Techniques are described for content-adaptive downsampling of digital images and videos for computer vision operations, such as semantic segmentation. A computer vision system comprises a memory, one or more processors operably coupled to the memory and a downsampling module configured for execution by the one or more processors to perform, based on a non-uniform sampling model trained to predict content-aware sampling parameters, downsampling input image data to generate downsampled image data. A segmentation module is configured for execution by the one or more processors to perform segmentation on the downsampled image to produce a segmentation result, such as a feature map that assigns pixels of the downsampled image data to object classes. An upsampling module is configured for execution by the one or more processors to perform upsampling according to the segmentation result to produce upsampled image data.

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for generating rare pose data. One of the methods includes obtaining a three-dimensional model of a dynamic object, wherein the dynamic object has multiple movable elements that define a plurality of poses of the dynamic object. A plurality of template poses of the dynamic object are used to generate additional poses for the dynamic object including varying angles of one or more key joints of the dynamic object according to the three-dimensional model. Point cloud data is generated for the additional poses generated for the dynamic object.

An online system performs predictions for real-time tasks and near real-time tasks that need to be performed by a deadline. A client device receives a real-time machine learning based model associated with a measure of accuracy. If the client device determines that a task can be performed using predictions having less than the specified measure of accuracy, the client device uses the real-time machine learning based model. If the client device determines that a higher level of accuracy of results is required, the client device sends a request to an online system. The online system provides a prediction along with a string representing a rationale for the prediction.

An online system performs predictions for real-time tasks and near real-time tasks that need to be performed by a deadline. A client device receives a real-time machine learning based model associated with a measure of accuracy. If the client device determines that a task can be performed using predictions having less than the specified measure of accuracy, the client device uses the real-time machine learning based model. If the client device determines that a higher level of accuracy of results is required, the client device sends a request to an online system. The online system provides a prediction along with a string representing a rationale for the prediction.

State of the art systems and methods attempting to generate synthetic biosignals such as PPG generate patient specific PPG signatures and do not correlate with pathophysiological changes. Embodiments herein provide a method and system for generating synthetic time domain signals to build a classifier. The synthetic signals are generated using statistical explosion. Initially, a parent dataset of actual sample data of class and non-class subjects is identified, and statistical features are extracted. Kernel density estimate (KDE) is used to vary the feature distribution and create multiple data template from a single parent signal. PPG signal is again reconstructed from the distribution pattern using non-parametric techniques. The generated synthetic data set is used to build the two stage cascaded classifier to classify CAD and Non CAD, wherein the classifier design enables reducing bias towards any class.

This disclosure provides systems, methods and apparatuses for classifying traffic flow using a plurality of learning machines arranged in multiple hierarchical levels. A first learning machine may classify a first portion of the input stream as malicious based on a match with first classification rules, and a second learning machine may classify at least part of the first portion of the input stream as malicious based on a match with second classification rules. The at least part of the first portion of the input stream may be classified as malicious based on the matches in the first and second learning machines.

One example provides a computing system comprising a depth sensor comprising a plurality of pixels, and a storage machine holding instructions executable by a logic machine to, for each pixel, make K phase measurements to form a set of noisy phase measurements, determine a location at which a projection line that passes through the set of noisy phase measurements in a K-dimensional phase space passes through a lower dimensional plane, the projection line being parallel to a noise free phase evolution line, compare the location to a plurality of independent terms of a predetermined matrix of points in the lower dimensional plane, locate a corresponding set of noiseless phase orders by using a selected set of independent terms to reference a look-up table, determine a distance value for the pixel based upon the corresponding set of noiseless phase orders, and output the distance value for the pixel.

One example provides a computing system comprising a depth sensor comprising a plurality of pixels, and a storage machine holding instructions executable by a logic machine to, for each pixel, make K phase measurements to form a set of noisy phase measurements, determine a location at which a projection line that passes through the set of noisy phase measurements in a K-dimensional phase space passes through a lower dimensional plane, the projection line being parallel to a noise free phase evolution line, compare the location to a plurality of independent terms of a predetermined matrix of points in the lower dimensional plane, locate a corresponding set of noiseless phase orders by using a selected set of independent terms to reference a look-up table, determine a distance value for the pixel based upon the corresponding set of noiseless phase orders, and output the distance value for the pixel.

A method and an apparatus for length-aware local tiling in a sparse attention module in a transformer in heterogeneous devices are provided. The method includes that a heterogeneous device including one or more GPUs: divides a transformed sparsity mask into a plurality of first tiles and obtaining one or more effective first tiles from the plurality of first tiles, where each effective first tile includes at least one non-zero element; loads the one or more effective first tiles into a shared memory in the one or more GPUs and loads a plurality of elements in a first matrix corresponding to the one or more effective first tiles into the shared memory; and performs multiplication by a first sampled dense-dense matrix multiplication (SDDMM) kernel in the sparse attention module in the transformer by fetching the one or more effective first tiles and the plurality of elements from the shared memory.

In some aspects, systems and methods for efficiently clustering a large-scale dataset for improving the construction and training of machine-learning models, such as neural network models, are provided. A dataset used for training a neural network model configured can be clustered into a first set of clusters and a second set of clusters. The neural network model can be constructed with a number of nodes in a hidden layer that is based on the number of clusters in the first set of clusters. The neural network can be trained based on training samples selected from the second set of clusters. In some aspects, the trained neural network model can be utilized to satisfy risk assessment queries to compute output risk indicators for target entities. The output risk indicator can be used to control access to one or more interactive computing environments by the target entities.