Absstract of: US20260350343A1
0000 Apparatuses, systems, and techniques to perform neural networks. For example, one or more processors cause a second portion of a neural network to be loaded into memory concurrently with inference performed using a first portion of the neural network; and cause the loaded second portion to be used to perform inference using output of the first portion of the neural network.
Absstract of: US20260350380A1
0000 Various examples relate to dynamic parallelism for language model inference. A pool of processing units is reconfigured between different parallelism modes during inference. In a first mode for attention computation, contextual memory data is distributed across processing units. After attention computation, the units are reprovisioned into a second mode with multi-dimensional parallelism, having tensor parallelism and sparse computation dimensions. The second mode enables processing tokens through neural networks in a sparse computation architecture. For dense models without mixture-of-experts, the sparse computation dimension equals one, allowing graceful degradation to tensor-parallel execution. Collective communication operations may facilitate transitions between modes while maintaining mathematical equivalence with non-distributed computation.
Absstract of: WO2026232172A2
A method for utilizing a Structural Receptive Set (SRS) structure in a trained neural network model includes accessing learned weight tensors of a trained neural network layer comprising input channels and output channels. Semantic Core (SC) channels are identified from the output channels based on the learned weight tensors. For an SC channel, a Structure Auxiliary (AC) channel set is identified based on structural similarity among output-channel weight tensors. An SRS comprising a subset of the input channels structurally used to compute an SC/AC output group is determined for the SC channel and the AC channel set. An Invariant Structure Channel Mapping (ISCM) group comprising the SC channel, the AC channel set, and the SRS is generated and used to produce a sparse input-output dependency structure that enables the SC/AC output group to be computed from fewer than all input channels.
Absstract of: US20260350739A1
Systems and methods for performing a zoom operation on a display of a computing device, which include identifying a region of interest (ROI) of an image presented on the display based on a zoom user input, determining a scaling factor for a final zoomed presentation of the ROI based on an image rendering requirement and/or an inference speed of a frame generation convolutional neural network (CNN), using the CNN to generate the final zoomed presentation of the ROI to the determined scaling factor, down-scaling the final zoomed presentation to generate a sequence of intermediate size image frames each having dimensions smaller than dimensions of the final zoomed presentation of the ROI and greater than dimensions of the ROI, and rendering an animation of changes in a displayed size of the ROI using the sequence of intermediate image frames of the ROI.
Absstract of: WO2026227712A1
Provided is a reconstruction error-type fault detection method based on a graph neural network. The method comprises: obtaining time series data of industrial equipment, the time series data comprising features of a plurality of nodes; processing the time series data by means of a convolutional neural network to obtain a time feature of each node; constructing a graph structure on the basis of the time series data, the graph structure representing spatial association between the nodes; processing the graph structure and the time feature by means of a graph neural network to obtain a spatial feature of each node; reconstructing the spatial features by means of a decoder to obtain a reconstructed output; and determining a fault detection result on the basis of the reconstructed output and the time series data. The reconstruction error-type fault detection method based on the graph neural network improves the accuracy and robustness of fault detection, is suitable for real-time monitoring and predictive maintenance of complex industrial systems, and has important practical value.
Absstract of: US20260351048A1
A computer-implemented method receives an instruction that is optionally interpreted by using a trained computer-implemented neural network and accesses video-related user behavior information. The video-related user behavior information includes user interaction or attention with respect to objects, object-related attributes, or object-related actions that are represented by distinct subsets of pixels within at least one video. An action is determined based on the instruction and an inference of a preference, interest, or intent of the user that is derived from the video-related user behavior information, and optionally, the action is further based on non-video-related user behaviors. The action can be performed by invoking an agent through an API.
Absstract of: US20260350392A1
Disclosed herein are a method and apparatus for compressing learning parameters for training of a deep-learning model and transmitting the compressed parameters in a distributed processing environment. Multiple electronic devices in the distributed processing system perform training of a neural network. By performing training, parameters are updated. The electronic device may share the updated parameter thereof with additional electronic devices. In order to efficiently share the parameter, the residual of the parameter is provided to the additional electronic devices. When the residual of the parameter is provided, the additional electronic devices update the parameter using the residual of the parameter.
Absstract of: US20260350399A1
0000 An encoder neural network is described which can encode a data item, such as a frame of a video, to form a respective encoded data item. Data items of a first data sequence are associated with respective data items of a second sequence, by determining which of the encoded data items of the second sequence is closest to the encoded data item produced from each data item of the first sequence. Thus, the two data sequences are aligned. The encoder neural network is trained automatically using a training set of data sequences, by an iterative process of successively increasing cycle consistency between pairs of the data sequences.
Absstract of: US20260350357A1
0000 Systems and methods that use at least one neural network to infer content of individual frames in a sequence of images and to further infer changes to content in sequence of images over time to determine whether one or more anomalous events are present in sequence of images is described herein.
Absstract of: US20260350975A1
0000 Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training Vision Transformer (ViT) neural networks.
Absstract of: US20260350393A1
0000 Neural network performance is improved in terms of training speed and/or accuracy by encoding (mapping) inputs to the neural network into a higher dimensional space via a hash function. The input comprises coordinates used to identify a point within a d-dimensional space (e.g., 3D space). The point is quantized and a set of vertex coordinates corresponding to the point are input to a hash function. For example, for d=3, space may be partitioned into axis-aligned voxels of identical size and vertex coordinates of a voxel containing the point are input to the hash function to produce a set of encoded coordinates. The set of encoded coordinates is used to lookup D-dimensional feature vectors in a table of size T that have been learned. The learned feature vectors are filtered (e.g., linearly interpolated, etc.) based on the coordinates of the point to compute a feature vector corresponding to the point.
Absstract of: US20260350344A1
Disclosed are a compilation and inference method for a neural network model, and a device and a medium. The compilation method comprises: acquiring a first model corresponding to a neural network model to be subjected to compilation; using, as a target network layer, any one of at least one network layer which includes a preset operation in the first model, and converting the preset operation of the target network layer into a convolution operation which is supported by a neural network processor, so as to obtain a converted first network layer, wherein the preset operation is an operation comprising a multiply—accumulate operation; on the basis of the converted first network layer, obtaining an optimization model corresponding to the first model; and according to the optimization model, generating a target instruction sequence that is executable by the neural network processor and corresponds to said neural network model.
Absstract of: US20260350360A1
Techniques for mapping human-understandable attributes to a neural network are disclosed. A system generates human-understandable explanations for decisions or predictions of a neural network by mapping human-understandable attributes to the trained neural network. The system may map a human-understandable attribute to a particular label, indicating a correlation between the particular human-understandable attribute and the particular label. In addition, or in the alternative, the system may map the human-understandable attribute to a particular layer of the neural network, such as a particular hidden layer, indicating a correlation between the human-understandable attribute and the particular layer.
Absstract of: US20260350783A1
0000 Disclosed herein is a medical system (100, 300). The execution of machine executable instructions (120) causes a computational system to (104): receive (200) an overview image (128), receive (202) object locators (130) in response to inputting the overview image into a first neural network (122), construct (204) an object sub image (132) using the object locator, receive (206) object positional data (134) in response to inputting the object sub image for each of the multiple predetermined objects into an object specific neural network (124), construct (208) multiple search images (136) using the object positional data for the multiple predetermined objects, receive (210) locator data (138) descriptive of a position of an item in response to inputting each of the multiple search images into an item specific neural network (126), detect (212) a safety state (140) by comparing the locator data and the object positional data of each of the predetermined objects with respect to a predetermined criterion; and provide (214) a warning signal (144) if the safety state is detected
Absstract of: US20260346145A1
0000 Apparatuses, systems, and techniques to provide emotion-based experience adjustment. In at least one embodiment, one or more neural networks are used to adjust one or more properties of one or more games based on an emotional state of a user.
Absstract of: US20260350385A1
Systems, methods, and computer-readable storage media for generating tests offering improved screening of subjects. A system can receive historical data associated with multiple variables and identify a cutoff value for each variable. The system can also generate a neural network using the historical data. The system can then generate, using the cutoff values and the neural network, one or more combination tests using multiple variables and the cutoff values for those variables. When a new sample is received, the system can then execute the one or more combination tests using the new sample's data, resulting in a prediction regarding the new sample. The one or more combination tests can then be tested and combined to form an ensemble test.
Absstract of: US20260350770A1
Apparatuses, systems, and techniques are presented to determine visual properties of an object. In at least one embodiment, shape and visual properties of an object can be determined using differently illuminated images fed to a pipeline of neural networks.
Absstract of: WO2026229971A1
Systems and methods for automatically analyzing movements to identify anomalies and incorrect actions. An artificial intelligence-based digital system is provided that can monitor PT exercises. The systems and methods include a scalable solution to detect incorrect actions or anomalies in PT exercises in real time without using manual definitions. Furthermore, the system includes explanations for detected anomalies, enhancing the real-time feedback and adjustment of treatment plans. Providing real-time feedback improves treatment results by ensuring accurate performance of the exercise by the patient. The systems and methods can include extracting skeleton sequences of an exerciser performing an exercise from 2D videos. A neural network classifies the exercise being performed and identifies the side of the body performing the exercise. The exercise performed by the exerciser can be compared to an expert version of the exercise, and anomalies in the exercise performed by the exerciser can be detected.
Absstract of: US20260350358A1
0000 A method is described for designing systems that provide efficient implementations of feedforward, recurrent, and deep networks that process dynamic signals using temporal filters and static or time-varying nonlinearities. A system design methodology is described that provides an engineered architecture. This architecture defines a core set of network components and operations for efficient computation of dynamic signals using temporal filters and static or time-varying nonlinearities. These methods apply to wide variety of connected nonlinearities that include temporal filters in the connections. Here we apply the methods to synaptic models coupled with spiking and/or non-spiking neurons whose connection parameters are determined using a variety of methods of optimization.
Absstract of: CA3329704A1
The present invention relates to a multi-node cluster-based inference method through GPU separation allocation of a pre-trained layer and a fine-tuning layer of multiple deep learning models. According to an embodiment of the present disclosure, the method includes: receiving, by an endpoint unit of an inference server, an input value from a client; distributing, by the endpoint unit of the inference server, the received input value, and transmitting the first input value to a first computation node including a container in which a neural network bundle of a first stage is loaded; performing, by a first container of the first computation nodes, an operation through a neural network layer of a GPU by using the received first input value as an input, and generating a first output value; and selecting, by the first container, a container in which the neural network bundle of the next stage is loaded, and transmitting the second output value to the computation node that includes the container in which the next stage is to be executed or the container in which to execute the next stage.
Absstract of: US20260350394A1
0000 A computer-implemented method for training a neural network. The neural network detects objects in images. The training includes: ascertaining a feature representation of a training image using a backbone of the neural network; ascertaining head outputs based on the feature representation and using heads of the neural network, wherein the head outputs, for regions of the training image, each characterize a presence or absence of an object or characterize an extension of the object or characterize a position of the object or characterize the extension and the position; ascertaining first loss values, wherein each first loss value corresponds to a head output and characterizes a deviation of the head output from a desired head output; adjusting a parameter(s) of the backbone and/or at least one parameter of a head based on a second loss value, wherein the second loss value characterizes a sum of the first loss values.
Absstract of: EP4836429A2
0001 A system and method of calibrating moving cameras capturing a sporting event is disclosed herein. A computing system retrieves a broadcast video feed for a sporting event. The broadcast video feed includes a plurality of video frames. The computing system labels, via a neural network, components of a playing surface captured in each video frame. The computing system matches a subset of labeled video frames to a set of templates with various camera perspectives. The computing system fits a playing surface model to the set of labeled video frames that were matched to the set of templates. The computing system identifies camera motion in each video frame using an optical flow model. The computing system generates a homography matrix for each video frame based on the fitted playing surface model and camera motion. The computing system calibrates each camera based on the homography matrix generated for each video frame.
Absstract of: WO2025141279A1
The invention relates to a continuous learning method for a facial comparison biometric neural network, wherein the method comprises the following operations: a) receiving an already-trained facial comparison biometric neural network; b) separating the neural network from operation a) into a characterisation block and an output block; c) separating the characterisation block into an extraction block and an adaptation block; d) receiving two separate sets of training data formatted as the inputs to the neural network from operation a); e) supplying the second set to the extraction block (230) and deriving a third set therefrom; and f) training the neural network from operation a) by freezing the extraction block (230) and using the first set of data with the extraction block (230) in its frozen state, and using the third set by supplying it directly to the adaptation block.
Absstract of: EP4835746A1
Embodiments of this application disclose a large language model inference method and apparatus to reduce bandwidth consumption of large language model inference. The method in embodiments of this application includes: A parallel inference system receives an inference task of a large language model, where the inference task includes a compute-intensive task and a self-attention compute task. A processor executes the compute-intensive task to generate intermediate data and sends the intermediate data to a near-storage computing module, where the compute-intensive task includes one or more of the following tasks of the large language model: a feed-forward neural network compute task, a projection task, and a layer normalization task. The near-storage computing module executes the self-attention compute task based on the intermediate data to generate a near-storage computation result. The processor generates, based on the near-storage computation result, an inference result corresponding to the inference task.
Nº publicación: EP4835739A2 04/11/2026
Applicant:
FRAUNHOFER GES FORSCHUNG [DE]
Fraunhofer-Gesellschaft zur F\u00F6rderung der angewandten Forschung e.V.
Absstract of: EP4835739A2
0001 Embodiments according to the invention comprise a decoder for decoding parameters of a neural network, wherein the decoder is configured to obtain a plurality of neural network parameters of the neural network on the basis of an encoded bitstream. Furthermore, the decoder is configured to obtain, e.g. to receive; e.g. to extract from an encoded bitstream, a node information describing a node of a parameter update tree, wherein the node information comprises a parent node identifier, which is, for example, a unique parent node identifier, for example an integer number, a string, and/or a cryptographic hash, and wherein the node information comprises a parameter update information, e.g. one or more update instructions, for example a difference signal between initial neural network parameters and a newer version thereof, e.g. corresponding to a child node of the update tree. 0002 Moreover, the decoder is configured to derive one or more neural network parameters using parameter information of a parent node (the parameter information comprising, for example a node information of the parent node, the node information for example comprising a parameter update information and a parent node identifier of the parent node, e.g. for a recursive reconstruction or recursive determination or recursive calculation or recursive derivation of the one or more neural network parameters and/or for example comprising a node parameter of the parent node, e.g. neural network parameters a