Alerta

METHODS AND SYSTEMS FOR VERIFYING MODELS AGAINST CONVOLUTIONAL PERTURBATIONS VIA PARAMETERIZED KERNELS

Absstract of: US20260268135A1

The present disclosure is directed to methods and systems that provide for efficient verification of neural networks against convolutional perturbations such as blurring and sharpening. Perturbation kernels are constructed based on user input, and parameterized versions of the perturbation kernels are derived as one or more matrices. Results of convolved input data for the user model are precomputed with the matrix (ices), and a robustness verification analysis of the user model is performed with prepended threshold layers. A determination is made regarding robustness according to a perturbation specification, and based upon that determination, an indication of robustness is provided to the user and/or counterexamples can be extracted and returned to modify the user model such that the model can be modified based on the counterexamples. The data analyzed by the methods and systems can include images, and thus kernels can include camera shake, box blur, and sharpen kernels.

Distributed System for Identifying the Appropriate Heart Valve Implant Using an Ensemble Machine-Learning Algorithm, Synthetic Data, and Edge Computing

Absstract of: US20260269081A1

0000 A system and method for and method for selecting an implantable heart valve for patients using is disclosed. The system includes analyzing population-level environmental exposure data to categorize geographic locations by cardiovascular risk. A synthetic clinical dataset is generated using a generative adversarial network along with predefined machine-learning parameters. The user device calculates an exposomic feature value by weighting pollutant concentrations based on residence durations. The user device is then trained using a machine-learning model, selected from random forest models, gradient-boosted decision tree models, support vector machines, artificial neural networks, and elastic net regression. After training, the model is executed on the user device using an edge computing approach with patient-specific data and exposomic feature values to generate heart valve intervention options that minimize the difference between the predicted remaining lifespan of the patient and the valve's operational lifespan while reducing the risk of failure or complications.

IMAGE ENCODING/DECODING METHOD USING NEURAL NETWORK-BASED IN-LOOP FILTER, DEVICE, AND RECORDING MEDIUM STORING BITSTREAM

Absstract of: US20260270423A1

An image decoding method using a neural network-based in-loop filter may comprise obtaining a first image feature from an input image, obtaining a block information feature of the input image from block information of the input image, obtaining a second image feature by removing noise and distortion of the first image feature based on the block information feature, and reconstructing the input image based on the second image feature. The block information may comprise at least one of a block boundary map indicating a block partition structure of the input image or a block distribution map indicating encoding information of the input image.

GENERATING CANDIDATE PROTEINS USING GENERATIVE NEURAL NETWORKS AND STRATEGIES FOR SEARCHING PROTEIN SPACES

Absstract of: WO2026187661A1

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for performing an automated search through a space of proteins to generate data defining a library of candidate proteins for performing a functional task. One of the methods includes performing an unconstrained search of the space of proteins to identify a first collection of candidate proteins. The method additionally includes performing a constrained search of the space of proteins to identify a second collection of candidate proteins, in which the constrained search uses a search space that is limited by one or more constraints and comprises only a proper subset of the space of proteins. The method includes generating the library of candidate proteins by combining the first collection generated by the unconstrained search and the second collection generated by the constrained search, and outputting the library of candidate proteins for performing the functional task.

DATA PROCESSING METHOD AND APPARATUS

Absstract of: WO2026184444A1

A data processing method, comprising: performing word segmentation on an input sequence, and mapping a word segmentation result into a first vocabulary index; and processing the first vocabulary index by means of a first neural network model to obtain an inference result, wherein in a first mixture-of-experts layer of the first neural network model, on the basis of the first vocabulary index, a first lookup table stored in a memory is queried to obtain outputs of expert networks in the first mixture-of-experts layer, the first lookup table comprises the mapping relationship between indexes in a vocabulary and results obtained by performing inference on embedding representations of the indexes by the expert networks in the first mixture-of-experts layer, and the vocabulary is a word set used by the first neural network model. In the method, expert networks in mixture-of-experts layers do not need to be loaded during inference, thereby reducing video random-access memory usage and communication traffic, eliminating expert network computation in the mixture-of-experts layers, and improving inference efficiency.

NEURAL NETWORK FUNCTIONS FOR POSITIONING OF A USER EQUIPMENT

Absstract of: US20260268112A1

0000 In an aspect, a BS obtains at least one neural network function configured to facilitate a UE to derive a likelihood of at least one set of positioning measurement features being present at a candidate set of positioning estimates for the UE, the at least one neural network function being generated dynamically based on machine-learning associated with one or more historical measurement procedures. The BS transmits the at least one neural network function to the UE. In another aspect, the UE obtains positioning measurement data associated with a location of the UE (e.g., locally the UE, or remotely from the BS). The UE determines a positioning estimate for the UE based at least in part upon the positioning measurement data and the at least one neural network function.

SIMULATION-BASED TRAINING METHOD FOR MODEL, ELECTRONIC DEVICE, STORAGE MEDIUM, AND PROGRAM PRODUCT

Absstract of: WO2026184150A1

The embodiments of the present disclosure relate to the technical field of artificial intelligence, and provide a simulation-based training method for a model, an electronic device, a storage medium, and a program product. The method comprises: transmitting, to a target graphics processing unit, training data determined according to a training instruction, the training data comprising data of at least one neural network layer of a model indicated by the training instruction; for any neural network layer among the at least one neural network layer, acquiring a training duration for the target graphics processing unit to train said neural network layer; and generating a simulation-based training result of the model according to the training duration of the at least one neural network layer. The technical solution of the embodiments of the present disclosure reduces simulation-based training consumption of real computing resources.

WIRELESS SIGNAL ADJUSTMENT

Absstract of: US20260271132A1

0000 Apparatuses, systems, and techniques to adjust one or more discontinuous wireless communication patterns. In at least one embodiment, a processor includes one or more circuits to use one or more neural networks to adjust one or more discontinuous wireless communication patterns.

TECHNIQUES FOR MODIFYING AND TRAINING A NEURAL NETWORK

Absstract of: US20260268651A1

Apparatuses, systems, and techniques are described herein to speed up inferencing in a neural network by copying output from one layer of the neural network to another computing resource based on dependencies among layers in the network. In at least one embodiment, a processor comprising one or more circuits causes two or more subsequent layers of one or more neural networks to be performed on separate computing resources from a previous layer of the one or more neural networks.

AUTOMATED CLASSIFIER FOR DIAGNOSING ACUTE OTITIS MEDIA

Absstract of: AU2025228694A1

A system for assessing AOM includes a computing device having a processor apparatus, wherein the processor apparatus implements a diagnostic classifier component that comprises a. trained neural network, the processor apparatus being structured and configured to receive tympanic membrane image data representing one or more images of a tympanic membrane of the patient, provide the tympanic membrane image data to the diagnostic classifier component, and process the tympanic membrane image data in the diagnostic classifier component to determine: (i) a plurality of tympanic membrane features from the tympanic membrane image data, (ii) a diagnosis of whether the patient has AOM based on the plurality of tympanic membrane features, (iii) a. confidence in the diagnosis, and (iv) an identification one or more of the tympanic membrane features determined to be predominant contributing features that led to the diagnosis.

NEURAL MODELER OF AUDIO SYSTEMS

Absstract of: US20260268924A1

0000 A process trains a neural network that digitally models an audio system. The process couples a test signal into an input of a reference audio system. The process further electronically collects an output of the reference audio system responsive to the test signal as captured information. Moreover, the process trains a neural network using at least some of the captured information such that the overall output of the neural network converges towards an output representative of the reference audio system. The process also outputs to a graphical user interface, a graphical representation associated with the trained neural network, the graphical representation visually displaying at least one virtual control. Here, upon coupling a musical instrument to the trained neural network, a digital signal representing a musical instrument signal from the musical instrument is processed through the trained neural network in the time domain with an algorithmic latency under 20 milliseconds.

NEURAL NETWORK FOR BULK SORTING

Absstract of: AU2026220313A1

A bulk sorting system for sorting objects (1) in bulk is provided. The bulk sorting system comprises: at least one radiation source (10) arranged to radiate the objects, at least one optical sensor (12) arranged to capture reflected radiation (22) of 5 the objects and acquire the reflected radiation as multi- or hyperspectral data (24); a processing circuit (16) configured to analyze the reflected radiation of the objects by inputting the multi- or hyperspectral data into a convolutional neural network (CNN) (18) with at least two convolutional layers in order to either detect and classify the objects in the multi- or hyperspectral data and/or semantically segment the multi- or 10 hyperspectral data; and a mechanical sorter (20) configured to sort the objects according to their classification and/or segmentation using the analysis of the processing circuit such that different overlapping and/or stacked objects are separated or treated as a single group of objects. To be published with Fig. 1. ug u g Fig. 1 ug u g

SELF-LEARNING IN DISTRIBUTED ARCHITECTURE FOR ENHANCING ARTIFICIAL NEURAL NETWORK

Absstract of: US20260268648A1

A vehicle having the first ANN model initially installed therein to generate outputs from inputs generated by one or more sensors of the vehicle. The vehicle selects an input based on an output generated from the input using the first ANN model. The vehicle has a module to incrementally train the first ANN model through unsupervised machine learning from sensor data that includes the input selected by the vehicle. Optionally, the sensor data used for the unsupervised learning may further include inputs selected by other vehicles in a population. Sensor inputs selected by vehicles are transmitted to a centralized computer server, which trains the first ANN model through supervised machine learning from sensor received inputs from the vehicles in the population and generates a second ANN model as replacement of the first ANN model previously incrementally improved via unsupervised machine learning in the population.

DEFECT DETECTION METHOD AND APPARATUS FOR CERAMIC MODULE, AND DEVICE AND STORAGE MEDIUM

Absstract of: WO2026184676A1

A defect detection method and apparatus for a ceramic module, and a device and a storage medium. The method comprises: acquiring an original module image corresponding to a target ceramic module to be subjected to defect detection (S101); on the basis of the original module image and a pre-trained target colloid annotation model, determining a colloid region image corresponding to the target ceramic module (S102), wherein the target colloid annotation model is obtained by means of training in advance on the basis of a U-shaped neural network model, and uses a convolutional block attention module; and on the basis of a module template image and the colloid region image of the target ceramic module, determining whether a colloid penetration defect occurs in the target ceramic module (S103).

FUNCTIONAL ACTIVATION-BASED ANALYSIS OF DEEP NEURAL NETWORKS

Absstract of: US20250148285A1

0000 Functional activation-based analysis of deep neural networks uses a structured set of inputs (e.g., input datasets corresponding to different knowledge or datatype domains) are sequentially provided to a pretrained neural network (e.g., according to a block-sequence). The output values for each node in the neural network are recorded and stored as a time-series of layer output values. A statistical analysis of the time-series of layer output values may be fit as a function of the structured set of inputs to generate neural network analysis data that indicate activations of layers within the neural network based on the inputs.

RISK ASSESSMENT OF ROTOR ANGLE INSTABILITY IN A POWER NETWORK

Absstract of: EP4804364A1

Rotor angle instability is a key criterion of dynamic stability in power networks. State-of-the-art machine-learning approaches are difficult to scale and have limited inputs with which to make predictions as to rotor angle instability. Accordingly, disclosed embodiments utilize a machine-learning model that is applied to bus voltage angles, which are local quantities available at every bus in the power network, to derive a prediction of the risk of rotor angle stability in the power network. These predictions may be biased in order to avoid false negatives. The machine-learning model may be a message-passing neural network. The resulting predictor is capable of quickly and reliably flagging potential rotor instability within a power network.

GENERATING AUDIO USING AUTO-REGRESSIVE GENERATIVE NEURAL NETWORKS

Absstract of: EP4804179A2

0001 Methods, systems, and apparatus, including computer programs encoded on computer storage media, for generating a prediction of an audio signal. One of the methods includes receiving a request to generate an audio signal; obtaining a semantic representation of the audio signal; generating, using one or more generative neural networks and conditioned on at least the semantic representation, an acoustic representation of the audio signal; and processing at least the acoustic representation using a decoder neural network to generate the prediction of the audio signal.

METHOD FOR WAKING UP DEVICE, AND SYSTEM AND CHIP

Nº publicación: EP4804020A1 09/09/2026

Applicant:

HUAWEI TECH CO LTD [CN]
Huawei Technologies Co., Ltd.

Absstract of: EP4804020A1

This application provides a device wake-up method, a system, and a chip. The method may include: processing a captured first image, determining a motion region image in the first image, and determining, by using a level-1 neural network, whether a target object exists in the motion region image; and when it is determined that the target object exists in the motion region image, performing reconfirmation, by using a level-2 neural network cascaded with the level-1 neural network, on a result output by the level-1 neural network. It can be learned that the target object is determined by using the level-1 neural network and the level-2 neural network, so that a false alarm frequency can be reduced, and system power consumption can be reduced.