Kabushiki Kaisha Toshiba

Aftab Khan

Usman Raza

Shengyang Li

Timothy David Farnham

Nan Jiang

Adnan Aijaz

Parag Kulkarni

Peizheng Li

Theo Spyridopoulos

Abdussalam Ahmed Elturki

Michael Jones

Marko Dogramadzi

Pietro E Carnelli

A method for communicating a plurality of numerical parameter updates of a machine learning model from a first node to a second node. The method includes dividing each of the parameter updates into a respective primary segment and one or more respective additional segments, wherein the primary segment of each parameter update is the segment that has the greatest influence on the value of that parameter update. The method further includes constructing primary packet containing the primary segments of each of the plurality of parameter updates, and one or more additional packets including the one or more additional segments of the plurality of parameter updates. The method further includes transmitting the plurality of packets from the first node, wherein the primary packet is transmitted with a higher priority than any of the one or more additional packets.

Communicating machine learning model parameters

A device for optimizing model performance includes: a memory resource storing a local model on an electronic device; and a first processor. The local model is obtained by use of a machine learning process. The first processor: obtains data at the electronic device; applies the local model to the data or further data derived from the data to obtain at least one local model result; generates a measure of the quality of the local model result, and sends the data, further data derived from the data and/or the local model result through a communication device to a second processor for use in updating a global model based on the measure of the quality of the local model result.

Device and method for optimising model performance

A method for detecting an attack on a distributed artificial intelligence deployment comprising a plurality of worker devices. Each of the plurality of worker devices comprises a local machine learning model. Each local machine learning model comprises a plurality of layers. The method comprises calculating a first inference from first input data using a first machine learning model comprising layers of the plurality of layers of one or more of the local machine learning models and calculating additional inferences from the first input data using one or more additional machine learning models. Each of the additional machine learning models comprises at least one of the layers used in the first machine learning model and at least one layer from the pluralities of layers of the one or more local machine learning models that is not used by the first machine learning model. The method further comprises calculating differences between the first inference and each of the one or more additional inferences.

Detection of model attacks in distributed AI

A computer implemented method of distributed learning in a system comprising a parameter server configured to maintain a global parameter set of a model to be trained and a plurality of workers. The method comprises transmitting a current global parameter set to a worker. The worker performs a training step based on training data available to the worker, thereby generating a local set of parameters of the model, determines a likelihood of the local set of parameters being suitable for improving the global parameter set and omits transmission of the local set of parameters to the parameter server if it is determined that the local set of parameters is not likely suitable for improving the global parameter set.

System and method for distributed learning

According to an embodiment there is provided a method for managing wireless communication between a sending wireless device and a receiving wireless device in a wireless network, the method comprising the sending wireless device: transmitting to the receiving wireless device one or more uplink communications using a first transmission setting; determining whether a link quality of the communication link falls outside of a predefined range; and in response to determining that the link quality falls outside of the predefined range: determining at least first and second probing transmission settings that differ from each other and from the first transmission setting; transmitting a first probing transmission using the first probing transmission setting and a second probing transmission using the second probing transmission setting; receiving from the receiving wireless device a selection of one of the first and second probing transmission settings; and adopting the selected probing transmission setting for use with future transmissions to the receiving wireless device.

Transmission parameter probing and searching for agile link quality adaptation

A neural network system for inferring a location of a target from a plurality of localization parameters derived from a wireless signal and a method of training thereof. The neural network system comprises first and second neural networks. The plurality of localization parameters comprise one or more parameters relating to a velocity of a target and one or more other parameters. The first neural network is trained to infer a set of candidate locations of a target from values of the one or more other parameters. The second neural network is trained to infer a location of the target from values of the one or more parameters relating to a velocity of the target and a set of candidate locations of the target.

Neural network localization system and method

The present disclosure provides an error detector for determining an error vector between a radio trajectory and an image trajectory. The error detector includes: an input for monitoring a radio trajectory of an object from a radio signal and an image trajectory of an object from an image over an observation area; a correlation module arranged to correlate the radio trajectory with the image trajectory; an error module arranged to determine an error vector between the radio trajectory and the image trajectory; and an output arranged to transmit the error vector for use in determining an estimated trajectory of a target based on a target trajectory from a radio signal.

Compensating radio tracking with comparison to image based tracking

A method for operating a real-time control system comprising a first system configured to generate an information signal and a second system configured to use the information signal, wherein the second system comprises a first buffer for storing a previously received information signal. The method comprising: transmitting, by the first system, a first communication packet comprising the first information signal and generating, by the second system, a predicted first information signal for use in the first time slot. The predicted first information signal being generated by: retrieving the previously received information signal from the first buffer; generating a first prediction using a short-term predictor; and concurrently generating a second prediction using a long-term predictor; and setting the predicted first information signal equal to the first prediction unless the second prediction is available. The method further comprising using the predicted first information signal in response to determining that the first information signal was not received by the second system in the first time slot.

Edge-intelligence for stability guaranteed real-time control systems

A method of operating a network comprising an edge node and a server. The method comprises obtaining, by the edge node, a plurality of data samples, determining, by the edge node, a plurality of output labels by applying a first machine learning model using an input memory having a first input memory size to the plurality of data samples, calculating, by the edge node, an error term based on the confidence score of a first output label from the plurality of output labels, determining, by the edge node, based on the error term, whether to modify the first input memory size of the machine learning model and, if so, generating a second machine learning model based on the first machine learning model and a second input memory size.

Method for adaptive context length control for on-line edge learning

According to an embodiment there is provided a method of skill classification comprising receiving data indicative of an activity performed by a person, classifying the type or types of activity performed by the person based on the received data, wherein classifying provides an indication of an activity type or activity types as well as an indication of the confidence that an activity has been classified correctly and classifying a skill level associated with a classified activity or classified activities on the basis of the indication of confidence.

System and method for generalized skill assessment using activity data

A system and computer-implemented method for controlling transmission settings for one or more wireless devices in a wireless network. The method comprises: monitoring uplink communications from a sending wireless device to a receiving wireless device; determining one or more link quality metrics comprising a frame loss ratio and one or more signal strength metrics for the uplink communications; determining a link quality between the sending wireless device and the receiving wireless device based on the link quality metrics; determining whether the link quality falls outside of a predefined range; and in response to determining that the link quality falls outside of the predefined range, sending an instruction to the sending wireless device to update one or more transmission parameters to adjust a link budget for future transmissions by the sending wireless device to the receiving wireless device.

Link quality awareness for uplink transmissions from nodes to gateways

According to an embodiment there is provided a method of radio resource allocation performed in a gateway or network server communicatively connected to a plurality of wireless nodes. The method comprises computing, for each spreading factor out of a number of spreading factors, a probability of successful receipt of an uplink communication from a wireless node from the plurality of wireless nodes based on spreading factor settings of the plurality of wireless nodes, determining, for a given link level budget, a spreading factor that provides a maximum likelihood of successfully completing uplink communications and determining, for the determined spreading factor, a transmission power that provides the given link level budget.

Collision aware radio resource allocation

A computer implemented method of matching a series of geolocation measurements to a map using a Viterbi algorithm by determining transition probabilities between pairs of candidate points for consecutive geolocation measurements. The map comprises a plurality of paths. The method comprises segmenting the map into a plurality of segments and, when a distance of the consecutive geolocation measurements is less than or equal to a segment length, determining the transition probability between two candidate points based on a Euclidian distance between the two candidate points and a route distance between the two candidate points. The route distance is equal to the Euclidean distance plus a first predetermined value when the two candidate points are not in the same segment as each other.

Growing community of inventors

Bristol, United Kingdom

Aftab Khan