Neural networks

SYSTEM AND METHOD FOR NEURAL NETWORK ORCHESTRATION

Absstract of: US2025384661A1

Methods and systems for training one or more neural networks for transcription and for transcribing a media file using the trained one or more neural networks are provided. One of the methods includes: segmenting the media file into a plurality of segments; inputting each segment, one segment at a time, of the plurality of segments into a first neural network trained to perform speech recognition; extracting outputs, one segment at a time, from one or more layers of the first neural network; and training a second neural network to generate a predicted-WER (word error rate) of a plurality of transcription engines for each segment based at least on outputs from the one or more layers of the first neural network.

METHOD AND DEVICE FOR TRAINING A NEURAL NETWORK BASED DECODER

Absstract of: US2025384277A1

A device and a method for training a neural network based decoder. The method includes during the training, quantizing, using a training quantizer, parameters representative of the coefficients of the neural network based decoder. A method and device are also provided for encoding at least parameters representative of the coefficients of a neural network based decoder. Provided also are a method for generating an encoded bitstream including an encoded neural network based decoder, a neural network based encoder and decoder, and a signal encoded using the neural network based encoder.

TRAINING ENCODER MODEL AND/OR USING TRAINED ENCODER MODEL TO DETERMINE RESPONSIVE ACTION(S) FOR NATURAL LANGUAGE INPUT

Absstract of: US2025384350A1

Systems, methods, and computer readable media related to: training an encoder model that can be utilized to determine semantic similarity of a natural language textual string to each of one or more additional natural language textual strings (directly and/or indirectly); and/or using a trained encoder model to determine one or more responsive actions to perform in response to a natural language query. The encoder model is a machine learning model, such as a neural network model. In some implementations of training the encoder model, the encoder model is trained as part of a larger network architecture trained based on one or more tasks that are distinct from a “semantic textual similarity” task for which the encoder model can be used.

AUTOMATED CREATION OF DIGITAL TWINS USING GRAPH-BASED INDUSTRIAL DATA

Absstract of: US2025384266A1

A computer-implemented method for automatically creating a digital twin of an industrial system having one or more devices includes accessing a triple store that includes an aggregated ontology of graph-based industrial data synchronized with the one or more devices. The triple store is queried for a specified device to extract, from the graph-based industrial data, structural information of the specified device defined by a tree comprising a hierarchy of nodes. For each node, a neural network element is assigned based on a mapping of node types to pre-defined neural network elements. The assigned neural network elements are combined based on the tree topology to create a digital twin neural network. The triple store is then queried to extract, form the graph-based industrial data, real-time process data gathered from the specified device at runtime and use the extracted real-time process data to tune parameters of the digital twin neural network.

MACHINE LEARNING SPARSE COMPUTATION MECHANISM FOR ARBITRARY NEURAL NETWORKS, ARITHMETIC COMPUTE MICROARCHITECTURE, AND SPARSITY FOR TRAINING MECHANISM

Absstract of: US2025384257A1

An apparatus to facilitate processing of a sparse matrix for arbitrary graph data is disclosed. The apparatus includes a graphics processing unit having a data management unit (DMU) that includes a scheduler for scheduling matrix operations, an active logic for tracking active input operands, and a skip logic for tracking unimportant input operands to be skipped by the scheduler. Processing circuitry is coupled to the DMU. The processing circuitry comprises a plurality of processing elements including logic to read operands and a multiplication unit to multiply two or more operands for the arbitrary graph data and customizable circuitry to provide custom functions.

COMPUTER-IMPLEMENTED METHOD AND ARRANGEMENT FOR OPERATING A RECOMMENDATION SYSTEM, COMPUTER-READABLE DATA CARRIER AND COMPUTER PROGRAM PRODUCT

Absstract of: WO2025257015A1

The invention relates to a computer-implemented method for operating a recommendation system which involves a) converting at least one part of at least one 3D CAD model capable of acquiring so-called "product manufacturing information", PMI, data into a graph representation in such a way that the graph representation comprises existing PMI data of the 3D CAD model, b) training a so-called "graph neural network", GNN, model at least at a first, in particular initial, point in time, the GNN model being trained on the basis of at least the graph representation, c) generating at least one recommendation output concerning at least the part of the 3D CAD model, in particular an output by the recommendation system on the basis of the graph representation and the GNN model. Furthermore, the invention relates to an arrangement for carrying out the method, to a computer-readable data carrier, and to a computer program product.

OBJECTIVE-CONDITIONED GENERATIVE NEURAL NETWORKS

Absstract of: WO2025260090A1

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training a generative neural network. In particular, the generative neural network is trained on an objective function that includes multiple different objectives, with two or more of the objectives being reward objectives.

ANOMALY DETECTING EDGE DEVICE WITH QUANTIZED DEEP NEURAL NETWORK ON FPGA, ASIC, OR SOC

Absstract of: WO2025255347A1

Systems and methods are described for an edge device with a chip-implemented, reduced-bit-quantized neural network (NN)-based anomaly detector, including a preprocessing logic block for preprocessing the sensor data, and a quantized, processing-in-memory (PIM) based NN block. The quantized PIM-based NN block includes quantized multiply-accumulate units (MACs) with quantized PIM logic that uses quantized-arithmetic-result lookup tables. The quantized MACs are arranged to function as quantized, PIM-based NN block kernels. Optionally, the quantized PIM- based NN block is implemented as an autoencoder with a quantized NN encoding and a quantized NN decoding section. The quantized NN encoder section performs NN-based, quantized dimensionality reduction of the processed sensor data, outputting a quantized lower-dimensional latent representation. The quantized NN decoder section reconstructs the quantized lower-dimensional latent representation, producing a quantized reconstructed input pattern. The quantized NN encoder and decoder sections include quantized MACs with PIM logic using quantized- arithmetic-result lookup tables.

INFORMATION PROCESSING METHOD AND APPARATUS, DEVICE, AND STORAGE MEDIUM

Absstract of: WO2025251721A1

According to embodiments of the present disclosure, provided are an information processing method and apparatus, a device, and a storage medium. The method comprises: determining a target processing layer from among a plurality of processing layers of a graph neural network, wherein a processing procedure corresponding to the target processing layer comprises a plurality of sub-processing procedures independent of each other; splitting the target processing layer into a plurality of sub-layers corresponding to the plurality of sub-processing procedures; using a plurality of processing devices to respectively process the plurality of sub-processing procedures corresponding to the plurality of sub-layers; and using a plurality of output results of processing of the plurality of processing devices in a summary layer corresponding to the plurality of sub-layers to determine a target output result for the target processing layer.

SPARSITY CONTROL BASED ON HARDWARE FOR DEEP-NEURAL NETWORKS

Absstract of: US2025378334A1

Systems, methods, computer program products, and apparatuses to transform a weight space of an inference model to increase the compute efficiency of a target inference platform. A density of a weight space can be determined, and a transformation parameter derived based on the determined density. The weight space can be re-ordered based on the transformation parameter to balance the compute load between the processing elements (PEs) of the target platform, and as such, reduce the idle time and/or stalls of the PEs.

PERSONALIZED GENERATIVE NEURAL NETWORKS

Absstract of: WO2025254658A1

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for generating personalized output data items using a generative neural network. In one aspect, one of the method includes maintaining a respective set of adaptation parameters of a generative neural network for each of a plurality of predetermined content providers; obtaining a new input associated with a new content provider; using a classifier model to generate a respective score for each of the plurality of predetermined content providers; generating a new set of adaptation parameters based on determining a weighted combination of the respective set of adaptation parameters for each of a subset of the plurality of predetermined content providers; and generating, by the generative neural network and based on the new set of adaptation parameters and the new input, a new personalized output data item.

RPA RPA workflow managing system and method based on an artificial neural network

Absstract of: KR20250172052A

RPA(Robtic Process Automation) 워크플로우 관리 시스템은, 적어도 하나 이상의 클라이언트 컴퓨팅 장치와 RPA 관리 컴퓨팅 장치를 포함하고, 클라이언트 컴퓨팅 장치의 프로세서는, 제1 RPA 워크플로우의 개시 조건이 충족되는 것에 응답하여 제1 RPA 로봇이 상기 제1 RPA 워크플로우를 개시하는 단계; 제1 RPA 워크플로우와 관련되는 제1 시점까지의 워크플로우 영상을 획득하고, 획득된 워크플로우 영상을 인코딩하여, 워크플로우 영상에 대한 제1 잠재 벡터를 생성하고, 제1 RPA 로봇과 통신가능한 RPA 관리 컴퓨팅 장치로 상기 잠재 벡터를 전송하는 단계를 수행하고, RPA 관리 컴퓨팅 장치의 마스터 프로세서는 제1 잠재 벡터와 미리 마련된 기준 잠재 벡터 간의 편차에 따라 적응적으로 상기 제1 잠재 벡터를 RPA 작업 예측 모델에 적용하여, 상기 제1 시점보다 후행하는 제2 시점에서의 예측 작업 영상을 생성하는 단계를 수행한다.

PREDICTION OF EQUIPMENT-RELATED EVENTS USING MACHINE LEARNING

Absstract of: WO2025250329A1

Techniques are disclosed herein for machine-learning (ML)-assisted event prediction for industrial machines (106). A first set of embeddings (1986a) can be generated based on labeled first event data, which can be labeled with classifiers (1989b) determined based on signaling channel (110b) information for the first event data. A neural network (300) can be trained, using the classifiers (1989b), to generate (i) a similarity score for the first set of embeddings and the second set of embeddings and (ii) a classifier (1989b) recommendation for the second set of embeddings. The second set of embeddings can be generated based on data collected using condition monitoring sensors (104a) for a particular industrial machine. Accordingly, the system (200) can generate alerts, recommendations, and/or notifications based on the automatically classified data encoded in the second set of embeddings.

SPEAKER IDENTIFICATION, VERIFICATION, AND DIARIZATION USING NEURAL NETWORKS FOR CONVERSATIONAL AI SYSTEMS AND APPLICATIONS

Absstract of: US2025372084A1

Disclosed are apparatuses, systems, and techniques that may use machine learning for implementing speaker recognition, verification, and/or diarization. The techniques include applying a neural network (NN) to a speech data to obtain a speaker embedding representative of an association between the speech data and a speaker that produced the speech. The speech data includes a plurality of frames and a plurality of channels representative of spectral content of the speech data. The NN has one or more blocks of neurons that include a first branch performing convolutions of the speech data across the plurality of channels and across the plurality of frames and a second branch performing convolutions of the speech data across the plurality of channels. Obtained speaker embeddings may be used for various tasks of speaker identification, verification, and/or diarization.

SYSTEMS AND METHODS FOR DEFINING CONFIDENCE IN DEEP LEARNING MODEL PREDICTION

Absstract of: US2025371340A1

The invention relates to a technique for improving confidence estimates associated with neural networks. The technique involves computing neuron activation statistics during training, evaluating neuron activations during inferencing and determining how the activations compare with the previously computed statistics (e.g. whether prediction activations are within the bounds of the training activation statistics). The comparison may be used to compute a confidence value for the neural network.

METHODS AND SYSTEMS FOR PROBABILISTIC RELATIONAL LEARNING

Absstract of: WO2025247606A1

Methods and systems for fully quantifying predictive uncertainty for recommender systems are disclosed. Embodiments include providing a fully Bayesian multiway neural network model, assigning a prior distribution over a plurality of weight parameters of a model parameter set of the neural network model such that a corresponding posterior distribution represents an epistemic uncertainty of the neural network model; incorporating a mixture density model for a target layer of the neural network model, wherein the mixture density model models aleatoric uncertainty of input data; estimating the posterior distribution of the weight parameters given training data by updating the prior distribution based on data likelihood of the training data; sampling a plurality of empirical samples from the estimated posterior distribution of the weight parameters; performing inference for each stochastic realization of the neural network model with the sampled weight parameters on an inference sample to generate one or more predictive distributions; drawing samples from each of the one or more predictive distributions as final predictions for the inference sample to obtain a matrix of predictions for the inference sample, wherein the matrix represents a joint distribution of both the aleatoric and epistemic uncertainty of the recommender system.

PREDICTION OF EQUIPMENT-RELATED EVENTS USING MACHINE LEARNING

Absstract of: US2025370446A1

Techniques are disclosed herein for machine-learning (ML)-assisted event prediction for industrial machines. A first set of embeddings can be generated based on labeled first event data, which can be labeled with classifiers determined based on signaling channel information for the first event data. A neural network can be trained, using the classifiers, to generate (i) a similarity score for the first set of embeddings and the second set of embeddings and (ii) a classifier recommendation for the second set of embeddings. The second set of embeddings can be generated based on data collected using condition monitoring sensors for a particular industrial machine. Accordingly, the system can generate alerts, recommendations, and/or notifications based on the automatically classified data encoded in the second set of embeddings. Incremental training techniques are disclosed for further training the neural network to minimize false positives and/or false negatives.

DECODING QUANTUM ERROR CORRECTION CODES USING TRANSFORMER NEURAL NETWORKS

Absstract of: WO2024157242A1

Transformer neural network based decoder for decoding Quantum Error Correction Codes (QECC), comprising, an input layer, a plurality of decoding layers, and an output layer. The input layer is adapted to receive initial noise estimation computed by a noise estimator for noise injected to syndrome bits of codewords encoded using QECC and transmitted over transmission channel(s) subject to interference, and create embeddings for the syndrome bits. The decoding layers adapted to compute an estimated logical operator matrix of each codeword, each comprises a self-attention layer constructed according to a mask indicative of a relation between the embeddings derived from a parity-check matrix of the error correction code. The plurality of decoding layers are trained using a combined loss function directed to minimize LER, BER, and error rate of the noise estimator. The output layer is adapted to produce a vector representing predicted soft error of the codeword's logical operator matrix.

METHODS AND SYSTEMS FOR PROBABILISTIC RELATIONAL LEARNING

Absstract of: EP4657316A1

Methods and systems for fully quantifying predictive uncertainty for recommender systems are disclosed. Embodiments include providing a fully Bayesian multiway neural network model, assigning a prior distribution over a plurality of weight parameters of a model parameter set of the neural network model such that a corresponding posterior distribution represents an epistemic uncertainty of the neural network model; incorporating a mixture density model for a target layer of the neural network model, wherein the mixture density model models aleatoric uncertainty of input data; estimating the posterior distribution of the weight parameters given training data by updating the prior distribution based on data likelihood of the training data; sampling a plurality of empirical samples from the estimated posterior distribution of the weight parameters; performing inference for each stochastic realization of the neural network model with the sampled weight parameters on an inference sample to generate one or more predictive distributions; drawing samples from each of the one or more predictive distributions as final predictions for the inference sample to obtain a matrix of predictions for the inference sample, wherein the matrix represents a joint distribution of both the aleatoric and epistemic uncertainty of the recommender system.

SYSTEMS AND METHODS FOR DEFINING CONFIDENCE IN DEEP LEARNING MODEL PREDICTION

Absstract of: EP4657324A1

The invention relates to a technique for improving confidence estimates associated with neural networks. The technique involves computing neuron activation statistics during training, evaluating neuron activations during inferencing and determining how the activations compare with the previously computed statistics (e.g. whether prediction activations are within the bounds of the training activation statistics). The comparison may be used to compute a confidence value for the neural network.

SUBGRAPH PATTERN EXTRACTION

Absstract of: US2025363328A1

Aspects of the disclosed technology provide solutions for extracting subgraph patterns in graph-structured data and encoding them as embeddings using a graph neural network (GNN). In some aspects, a process of the disclosed technology can include steps for receiving an input graph comprising a plurality of nodes and edges, the input graph representing relationships among a plurality of entities, parameterizing a graph neural network model based on a set of pattern graphs, and identifying, for at least a portion of the nodes in the input graph, rooted homomorphisms between the pattern graphs and local subgraphs rooted at the respective nodes. Systems and machine-readable media are also provided.

MACHINE LEARNING METHODS FOR PREDICTING PROPERTIES OF PROTEINS AND LIGANDS

Absstract of: US2025364082A1

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for generating a predicted property score of a protein and a ligand. In one aspect, a method comprises: obtaining a network input that characterizes a protein and a ligand; processing the network input characterizing the protein and the ligand using an embedding neural network to generate a protein-ligand embedding representing the protein and the ligand, wherein the embedding neural network has been jointly trained with a generative model that is configured to: receive an input protein-ligand embedding; and generate, while conditioned on the input protein-ligand embedding, a predicted joint three-dimensional (3D) structure of an input protein and an input ligand represented by the input protein-ligand embedding; and generating a property score that defines a predicted property of the protein and the ligand using the protein-ligand embedding.

INTEGRATED AI-POWERED ADAPTIVE ROBOTIC SURGERY SYSTEM

Absstract of: US2025359955A1

A robotic surgical system includes a robotic manipulator configured to perform surgical procedures under direct surgeon control. A surgical camera system captures real-time intraoperative video. An external imaging interface receives multimodal imaging data, including preoperative and intraoperative data from at least one of magnetic resonance imaging (MRI), computed tomography (CT), ultrasound, and fluoroscopy. An artificial intelligence (AI module has a trained neural network and a deep learning model trained on multi-institutional annotated surgical datasets, The AI module is configured to execute one or more of: fuse acquired video and imaging data into temporally and spatially coherent anatomical visualizations; generate continuously updating overlays aligned with the surgical field, with segmented anatomical features; projected tissue boundaries, proximity indicators for instruments, and predictive deformation trends; provide dynamic predictive trend visualization indicating zones of future anatomical complexity or risk; register and align preoperative imaging data with intraoperative imaging data in real time; adapt overlay presentation in response to tissue deformation without actuating the robotic manipulate or; and passively augment visual feedback without initiating any autonomous actuation of surgical instruments.

A MULTI-MODAL CLASSIFIER SYSTEM FOR MISSENSE MUTATION PATHOGENICITY PREDICTION

Absstract of: WO2025243286A1

The present invention relates to a computer-implemented multi-module classifier method and system for providing a pathogenicity classification score of a variant of a protein of interest. The classifier comprises a sequence module based on a protein language model (PLM); a structure module based on a graph neural network (GNN); a property module; and a unified head module based on a machine learning model. The invention further relates to methods for preparing, training, and implementing the multi-module classifier system.

PREDICTING PROPERTIES OF PROTEINS AND LIGANDS

Nº publicación: EP4654205A1 26/11/2025

Applicant:

ISOMORPHIC LABS LTD [GB]
Isomorphic Labs Limited

Absstract of: EP4654205A1

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for generating a predicted property score of a protein and a ligand. In one aspect, a method comprises: obtaining a network input that characterizes a protein and a ligand; processing the network input characterizing the protein and the ligand using an embedding neural network to generate a protein-ligand embedding representing the protein and the ligand, wherein the embedding neural network has been jointly trained with a generative model that is configured to: receive an input protein-ligand embedding; and generate, while conditioned on the input protein-ligand embedding, a predicted joint three-dimensional (3D) structure of an input protein and an input ligand represented by the input protein-ligand embedding; and generating a property score that defines a predicted property of the protein and the ligand using the protein-ligand embedding.