Abstract: Techniques are described for a time series probabilistic forecasting framework that combines recurrent neural networks (RNNs) with a flexible, nonparametric representation of the output distribution. The representation is based on the nonparametric quantile function (instead of, for example, a parametric density function) and is trained by minimizing a continuous ranked probability score (CRPS) derived from the quantile function. Unlike methods based on parametric probability density functions and maximum likelihood estimation, the techniques described herein can flexibly adapt to different output distributions without manual intervention. Furthermore, the nonparametric nature of the quantile function provides a significant boost in the approach's robustness, making it more readily applicable to a wide variety of time series datasets.

Abstract: The present disclosure relates to a content recommendation method, device and system. The method includes: recommending a content in a content set to a user based on an average recommendation probability; collecting feedback information on the recommended content from the user's client, wherein the feedback information includes display information and click information, the display information including displaying times and displaying timing of the recommended content on the client, and the click information including clicking times and clicking timing of the recommended content on the client; and determining a sequence of preferred contents from the contents in the content set according to the feedback information, so as to recommend a content to the user based on the sequence of preferred contents.

Abstract: Systems, methods, tangible non-transitory computer-readable media, and devices for object detection, tracking, and motion prediction are provided. For example, the disclosed technology can include receiving sensor data including information based on sensor outputs associated with detection of objects in an environment over one or more time intervals by one or more sensors. The operations can include generating, based on the sensor data, an input representation of the objects. The input representation can include a temporal dimension and spatial dimensions. The operations can include determining, based on the input representation and a machine-learned model, detected object classes of the objects, locations of the objects over the one or more time intervals, or predicted paths of the objects. Furthermore, the operations can include generating, based on the input representation and the machine-learned model, an output including bounding shapes corresponding to the objects.

Abstract: Methods and systems for training a machine learning algorithm (MLA) comprising: acquiring a first set of training samples having a plurality of features, iteratively training a first predictive model based on the plurality of features and generating a respective first prediction error indicator. Analyzing the respective first prediction error indicator for each iteration to determine an overfitting point, and determining at least one evaluation starting point. Acquiring an indication of a new set of training objects, and iteratively retraining the first predictive model with at least one training object from the at least one evaluation starting point to obtain a plurality of retrained first predictive models and generating a respective retrained prediction error indicator. Based on a plurality of retrained prediction error indicators and a plurality of the associated first prediction error indicators, selecting one of the first set of training samples and the at least one training object.

Abstract: The present teaching generally relates to removing perturbations from predictive scoring. In one embodiment, data representing a plurality of events detected by a content provider may be received, the data indicating a time that a corresponding event occurred and whether the corresponding event was fraudulent. First category data may be generated by grouping each event into one of a number of categories, each category being associated with a range of times. A first measure of risk for each category may be determined, where the first measure of risk indicates a likelihood that a future event occurring at a future time is fraudulent. Second category data may be generated by processing the first category data and a second measure of risk for each category may be determined. Measure data representing the second measure of risk for each category and the range of times associated with that category may be stored.

Abstract: A technique to prune weights of a neural network using an analytic threshold function h(w) provides a neural network having weights that have been optimally pruned. The neural network includes a plurality of layers in which each layer includes a set of weights w associated with the layer that enhance a speed performance of the neural network, an accuracy of the neural network, or a combination thereof. Each set of weights is based on a cost function C that has been minimized by back-propagating an output of the neural network in response to input training data. The cost function C is also minimized based on a derivative of the cost function C with respect to a first parameter of the analytic threshold function h(w) and on a derivative of the cost function C with respect to a second parameter of the analytic threshold function h(w).