Huawei Technologies Co., Limited

Natan Peterfreund

Itsik Dvir

Dror Irony

David Drezner

Zuguang Wu

Amiram Allouche

Ady Ecker

Roman Talyansky

Zach Melamed

Uri Verner

A distributed system for training a classifier is provided. The system comprises machine learning (ML) workers and a parameter server (PS). The PS is configured for parallel processing to provide the model to each of the ML workers, receive model updates from each of the ML workers, and iteratively update the model using each model update. The PS contains gradient datasets associated with a respective ML worker, for storing a model-update-identification (delta-M-ID) indicative of the computed model update and the respective model update, a global dataset that stores, the delta-M-ID, an identification of the ML worker (ML-worker-ID) that computed the model update, and a model version that marks a new model in PS that is computed from merging the model update with a previous model in PS; and a model download dataset that stores the ML-worker-ID and the model version of each transmitted model.

Systems and methods for fault tolerance recover during training of a model of a classifier using a distributed system

The invention relates to an image processing apparatus for compressing or decompressing a segment of an image. The segment includes a plurality of pixels, each pixel includes a pixel value and a pixel position defined by a first coordinate system. The pixel values of the plurality of pixels form a pixel value vector. The apparatus includes processing circuitry configured to compress and/or decompress the segment. Compressing the segment includes computing a plurality of expansion coefficients by expanding the pixel value vector into a plurality of basis vectors that are discrete approximations of solutions of a boundary value problem of the Helmholtz equation on the segment of the image in a second coordinate system rotated relative to the first coordinate system. Decompressing the segment includes computing the pixel value vector by forming a linear combination of the basis vectors using the plurality of expansion coefficients.

Image and video processing apparatuses and methods

There is provided an apparatus, methods, to generate a set of transform coefficients for coding a block in a frame or portion thereof, comprising: a media encoder coupled to a data interface adapted to receive a frame or portion thereof, the media encoder is adapted to: select a rotational symmetry mask for the block from a plurality of rotational symmetry masks which define a plurality of different rotational symmetries in a multi-dimensional space having a size and a shape as the block; split the block to two complementary portions based on the rotational symmetry mask; generate a pair of rotational symmetry blocks each having one of the two complementary portions; and compute at least one transform coefficient for at least one member of the pair of rotational symmetry blocks.

Systems and methods for processing a block of a digital image

The disclosure relates to an image processing apparatus for compressing or decompressing a segment of an image, the segment being non-rectangular and comprising a plurality of pixels, each pixel comprising a pixel value, the pixel values of the plurality of pixels forming a pixel value vector, the apparatus comprising: a processor configured to compress the segment or configured to decompress the segment, wherein compressing the segment comprises computing a plurality of expansion coefficients by expanding the pixel value vector into a plurality of basis vectors, wherein the basis vectors are discrete approximations of solutions of a boundary value problem of the Helmholtz equation on the segment of the image; and wherein decompressing the segment comprises computing the pixel value vector by forming a linear combination of the basis vectors according to the plurality of expansion coefficients.

There is provided an apparatus, methods, to generate a set of transform coefficients for coding a block in a frame or portion thereof, comprising: a media encoder coupled to a data interface adapted to receive a frame or portion thereof, the media encoder is adapted to: select a rotational symmetry mask for the block from a plurality of rotational symmetry masks which define a plurality of different rotational symmetries in a multi dimensional space having a size and a shape as the block; split the block to two complementary portions based on the rotational symmetry mask; generate a pair of rotational symmetry blocks each having one of the two complementary portions; and compute at least one transform coefficient for at least one member of the pair of rotational symmetry blocks.

There is provided an apparatus for generating a set of transform coefficients of a block in a frame or portion thereof. The apparatus includes a media encoder coupled to a data interface configured to receive a frame or portion thereof. The media encoder is configured to: designate a rotational symmetry mask having a size and a shape as the block partitioned in the frame or portion thereof for processing the block; split the block to two complementary portions using the rotational symmetry mask; generate a pair of rotational symmetry blocks each having one of the two complementary portions; and compute a transform coefficient for each member of the pair of rotational symmetry blocks.

Systems and methods for mask based processing of a block of a digital image

An apparatus adapted, including a media encoder coupled to a data interface adapted to receive a frame or portion thereof. The media encoder is adapted to iteratively subdivide each block of a plurality of blocks partitioned in a frame or portion thereof and to process a subdivided block of the plurality of blocks, during each iteration of a plurality of iterations, by selecting a rotational symmetry mask fulfilling an efficiency measure for coding said block, the rotational symmetry mask selected from a plurality of rotational symmetry masks which define a plurality of different rotational symmetries in a multidimensional space, splitting the subdivided block to two complementary portions based on the rotational symmetry mask, generating a pair of rotational symmetry blocks each having one of said two complementary portions, and computing a plurality of transform coefficients from the pair of rotational symmetry blocks.

Growing community of inventors

Munich, Germany

Natan Peterfreund