective discovery of subtle alterations in patterned action. The systems and methods of the invention can be applied to drug or gene therapy classification, drug or gene therapy screening, disease study including early detection of the onset of a disease, toxicology research, side-effect study, learning and memory process study, anxiety study, and analysis in consumer behavior.

Computer vision systems and methods for text classification are provided. The system detects a plurality of text regions in an image and generates a bounding box for each detected text region. The system utilizes a neural network to recognize text present within each bounding box and classifies the recognized text, based on at least one extracted feature of each bounding box and the recognized text present within each bounding box, according to a plurality of predefined tags. The system can associate a key with a value and return a key-value pair for each predefined tag.

A method for predicting user purchase by a user of a first site includes: selecting a distribution representing a probability distribution (PD) of inter-purchase-times (IPTs) across the first site and a second other site for each user, assigning each purchase of each user to one of the first site and the second site according to a Stochastic model, combining the selected PD with the Stochastic model to generate a PD of IPTs for only the first online site, estimating parameters of the probability distribution of IPTs for the first site by applying a Statistical modeling approach to features of each user, applying a sequence of observed IPTs of a given user for the first site and the parameters of the given user to the selected distribution to generate a probability, and determining whether the next purchase occurs on the second site based on the probability.

A method for continuous acoustic signature recognition and classification includes a step of obtaining an audio input signal from a resonant microphone array positioned proximate to a target, the audio input signal having a plurality of channels. The target produces characterizing audio signals depending on a state or condition of the target. A plurality of features is extracted from the audio input signal with a signal processor. The plurality of features is classified to determine the state of the target. An acoustic monitoring system implementing the method is also provided.