Resumen de: WO2022161891A1

The invention relates to a method for determining the architecture of an encoder (2) of a convolutional neural network (1), the neural network (1) being configured to process multiple different image processing tasks (t1, t2, t3), the method comprising the steps of: - for each image processing task (t1, t2, t3)), calculating characteristic scale distribution based on training data (S10); - generating multiple encoder architecture candidates, each encoder architecture of said encoder architecture candidates comprising at least one shared encoder layer (SL) which provides computational operations for multiple image processing tasks and multiple branches (b1, b2, b3) which span over one or more encoder layers which provide at least partly different computational operations for said image processing tasks, wherein each branch (b1, b2, b3) is associated with a certain image processing task (t1, t2, t3) (S11); - calculating receptive field sizes of the encoder layers of said multiple encoder architectures (S12); - calculating multiple assessment measures, each assessment measure referring to a combination of a certain encoder architecture of said multiple encoder architectures and a certain image processing task (t1, t2, t3), each assessment measure including information regarding the quality of matching of characteristic scale distribution of the image processing task (t1, t2, t3) associated with the assessment measure to the receptive field sizes of the encoder layers of the enco

Resumen de: US2022244684A1

A continuous-time recurrent neural network (CTRNN) is described that exploits the nonlinear dynamics of micro-electro-mechanical system (MEMS) devices to model a neuron in accordance with a neuron rate model that is the basis for dynamic field theory. Each MEMS device in the CTRNN is configured to simulate a neuron population by exploiting the characteristics of bi-stability and hysteresis inherent in certain MEMS device structures. In an embodiment, the MEMS device is a microbeam or cantilevered microbeam device that is excited with an alternating current (AC) voltage at or near an electrical resonance frequency associated with the MEMS device. In another embodiment, the MEMS device is an arched microbeam device that is excited with a direct current voltage and exhibits snap-through behavior due to the physical design of the structure. A CTRNN can be implemented using a number of MEMS devices that are interconnected, the connections associated with varying connection coefficients.

Resumen de: WO2022165366A1

Examples of a system and methods for quantifying patient improvement via artificial intelligence are disclosed. In general, via at least one processing element, a machine learning model such as a Siamese neural network is trained in view of a cost function to learn on average a maximum difference in outcomes between a patient at different points in time. Given the architecture of the neural network, a plurality of outcome measures generated for a given point in time can be condensed into a single score.

Resumen de: WO2022163972A1

In an operation method for a video-based behavior recognition device, according to an embodiment of the present invention, a synthesized channel frame provision unit may generate highlight information by comparing channel frames corresponding to the respective channels among a plurality of channels, and synthesize the channel frames and the highlight information to provide a synthesized channel frame. A neural network unit may provide a middle frame on the basis of the synthesized channel frame and a multi-frame convolution neural network. A behavior recognition result provision unit may provide a behavior recognition result on the basis of the middle frame and a weighted value generated according to the middle frame. In the operation method for a video-based behavior recognition device, according to the present invention, the behavior recognition result is provided on the basis of the multi-frame convolution neural network and the synthesized channel frame synthesized from the channel frames provided for the respective channels, and, thereby, an event occurence in a video can be more effectively detected.

Resumen de: US2022245488A1

This disclosure provides an accurate and personalized recommendation method based on a knowledge graph, which includes following steps: acquiring relevant knowledge of objects from a knowledge base according to historical behaviors of a user, and constructing a knowledge graph; initializing a vector representation of each node and its connection, and determining a receptive field of the node; generating training samples according to the historical behaviors of the user, and initializing a vector representation of all users and objects; acquiring a receptive field of an entity in the knowledge graph corresponding to the object in the training sample, then inputting the receptive field and the training sample to a graph neural network model to obtain predicted values of a possibility of an interaction between the user and the object. According to the disclosure, a sparsity of the historical behavior information of the original user is compensated with the knowledge graph information, and the user and objects are depicted in multi-dimension, so that the personalized recommendation is more accurate.

Resumen de: US2022245458A1

Disclosed herein are an apparatus and method for converting a neural network. The method includes separating neural network data of a source framework to form a tree structure by analyzing the same, converting the neural network data in a tree structure to a neural network optimized for a target framework, classifying training data based on the result of analysis of the neural network data of the source framework, converting the classified training data to the training data structure of the target framework, and creating a neural network and training data of the target framework by combining the converted neural network and the converted training data.

Resumen de: US2022245447A1

Systems and methods are provided for quantization aware training of a neural network for heterogeneous hardware platform. In the method, the system acquires hardware profiles with respect to a plurality of hardware components of a heterogeneous hardware platform. The system determines a plurality of hardware configurations based on the hardware profiles. The system acquires a set of training data and performing a quantization aware training using the training data on a network model based on the hardware configurations. The system obtains the network model with model weights for the heterogeneous hardware platform.

Resumen de: US2022245440A1

A method of using a computing device to train a neural network to recognize features in variate time series data that includes receiving, by a computing device, variate time series data. The computing device further receives results associated with the variate time series data. The computing device determines an anchor of the variate time series data. The computing device additionally determines one or more portions of the variate time series data which lead to a positive result. The computing device further determines one or more portions of the variate time series data which lead to a negative result. The computing device trains a neural network to interpret results of future variate time series data based upon the anchor, the one or more portions of the variate time series data which lead to the positive result, and the one or more portions of the variate time series data which lead to the negative result.

Resumen de: US2022245457A1

Various embodiments include methods and devices for neural network pruning. Embodiments may include receiving as an input a weight tensor for a neural network, increasing a level of sparsity of the weight tensor generating a sparse weight tensor, updating the neural network using the sparse weight tensor generating an updated weight tensor, decreasing a level of sparsity of the updated weight tensor generating a dense weight tensor, increasing the level of sparsity of the dense weight tensor the dense weight tensor generating a final sparse weight tensor, and using the neural network with the final sparse weight tensor to generate inferences. Some embodiments may include increasing a level of sparsity of a first sparse weight tensor generating a second sparse weight tensor, updating the neural network using the second sparse weight tensor generating a second updated weight tensor, and decreasing the level of sparsity the second updated weight tensor.

Resumen de: US2022245387A1

A method for semantic keypoint detection is described. The method includes linking, using a keypoint graph neural network (KGNN), semantic keypoints of an object within a first image of a video stream into a 2D graph structure corresponding to a category of the object. The method also includes embedding descriptors within the semantic keypoints of the 2D graph structure corresponding to the category of the object. The method further includes tracking the object within subsequent images of the video stream using the embedded descriptors within the semantic keypoints of the 2D graph structure corresponding to the category of the object.

Resumen de: US2022245386A1

A system comprises a processor and a memory storing instructions. The processor receives an image for processing using a reinforcement learning based agent comprising a neural network trained using a reward function. The image includes N lane lines of a roadway, where N is a positive integer. The instructions configure the processor to traverse the image using the agent at least N times from a first end of the image to a second end of the image by: incrementally moving the agent from a first side of the image to a second side of the image after each traversal; and maximizing rewards for the agent using the reward function during each traversal of the image using the agent. The instructions configure the processor to identify the N lane lines of the roadway as a single cluster of lane lines after traversing the image using the agent at least N times.

Resumen de: WO2022161599A1

Computer implemented methods for training a Student Neural Network, SNN, (600, 700, 800), and for managing an environment of a communication network using a trained SNN (900, 1000), are disclosed. The SNN is for generating an action prediction matrix for an environment in a communication network, the action prediction matrix comprising action predictions for a plurality of nodes or resources in the environment. The training method (600, 700, 800) comprises using a Reinforcement Learning process to train a Teacher Neural Network, TNN, to generate an action prediction for a resource or node in the environment (610), and using the trained TNN to generate a first training data set including action predictions for individual nodes or resources (620). The training method further comprises generating a second training data set from the first training data set (630) such that the second training data set includes action prediction matrices, and using the second training data set to update values of the parameters of the SNN (640).

Resumen de: US2022248237A1

A network metrics repository stores cell performance metrics and rule-based data measured during operation of a communication network. A policy neural network circuit has an input layer having input nodes, a sequence of hidden layers, and at least one output node. A processor trains the policy neural network circuit to approximate a baseline rule-based policy for controlling a tilt angle of a remote electrical tilt (RET) antenna based on the rule-based data. The processor provides a live cell performance metric to input nodes, adapts weights that are used by the input nodes responsive to output of the output node, and controls operation of the tilt angle of the RET antenna based on the output The output node provides the output responsive to processing a stream of cell performance metrics through the input nodes. The processor controls operation of the RET antenna based on the output.

Resumen de: US2022248179A1

The present invention discloses a method for identifying travel classification based on smartphone travel surveys. The method takes individual volunteers as the object, GPS information collected by smartphones and instant recall verification features of respondents as the data source. First, travel features, individual features, and family features are determined by GPS data and questionnaire information of volunteers. Then, a training data set is determined by the equal proportion method. Finally, an artificial neural network combined with a particle swarm optimization algorithm is used to detect travel classification from the survey data. The method described in the invention can realize automatic identification of six travel classifications. It is conducive to replace the traditional resident travel survey with the advanced survey method under the environment of big data. At the same time, it provides data foundation for urban management and planning.

Resumen de: US2022247800A1

Co-browsing allows a providing party to access visual content on a computing device for sharing with one or more other parties. The parties receiving the shared image may have dissimilar security authorizations. Accordingly, systems and methods are provided that enable shared content, such as a document, web page viewed in a browser, etc., to automatically be redacted to block those parties who are not authorized to view the content. For example, a neural network may be utilized to scan the document and provide specific redacted copies to the parties so each can view the image of the content with unauthorized content redacted.

Resumen de: US2022245782A1

A method for classifying an image of a displaying base plate includes: acquiring an image to be checked; from a first predetermined-type set, determining a type of the image to be checked. The first predetermined-type set includes: a first image type, a second image type and a third image type. An image of the first image type is a no-defect image, an image of the second image type is a blurred image, and an image of the third image type is a defect image. When the type of the image to be checked is the third image type, by using a first convolutional neural network, determining a defect data of the image to be checked, wherein the defect image refers to an image of a displaying base plate having a defect, and the defect data contains a defect type of the displaying base plate in the image to be checked.

Resumen de: US2022245809A1

Embodiments of the present disclosure include a method, device and computer readable medium involving receiving image data to detect tissue lesions, passing the image data through at least one first convoluted neural network, segmenting the image data, fusing the segmented image data, and detecting tissue lesions.

Resumen de: US2022245801A1

The technology disclosed relates to training a convolutional neural network (CNN) to identify and classify images of sections of an image generating chip resulting in process cycle failures. The technology disclosed includes creating a training data set of images of dimensions M×N using labeled images of sections of image generating chip of dimensions J×K. The technology disclosed can fill the M×N frames using horizontal and vertical reflections along edges of J×K labeled images positioned in M×N frames. A pretrained CNN is further trained using the training data set. Trained CNN can classify a section image as normal or depicting failure. The technology disclosed can train a root cause CNN to classify process cycle images of sections causing process cycle failure. The trained CNN can classify a section image by root cause of process failure among a plurality of failure categories.

Resumen de: US2022245768A1

A computer-implemented method of handling an emergency incident rcan include receiving information on an emergency incident that includes at least one image of the emergency incident, applying a Convolutional Neural Network (CNN) object recognition and classification process for identifying and marking objects on the at least one image that are related to the emergency incident and that may cause at least one secondary hazardous situation, processing the data relating to the identified and marked objects by applying a deep learning algorithm to the data in a Relational Network (RN) architecture, wherein the image on the basis of the identified and marked objects is correlated to a set of recognized objects in a database for classifying the emergency. A communication network, communication apparatus, and an emergency processing unit are also provided. Embodiments of such machines and systems can be configured to implement embodiments of the method.

Resumen de: US2022245404A1

Disclosed herein are examples of a wearable display system capable of determining a user interface (UI) event with respect to a virtual UI device (e.g., a button) and a pointer (e.g., a finger or a stylus) using a neural network. The wearable display system can render a representation of the UI device onto an image of the pointer captured when the virtual UI device is shown to the user and the user uses the pointer to interact with the virtual UI device. The representation of the UI device can include concentric shapes (or shapes with similar or the same centers of gravity) of high contrast. The neural network can be trained using training images with representations of virtual UI devices and pointers.

Resumen de: GB2603400A

A method for process control using predictive long short term memory includes obtaining historical post-process measurements taken on prior products of the manufacturing process; obtaining historical in-process measurements taken on prior workpieces during the manufacturing process; training a neural network to predict each of the historical post-process measurements, in response to the corresponding historical in-process measurements and preceding historical post-process measurements; obtaining present in-process measurements on a present workpiece during the manufacturing process; predicting a future post-process measurement for the present workpiece, by providing the present in-process measurements and the historical post-process measurements as inputs to the neural network; and adjusting at least one controllable variable of the manufacturing process in response to the prediction of the future post-process measurement.

Resumen de: WO2022158681A1

Disclosed are a door control device for controlling locking and unlocking of a door on the basis of face recognition, and a method for operating same. The present invention acquires a face image of a user through a camera mounted on a door, passes the acquired face image through a convolutional neural network to acquire an output vector, makes comparison between feature vectors of the faces of a plurality of users who are pre-specified as being permitted to enter and the output vector to identify a first feature vector having a maximum similarity to the output vector from among the feature vectors for the face images of the plurality of users, and then unlocks a locking device mounted on the door to allow the user to easily unlock the door only through face recognition when it is confirmed that an error between the first feature vector and the output vector is less than a preset reference value.

Resumen de: WO2022158665A1

An electronic device is disclosed. The electronic device according to the present disclosure comprises a processor, wherein the processor may: receive, from a first external device, text corresponding to a user's utterance and information about the first external device; obtain a plurality of weights for a plurality of elements related to the first external device by inputting the text corresponding to the user's utterance to a first neural network model; identify a second external device on the basis of the information about the first external device and the plurality of weights; transmit the user's utterance to the second external device; receive first response information about the user's utterance from the second external device; obtain second response information by inputting the first response information and the information about the first external device to a second neural network model; and transmit the second response information to the first external device.

Resumen de: WO2022156235A1

The present disclosure relates to a neural network training method and apparatus, an image processing method and apparatus, and an electronic device and a storage medium. The neural network training method comprises: inputting a sample image into a first segmentation network to obtain a first segmentation result of the sample image; inputting the sample image into a second segmentation network to obtain a second segmentation result of the sample image; and training the first segmentation network and the second segmentation network at least according to the first segmentation result and the second segmentation result. According to the neural network training method in the embodiments of the present disclosure, a first segmentation result obtained by a first segmentation network can be taken as predicted labeling information, such that when there are insufficient sample images with labeling information, the first segmentation network and a second segmentation network can be continuously trained, thereby improving the training effect and the precision of a neural network. When medical images with a relatively high labeling difficulty and a small number of samples are processed, the precision of the neural network can be improved, such that the neural network can accurately obtain, through segmentation, areas, in which targets are located, from the medical images.

Resumen de: WO2022156609A1

The embodiments of the present disclosure relate to a keyboard encryption method, and a device, a storage medium and a computer program product. The method comprises: performing feature extraction on a keyboard character image, so as to determine a feature of the keyboard character image; determining a fusion feature according to the feature of the keyboard character image and a feature of a wrong keyboard character; determining an encrypted keyboard character image according to the fusion feature, wherein the wrong keyboard character is different from the character corresponding to the keyboard character image; and constructing an encrypted keyboard on the basis of the encrypted keyboard character image, wherein a neural network recognition result of the encrypted keyboard character image is a wrong keyboard character, and the encrypted keyboard character image has the same visual recognition result as the keyboard character image. By using the method, the security of a keyboard can be improved, thereby effectively preventing the occurrence of malicious registration behavior.