Neural networks

NEURAL NETWORK TRAINING METHOD AND RELATED DEVICE

Absstract of: EP4790604A1

Embodiments of this application disclose a neural network training method. A first module in a federated neural network is deployed in each of a plurality of first devices. The first module includes a feature extraction module. A plurality of second modules in the federated neural network and early exit modules connected to the respective second modules are deployed in a second device. It can be learned that, in a federated learning process, the second device may include a plurality of early exit nodes, and each early exit node corresponds to one second module and a corresponding early exit module connected to the second module. In this way, after federated learning, when a target network is deployed in the first device, a structure of the target network may be in a plurality of forms based on the plurality of early exit nodes. In other words, target networks of different structures may be flexibly deployed in different first devices. For example, flexible scheduling may be performed based on resource statuses of different first devices, so that each first device can implement efficient data processing through a target network of an appropriate scale.

DEVICE FOR ACQUIRING MAGNETIC RESONANCE IMAGE ON BASIS OF DEEP LEARNING MODEL AND CONTROL METHOD THEREOF

Absstract of: EP4790443A1

The present disclosure provides an apparatus for restoring the quality of magnetic resonance images based on a deep learning model and a method of controlling the same. The method includes: obtaining a training image corresponding to each magnetic resonance image by applying at least one of a plurality of elements set in connection with the quality of the magnetic resonance image to a magnetic resonance signal corresponding to the magnetic resonance image; obtaining a training dataset including the magnetic resonance image as label data and the obtained training image as input data matching the label data; and training a neural network model based on the training dataset and context data corresponding to the training image. Obtaining the training image includes distorting the magnetic resonance signal by applying the at least one of the plurality of elements and obtaining the training image based on the distorted magnetic resonance signal.

NEURAL NETWORK CACHE EVICTION TO AVOID RESTORE

Absstract of: US20260228135A1

0000 Apparatuses, systems, and techniques to identify information to evict from a Key-Value (KV) cache. In at least one embodiment, information stored within one or more large language model (LLM) KV caches may be identified to cause an indication to generated of information stored within the one or more LLM KV caches that can be evicted without causing other information stored within the one or more LLM KV caches to be restored to the one or more LLM KV caches.

TRANSFORMER-BASED SHAPE MATCHING ACROSS IMAGES

Absstract of: WO2026165126A1

A system for matching shapes between images includes a transformer-based neural network with an encoder processing input images and a decoder processing a query shape from a first image. A parallel decoding module estimates corresponding shapes in a second image based on processed feature maps and query shapes from the first image. The system may include a recursive zoom-in module that iteratively refines the estimated corresponding shapes by zooming into regions around initial estimates to achieve sub-pixel accuracy. The transformer-based neural network can be trained using a data generation process that adapts based on training loss curves.

NEURAL PROCESSING CIRCUIT WITH CONCURRENT IMAGE PROCESSING AND NON-MAXIMUM SUPPRESSION

Absstract of: US20260228499A1

According to one example of the present disclosure, the neural processing unit may comprise a processing element array configured to perform operations of a neural network model and a post-processing unit configured to process data output from the processing element array. The post-processing unit includes a first computation circuit that extracts a subset of classes for each bounding box by comparing class scores of classes and a second computation circuit configured to extract one or more bounding boxes by comparing a class confidence score of each bounding box with a threshold confidence score.

MAPPING IMAGES TO WORDS FOR COMPOSED IMAGE RETRIEVAL

Absstract of: US20260228272A1

0000 Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for composed image retrieval. In one aspect, a method performed by one or more computers is described. The method includes: receiving a query including: (i) an image depicting a scene, and (ii) a text prompt describing a context of the scene; processing the image, using a visual encoder, to generate a visual embedding of the image; processing the visual embedding of the image, using a mapping neural network, to generate one or more language tokens of the image; generating multiple language tokens of the text prompt; processing the language tokens of the image and text prompt, using a language encoder, to generate a language embedding of the query; and selecting, from a number candidate images, one or more of the candidate images using the language embedding of the query.

OPTIMIZING PARAMETER ESTIMATION FOR TRAINING NEURAL NETWORKS

Absstract of: US20260228308A1

Apparatuses, systems, and techniques estimate parameters to train one or more neural networks based on uniqueuss of training data. In at least one embodiment, a subset of training data is selected and used to estimate parameters to train one or more neural networks, based on, for example, uniqueness of training data.

TRAINING END-TO-END WEAKLY SUPERVISED NETWORKS AT THE SPECIMEN (SUPRA-IMAGE) LEVEL

Absstract of: US20260228895A1

0000 Techniques for determining a presence of a pathology property in a supra-image are presented. The techniques can include receiving an electronic evaluation supra-image; providing the electronic evaluation supra-image to an electronic neural network that has been trained, using a training corpus of training supra-images and on an electronic computer, to determine the presence of the pathology property in a supra-image, each training supra-image including at least one image, each image corresponding to a plurality of components, wherein each training supra-image of the training corpus is associated with a respective electronic label indicating whether the pathology property is present, where the training corpus is sufficient to train the electronic neural network to determine a presence of the pathology property; receiving from the electronic neural network an output indicative of whether the pathology property is present in the evaluation supra-image; and providing the output.

PIPELINED MATRIX MULTIPLICATION AT A GRAPHICS PROCESSING UNIT

Absstract of: US20260228302A1

A graphics processing unit (GPU) schedules recurrent matrix multiplication operations at different subsets of CUs of the GPU. The GPU includes a scheduler that receives sets of recurrent matrix multiplication operations, such as multiplication operations associated with a recurrent neural network (RNN). The multiple operations associated with, for example, an RNN layer are fused into a single kernel, which is scheduled by the scheduler such that one work group is assigned per compute unit, thus assigning different ones of the recurrent matrix multiplication operations to different subsets of the CUs of the GPU. In addition, via software synchronization of the different workgroups, the GPU pipelines the assigned matrix multiplication operations so that each subset of CUs provides corresponding multiplication results to a different subset, and so that each subset of CUs executes at least a portion of the multiplication operations concurrently.

GENERALIZED ENERGY DISTANCE DIFFUSION MODELS

Absstract of: WO2026165379A1

A method for generating a data item using a generative neural network model comprises, at each of a plurality of iterations, the generative neural network model processing a network input based on a current version of a data representation, to form a network output. The generative neural network is trained to form the network output as a sample from a target probability distribution over the representation space. The data representation is a vector defined in a representation space, and in each iteration, the network output is used to generate an updated data representation. The data item is generated based on the updated data representation generated in the last iteration.

COMPREHENSIVE RESILIENCE EVALUATION SYSTEM FOR TRADITIONAL VILLAGE

Absstract of: US20260228305A1

A comprehensive resilience evaluation system for a traditional village is provided. The system employs a deep learning-based neural network model to perform semantic encoding on data for resilience aspects of a traditional village to be evaluated to extract semantic descriptive coding features from the multi-dimensional data for the resilience aspects of the village to be evaluated. Concurrently, the system retrieves village resilience-related data labeled with a first resilience evaluation label from a backend database as reference features. Through conducting a semantic query matching analysis based on multi-source resilience-related data between the village to be evaluated and various villages labeled with the first resilience evaluation label, the system intelligently evaluates whether the village to be evaluated possesses a resilience level corresponding to the first resilience evaluation label. Thus, the accuracy of the comprehensive resilience evaluation for the village to be evaluated can be effectively improved.

DYNAMIC CLASS-INCREMENTAL LEARNING WITHOUT FORGETTING

Absstract of: US20260228525A1

A processor-implemented method for dynamic class-incremental learning without forgetting includes receiving, by an artificial neural network (ANN), an input. The ANN extracts features of the input to generate a representation of the input. An embedding is generated by the ANN based on the representation and multiple similarity metrics. The ANN generates a new class without retraining the ANN. The new class is generated based on a comparison of the embedding and a set of prior embeddings.

OBJECT DETECTION FROM SYNTHETIC APERTURE RADAR USING A COMPLEX-VALUED CONVOLUTIONAL NEURAL NETWORK

Absstract of: US2025180734A1

Systems and methods are provided for object detection. A radar interface receives complex-valued data representing a region of interest from a synthetic aperture radar system. A complex-valued convolutional neural network includes a plurality of convolutional layers and provides an output indicating if objects are present in the region of interest. Each convolutional layer includes a complex-valued kernel that is applied to an input. The kernel includes a first set of weights that is applied to each of real and imaginary components of the input to provide respective first and second convolution products and a second set of weights applied to each of real and imaginary components of the input to provide respective third and fourth convolution products. A difference between the first and fourth convolution products provides a real output component and a sum of the second and third convolution products provides an imaginary output component.

CATEGORIZATION WITH GRAPH NEURAL NETWORK AND LANGUAGE MODEL

Absstract of: EP4769175A1

0001 Certain aspects of the disclosure provide techniques for categorization by a device. An example method includes receiving input information regarding a plurality of classification targets and a plurality of categories for classification of the plurality of classification targets; generating a plurality of embeddings for the input information using a first model, the plurality of embeddings including: a set of first embeddings associated with the plurality of classification targets, and a set of second embeddings associated with the plurality of categories; determining that a similarity score for the set of first embeddings and the set of second embeddings fails to satisfy a threshold; generating, based on the similarity score failing to satisfy the threshold and using a graph neural network (GNN), a classification of the plurality of classification targets in accordance with the plurality of categories; and outputting information regarding the classification.

METHOD FOR SOLVING NONLINEAR EQUATION SET, AND RELATED DEVICE

Absstract of: EP4787198A1

Embodiments of this application disclose a method for solving a system of nonlinear equations. In the method, during an i^th iteration of a first system of nonlinear equations, a neural network model can identify, based on input description information of the i^th iteration, system of equations feature information of the first system of nonlinear equations during the i^th iteration, to efficiently determine a first convergence policy that is appropriate for the first system of nonlinear equations in the current iteration process. In this way, a target convergence policy is determined according to the first convergence policy, and the first system of nonlinear equations is iteratively solved efficiently and accurately according to the target convergence policy, so that performance of iterative solving on a system of nonlinear equations is improved (for example, a solving speed is improved and solving time consumption is reduced), and a generalization problem caused by a convergence policy determined based on expert experience and according to a default convergence policy is avoided, thereby resolving a problem that a convergence policy determined in different scenarios has a poor effect and is difficult to converge when a system of nonlinear equations is solved.

WEAKLY SUPERVISED OBJECT DETECTION METHOD BASED ON MISCLASSIFICATION CORRECTION, AND RELATED DEVICE

Absstract of: WO2026157121A1

Disclosed in the present invention are a weakly supervised object detection method and system based on misclassification correction, an electronic device, and a storage medium. The method comprises: extracting region features of target candidate locations by using a convolutional neural network; assigning category labels to the extracted region features by means of a classifier; and designing a misclassification correction-driven label assignment module on the basis of confidence differences between categories so as to identify misclassification cases and correct same, and reassigning labels for training of an object detection model. In the method of the present invention, erroneous category labels generated in a training phase are corrected, significantly improving the classification performance of the weakly supervised object detection method.

METHOD, SYSTEM, AND APPARATUS FOR ANALYZING EVENT ON THE BASIS OF GENERATIVE TASK AND MULTIMODALITY

Absstract of: WO2026157818A1

The present application discloses a method, system, and apparatus for analyzing an event on the basis of a generative task and multimodality, which are used for quickly and accurately ascertaining the progress of an event. The method of the present application comprises: generating structured event information from multi-modal data, and obtaining an event relationship; constructing an event graph; extracting multi-modal features, and using a cross-modal attention mechanism to obtain multi-modal sentiment relationship vectors; on the basis of a preset graph neural network and in view of the event graph, obtaining sentiment feature vectors; calculating a contrastive learning loss value, and obtaining a target sentiment feature on the basis of the contrastive learning loss value; using the cross-modal attention mechanism to generate a target event graph; extracting a relationship feature, and obtaining an event prediction graph in view of attention weights and a preset time series model; extracting a causal chain and performing inference, and obtaining a trigger condition and an emotion factor after the inference; and combining the event prediction graph and the target event graph with a node feature sequence, the trigger condition, and the emotion factor, and obtaining an event analysis result.

ACCELERATING MULTI-VIEW SEARCH WITH DYNAMIC VIEW ROUTING

Absstract of: US20260220430A1

Methods, systems, and apparatuses, including computer programs encoded on computer storage media, for selecting relevant documents. A system processes a query using a query encoder neural network to generate query embeddings. A router neural network processes the query embeddings to generate a respective score for each embedding. The system selects a representative query embedding based on the scores and utilizes the representative query embedding and stored document embeddings to generate a relevance score for each of a plurality of documents. The system then selects a subset of the documents as relevant to the query based on the relevance scores.

HIERARCHICAL MEMORY ATTENTION NETWORK WITH MULTI-COMPONENT WEIGHTED ATTENTION AND ADAPTIVE RESOURCE ALLOCATION

Absstract of: US20260220425A1

Systems, methods, and computer-readable media are disclosed for providing hierarchical processing in neural network architectures (e.g., transformer-based models, recurrent networks, state space models, memory-augmented networks, and retrieval-augmented systems), which may improve computational efficiency and memory utilization.

Proxy-Based Multi-Modal Learning For Render Time Prediction

Absstract of: US20260220485A1

0000 Techniques for predicting render times are disclosed. A system accesses auxiliary rendered outputs that were generated from a coarse render pass of a 3D scene description using a set of hardware resources. The system accesses system performance metrics of the set of hardware resources associated with the coarse render pass. The system encodes visual feature representations from the auxiliary rendered outputs using a convolutional neural network. The system also encodes the system performance metrics and a set of rendering configuration parameters associated with a target rendering operation. The system generates, using a multi-modal fusion model, a fused representation of the encoded visual feature representations, the encoded system performance metrics, and the encoded first set of rendering configuration parameters, using an attention-based mechanism that weights contributions of the inputs to the model. The system predicts a render time for the target rendering operation based on the fused representation.

FINGERPRINT CARRIER STATE RECOGNITION METHOD AND APPARATUS, ELECTRONIC DEVICE, AND STORAGE MEDIUM

Absstract of: WO2026156746A1

Embodiments of the present application provide a fingerprint carrier state recognition method and apparatus, an electronic device, and a storage medium. The fingerprint carrier state recognition method comprises: acquiring a current fingerprint image frame; inputting feature information of the current fingerprint image frame, time information corresponding to the current fingerprint image frame, and a latent variable corresponding to the current fingerprint image frame into a recurrent neural network model to obtain carrier state information outputted by the recurrent neural network model and a latent variable corresponding to a next fingerprint image frame, wherein a fingerprint carrier is a medium carrying a fingerprint corresponding to a fingerprint image; and determining the state of the fingerprint carrier on the basis of the carrier state information. In the fingerprint carrier state recognition method provided by the present application, carrier state recognition is performed with reference to the time information of the current fingerprint image frame and a latent variable obtained from a historical fingerprint image, so that a high recognition accuracy rate can be achieved.

ELECTRONIC APPARATUS FOR PREDICTING SYSTEMIC BIOLOGICAL AGE AND OPERATION METHOD THEREOF

Absstract of: WO2026160517A1

An electronic apparatus for predicting systemic biological age, according to an embodiment of the present invention, comprises a memory and a processor that predicts systemic biological age, wherein the processor can acquire an eye image of an examinee, process the eye image by using an artificial neural network model that has been established, and generate a predicted result for the systemic biological age of the examinee.

SYSTEMS, METHODS AND TESTING APPARATUS FOR ASSESSING PLATELET SAMPLES USING MACHINE LEARNING FOR CANCER ESTIMATION

Absstract of: WO2026156436A1

An artificial intelligence-based method for detecting a disease comprises a meta- classification model comprising a plurality of base classification models each generating a respective output and wherein each of such outputs is input to a meta-model, the meta-model comprising an artificial neural network. Variants are described in respect of the meta-model stacked ensemble pipeline, dimensionality / complexity control, and specific architectures designed for tracking disease trajectory through identifying gene pathway crosstalk. The method is particularly suitable for detecting cancer in blood samples comprising tumor educated platelets.

AUGMENTING DEEP NEURAL NETWORKS WITH RESIDUAL MEMORIZATION

Absstract of: US20260221132A1

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for performing a machine learning task. In particular, the machine learning task is performed by augmenting a trained neural network with residual memorization.

SOMATIC VARIANT PREDICTION

Nº publicación: US20260221225A1 30/07/2026

Applicant:

AGENCY FOR SCIENCE TECH AND RESEARCH [SG]
Agency for Science, Technology and Research

Absstract of: US20260221225A1

0000 Method and systems for prediction of somatic mutations based on data of a potentially tumorous sample, by receiving nucleic acid sequencing data of the potentially tumorous sample; transforming the nucleic acid sequencing data into an image-like representation; processing the image-like representation by a trained neural network to predict somatic mutations in the nucleic acid sequencing data.