Microsoft Technology Licensing, LLC

Reinhard Sebastian Bernhard Nowozin

Pushmeet Kohli

Toby Leonard Sharp

Antony Ian Taylor Rowstron

Carsten Curt Eckard Rother

Benn Charles Thomsen

Ioan Alexandru Stefanovici

Alexander Lloyd Gaunt

Jamie Daniel Joseph Shotton

Andrew William Fitzgibbon

Antonio Criminisi

Aaron William Ogus

Andreas Georgiou

Richard John Black

Austin N Donnelly

John Michael Winn

David Lara Saucedo

James Hilton Clegg

Patrick Neil Anderson

Rokas Drevinskas

Ariel Gomez Diaz

Cheng Zhang

Sebastian Tschiatschek

Serguei Anatolievitch Legtchenko

Jeremy Martin Jancsary

Douglas Wayne Phillips

Anja Thieme

Hugh David Paul Williams

Jose Miguel Hernandez Lobato

Andromachi Chatzieleftheriou

John Christopher Dainty

Jeremy Jancsary

Danielle Charlotte Mary Belgrave

Konstantina Palla

Chao Ma

Ameera Patel

Iain Edward Buchan

Raluca Andreea Diaconu

Mengyang Yang

Shahram Izadi

Bruce Justin Lindbloom

Jonathan James Taylor

Danyal Khashabi

Bangpeng Yao

Po-Ling Loh

Kevin Schelten

Examples are disclosed that relate to encoding data on a data-storage medium. The method comprises obtaining a representation of a measurement performed on the data-storage medium, the representation being based on a previously recorded pattern of data encoded in the data-storage medium in a layout that defines a plurality of data locations. The method further comprises inputting the representation into a data decoder comprising a trained machine-learning function, and obtaining from the data decoder, for each data location of the layout, a plurality of probability values, wherein each probability value is associated with a corresponding data value and represents the probability that the corresponding data value matches the actual data value in the previously recorded pattern of data at a same location in the layout.

Machine-learning optimization of data reading and writing

A computer-implemented method comprising: outputting questions to a user via one or more user devices, and receiving back responses to some of the questions from the user via one or more user devices; over time, controlling the outputting of the questions so as to output the questions under circumstances of different values for each of one or more items of metadata, wherein the one or more items of metadata comprise one or more physical conditions of the user; monitoring whether or not the user responds when the question is output with the different metadata values; training the machine learning algorithm to learn a value of each of the items of metadata which optimizes a reward function, and based thereon selecting a circumstance when the user is exhibiting a particular physical condition to output subsequent questions.

Gathering data in a communication system

A computer-implemented method comprising: outputting questions to a user via one or more user devices, and receiving back responses to some of the questions from the user via one or more user devices; over time, controlling the outputting of the questions so as to output the questions under circumstances of different values for each of one or more items of metadata, wherein the one or more items of metadata comprise at least a time and/or a location at which a question was output to the user via the one or more user devices; monitoring whether or not the user responds when the question is output with the different metadata values; training the machine learning algorithm to learn a value of each of the items of metadata which optimizes a reward function, and based thereon selecting a time and/or location at which to output subsequent questions.

Examples are disclosed that relate to reading stored data. The method comprises obtaining a representation of a measurement performed on a data-storage medium, the representation being based on a previously recorded pattern of data encoded in the data-storage medium in a layout that defines a plurality of data locations. The method further comprises inputting the representation into a data decoder comprising a trained machine-learning function, and obtaining from the data decoder, for each data location of the layout, a plurality of probability values, wherein each probability value is associated with a corresponding data value and represents the probability that the corresponding data value matches the actual data value in the previously recorded pattern of data at a same location in the layout.

Memory facilitation using directed acyclic graphs is described, for example, where a plurality of directed acyclic graphs are trained for gesture recognition from human skeletal data, or to estimate human body joint positions from depth images for gesture detection. In various examples directed acyclic graphs are grown during training using a training objective which takes into account both connection patterns between nodes and split function parameter values. For example, a layer of child nodes is grown and connected to a parent layer of nodes using an initialization strategy. In examples, various local search processes are used to find good combinations of connection patterns and split function parameters.

Memory facilitation using directed acyclic graphs

A data-storage system comprises a head receiver configured to variably receive up to a number M of write heads. The data-storage system also includes an installed number N of write heads arranged in the head receiver, a substrate receiver configured to receive one or more data-storage substrates, and a positioner machine configured to adjust a relative placement of each of the M write heads with respect to at least one of the one or more data-storage substrates.

Disaggregated data storage and retrieval system

Resource allocation for machine learning is described such as for selecting between many possible options, for example, as part of an efficient training process for random decision tree training, for selecting which of many families of models best describes data, for selecting which of many features best classifies items. In various examples samples of information about uncertain options are used to score the options. In various examples, confidence intervals are calculated for the scores and used to select one or more of the options. In examples, the scores of the options may be bounded difference statistics which change little as any sample is omitted from the calculation of the score. In an example, random decision tree training is made more efficient while retaining accuracy for applications not limited to human body pose detection from depth images.

Resource allocation for machine learning

Model fitting from raw time of flight image data is described, for example, to track position and orientation of a human hand or other entity. In various examples, raw image data depicting the entity is received from a time of flight camera. A 3D model of the entity is accessed and used to render, from the 3D model, simulations of raw time of flight image data depicting the entity in a specified pose/shape. The simulated raw image data and at least part of the received raw image data are compared and on the basis of the comparison, parameters of the entity are computed.

Model fitting from raw time-of-flight images

Blind image deblurring with a cascade architecture is described, for example, where photographs taken on a camera phone are deblurred in a process which revises blur estimates and estimates a blur function as a combined process. In various examples the estimates of the blur function are computed using first trained machine learning predictors arranged in a cascade architecture. In various examples a revised blur estimate is calculated at each level of the cascade using a latest deblurred version of a blurred image. In some examples the revised blur estimates are calculated using second trained machine learning predictors interleaved with the first trained machine learning predictors.

Blind image deblurring with cascade architecture

Image restoration cascades are described, for example, where digital photographs containing noise are restored using a cascade formed from a plurality of layers of trained machine learning predictors connected in series. For example, noise may be from sensor noise, motion blur, dust, optical low pass filtering, chromatic aberration, compression and quantization artifacts, down sampling or other sources. For example, given a noisy image, each trained machine learning predictor produces an output image which is a restored version of the noisy input image; each trained machine learning predictor in a given internal layer of the cascade also takes input from the previous layer in the cascade. In various examples, a loss function expressing dissimilarity between input and output images of each trained machine learning predictor is directly minimized during training. In various examples, data partitioning is used to partition a training data set to facilitate generalization.

Image restoration cascade

Improved decision tree training in machine learning is described, for example, for automated classification of body organs in medical images or for detection of body joint positions in depth images. In various embodiments, improved estimates of uncertainty are used when training random decision forests for machine learning tasks in order to give improved accuracy of predictions and fewer errors. In examples, bias corrected estimates of entropy or Gini index are used or non-parametric estimates of differential entropy. In examples, resulting trained random decision forests are better able to perform classification or regression tasks for a variety of applications without undue increase in computational load.

Decision tree training in machine learning

Image demosaicing is described, for example, to enable raw image sensor data, where image elements have intensity values in only one of three color channels, to be converted into a color image where image elements have intensity values in three color channels. In various embodiments a trained machine learning component is used to carry out demosaicing optionally in combination with denoising. In some examples the trained machine learning system comprises a cascade of trained regression tree fields. In some examples the machine learning component has been trained using pairs of mosaiced and demosaiced images where the demosaiced images have been obtained by downscaling natural color digital images. For example, the mosaiced images are obtained from the demosaiced images by subsampling according to one of a variety of color filter array patterns.

Image demosaicing

A tractable model solves certain labeling problems by providing potential functions having arbitrary dependencies upon an observed dataset (e.g., image data). The model uses decision trees corresponding to various factors to map dataset content to a set of parameters used to define the potential functions in the model. Some factors define relationships among multiple variable nodes. When making label predictions on a new dataset, the leaf nodes of the decision tree determine the effective weightings for such potential functions. In this manner, decision trees define non-parametric dependencies and can represent rich, arbitrary functional relationships if sufficient training data is available. Decision trees training is scalable, both in the training set size and by parallelization. Maximum pseudolikelihood learning can provide for joint training of aspects of the model, including feature test selection and ordering, factor weights, and the scope of the interacting variable nodes used in the graph.

Decision tree fields to map dataset content to a set of parameters

Growing community of inventors

Cambridge, United Kingdom

Reinhard Sebastian Bernhard Nowozin