International Business Machines Corporation

Jianbin Tang

Jun Song Wang

Qing Fa Wang

Yong Hua Lin

Yonghua Lin

Umar Asif

Zhen Bo Zhu

Rong Yan

Stefan Harrer

Chao Zhu

Hai Zhan

Yini Wang

Ziyuan Wang

Feng Cao

Antonio Jose Jimeno Yepes

Qi Jun Wang

Stefan Von Cavallar

Xing Zhi Sun

Xin Hao Cheng

Subhrajit Roy

Yu Wang

Yu Meng Li

Arnon Amir

Hang Liu

Jeffrey L Rogers

Jun Chi Yan

Lenin Mehedy

Junsong Wang

Feng Juan Wang

Qian Zhao

Xi Liang

Xinhao Cheng

Erhan Bilal

Stefan Renard Maetschke

Tian Hao

Longxiang Gao

Deval Samirbhai Mehta

Chao E Zhang

Gandhi Sivakumar

Sarvesh S Patel

Kushal S Patel

Benjamin Scott Mashford

Filiz Isabell Kiral-Kornek

Peter Zhong

Xian Wei Yi

A first version of a model specified by a model execution request is executed, producing a first execution result. A second version of the model is selected according to an input data attribute specified by the model execution request. The second version of the model is executed, producing a first execution result. Using a natural language processing engine, responsive to the first execution result and the second execution result differing by more than a threshold amount, a natural language explanation of a difference between the first execution result and the second execution result is constructed.

Version based model result explainability

A method, a structure, and a computer system for privacy-preserving motion analysis. Embodiments may include identifying one or more joints of a user based on collected data and generating one or more 3D representations of the one or more joints of the user. Embodiments may further include anonymizing the one or more 3D representations, classifying one or more actions of the user based on the one or more 3D representations, wherein the classifying outputs an action score, and exporting at least one of the one or more actions and the action score.

Privacy-preserving motion analysis

A computer-implemented method of decoding video data in multiple resolution formats includes receiving an encoded video stream including a salient data and a non-salient data, the salient data having a higher resolution format than the non-salient data. The video stream is decoded into the non-salient data in a lower-resolution format and the salient data in the higher-resolution format. The non-salient data is reconstructed to a higher resolution format. The salient data and the reconstructed non-salient data are combined to form a video stream in the higher-resolution format of the salient data.

Video encoding through non-saliency compression for live streaming of high definition videos in low-bandwidth transmission

A method, a computer program product, and a computer system determine abnormal motion from a patient. The method includes receiving sensory data of the patient and a location in which the patient is present. The sensory data includes video data over a period of time the patient is being monitored. The method includes generating contextual information based on the sensory data. The contextual information is indicative of surroundings of the patient and characteristics of the location. The method includes generating motion information based on the sensory data. The motion information is indicative of movement of the patient in the location. The method includes generating contextual motion data by incorporating the contextual information with the motion information. The method includes determining the abnormal motion based on the contextual motion data.

Seizure detection using contextual motion

A computer-implemented method of encoding video streams for low-bandwidth transmissions includes identifying a salient data and a non-salient data in a high-resolution video stream. The salient data and the non-salient data is segmented. The non-salient data is compressed to a lower resolution. The salient data and the compressed non-salient data are transmitted in a low-bandwidth transmission.

A method, a structure, and a computer system for privacy-preserving motion analysis. Embodiments may include collecting data corresponding to a user with one or more sensors and identifying one or more joints of the user based on the data. Embodiments may additionally include generating one or more 3D representations of the one or more joints of the user and anonymizing the one or more 3D representations by applying thereto a joint-centering and a random shuffling. Embodiments may further include classifying one or more actions of the user based on analysing the one or more 3D representations, and exporting at least one of the data and the one or more actions.

Weighted population code in neuromorphic systems is provided. According to an embodiment, a plurality of input values is received. For each of the plurality of values, a plurality of spikes is generated. Each of the plurality of spikes has an associated weight. A consumption time is determined for each of the plurality of spikes. Each of the plurality of spikes is sent for consumption at its consumption time.

Efficient generation of stochastic spike patterns in core-based neuromorphic systems

Aspects described herein include a method of generating three-dimensional (3D) spikes. The method comprises receiving a signal comprising time-series data and generating a first two-dimensional (2D) grid. Generating the first 2D grid comprises mapping segments of the time-series data to respective positions of the first 2D grid, and generating, for each position, a spike train corresponding to the respective mapped segment. The method further comprises generating a second 2D grid including performing, for each position, a mathematical operation on the spike train of the corresponding position of the first 2D grid. The method further comprises generating a third 2D grid including performing spatial filtering on the positions of the second 2D grid. The method further comprises generating a 3D grid based on a combination of the first 2D grid, the second 2D grid, and the third 2D grid. The 3D grid comprises one or more 3D spikes.

Generating three-dimensional spikes using low-power computing hardware

Methods, systems and computer program products for automatically generating structured training data based on an unstructured document are provided. Aspects include receiving an unstructured document and a corresponding structured document that includes labeled portions. Aspects also include generating a parsed document that has one or more extracted objects by applying a parsing tool to the unstructured document. Aspects also include identifying one or more matching extracted objects by applying a matching algorithm to the structured document and the parsed document. Each matching extracted object is an extracted object of the parsed document that corresponds to a labeled portion of the structured document. Aspects also include annotating a region of the unstructured document that corresponds to the bounding box of the respective matching extracted object with a respective label of the corresponding labeled portion of the unstructured document.

Automated generation of structured training data from unstructured documents

One or both of epilepsy seizure detection and prediction at least by performing the following: running multiple input signals from sensors for epilepsy seizure detection through multiple classification models, and applying weights to outputs of each of the classification models to create a final classification output. The weights are adjusted to tune relative output contribution from each classifier model in order that accuracy of the final classification output is improved, while power consumption of all the classification models is reduced. One or both of epilepsy seizure detection and prediction are performed with the adjusted weights. Another method uses streams from sensors for epilepsy seizure detection to train and create the classification models, with fixed weights once trained. Information defining the classification models with fixed weights is communicated to wearable computer platforms for epilepsy seizure detection and prediction. The streams may be from multiple people and applied to an individual person.

Epilepsy seizure detection and prediction using techniques such as deep learning methods

Provided are embodiments for a computer-implemented method. The method includes receiving a sequence of image data, transforming objects in each frame of the sequence of the image data into direction vectors, and clustering the direction vectors based at least in part on features of the objects. The method also includes mapping the direction vectors for the objects in each frame into a position-orientation data structure, and performing tracking using the mapped direction vectors in the position-orientation data structure. Also provided are embodiments of a computer program product and a system for performing object tracking.

Memory-efficient video tracking in real-time using direction vectors

A method, system and computer program product for accelerating a deep neural network (DNN) in a field-programmable gate array (FPGA) are disclosed. The method includes receiving a DNN net file and weights, converting the received DNN net file to one or more source files, generating an executable FPGA bit file using the one or more source files, and downloading the executable FPGA bit file from the DNN conversion platform to the FPGA. Converting of the received DNN net file and the weights to the one or more source files can further include analyzing the DNN net file to identify a plurality of neural layers, decomposing one or more neural layers of the plurality of neural layers to one or more operation blocks, instantiating the one or more source files, based on the one or more operation blocks.

Accelerate deep neural network in an FPGA

A method for digital image classification and localization includes receiving a digital image of a biological organism from an imaging apparatus, the digital image comprising a plurality of intensities on a 2-dimensional grid of points, generating a plurality of discriminative representations of the 2D digital image by extracting dominant characteristics of the image from three different viewpoints, where the plurality of discriminative representations form a 3-dimensional digital image, combining the 3D digital image with the 2D digital image in a convolutional neural network that outputs a 3-channel feature map that localizes image abnormalities in each of the three channels and includes a detection confidence that each abnormalities is a neoplasm, providing the 3-channel feature map to a controller of a robotic surgical device where the robotic surgical device uses the 3-channel feature map to locate the neoplasm within the biological organism in a surgical procedure for treating the neoplasm.

Image feature classification and localization using discriminative representations for robotic surgical control

There is provided a system and method for providing privacy viewing of an output from a electronic display, including a first polarizer configured to polarize a display signal at a first polarization angle; a second polarizer configured to polarize a noise signal at a second polarization angle orthogonal to the first polarization angle; a combiner configured to combine the polarized display signal and the polarized noise signal, and a third polarizer configured to receive the output of the electronic display and polarize the combined polarized display and noise signal at the first polarization angle. The system includes first and second polarization rotators configured to rotate the first and second polarization angles of the combined polarized display and noise signals prior to being output by the electronic display and prior to being received by the third polarizer.

Method and apparatus of private display device

Growing community of inventors

Melbourne, Australia

Jianbin Tang