Resumen de: US2023090720A1

A computer-implemented apparatus installed and executed in a computer to search an optimal design of a neural processing unit (NPU), a hardware accelerator used for driving a computer-implemented artificial neural network (ANN) is disclosed. The NPU comprises a plurality of blocks connected in a form of pipeline, and the number of the plurality blocks and the number of the layers within each block of the plurality blocks are in need of optimization to reduce hardware resources demand and electricity power consumption of the ANN while maintaining the inference accuracy of the ANN at an acceptable level. The computer-implemented apparatus searches for and then outputs an optimal L value and an optimal C value when a first set of candidate values for a number of layers L and a second set of candidate values for a number of channels C per each layer of the ANN is provided.

Resumen de: US2023087330A1

An apparatus, method and computer program is described comprising detecting a first object in a first image of a sequence of images using a neural network, wherein the means for detecting the first object provides an object area indicative of a first location of the first object; and tracking the first object, wherein the means for tracking the first object further comprises generating a predicted future location of the first object and generating an updated location of the first object using the neural network. The means for generating the predicted future location of the first object may, for example, receive said object area indicative of a first location of the first object and may receive said updated location information of the first object.

Resumen de: US2023093630A1

In general, at least one example of an embodiment can involve selecting a neural network from a plurality of neural networks based on an indication of resource availability and processing data using the selected neural network in accordance with the resource availability.

Resumen de: US2023087364A1

One embodiment provides for a computer-readable medium storing instructions that cause one or more processors to perform operations comprising determining a per-layer scale factor to apply to tensor data associated with layers of a neural network model and converting the tensor data to converted tensor data. The tensor data may be converted from a floating point datatype to a second datatype that is an 8-bit datatype. The instructions further cause the one or more processors to generate an output tensor based on the converted tensor data and the per-layer scale factor.

Resumen de: WO2023036017A1

A speech recognition method and system for power grid dispatching. The method comprises: acquiring an original speech signal in power grid dispatching; performing noise reduction pre-processing on the original speech signal; performing a fast Fourier transform (FFT) on the original speech signal subjected to noise reduction pre-processing; performing, by using a Mel frequency cepstral coefficient (MFCC), feature extraction on the original speech signal subjected to the FFT; combining a deep learning neural network (DNN) and a long short-term memory (LSTM) neural network to obtain a combined neural network DNN-LSTM algorithm, and performing, by using the algorithm, acoustic model training on the original speech signal subjected to feature extraction; and searching for an optimal text output result by using a decoder and on the basis of an acoustic model output result, a speech model and a dictionary.

Resumen de: US2023084700A1

Methods, systems and apparatus, including computer programs encoded on computer storage media, for training an action selection neural network. One of the methods includes receiving an observation characterizing a current state of the environment; determining a target network output for the observation by performing a look ahead search of possible future states of the environment starting from the current state until the environment reaches a possible future state that satisfies one or more termination criteria, wherein the look ahead search is guided by the neural network in accordance with current values of the network parameters; selecting an action to be performed by the agent in response to the observation using the target network output generated by performing the look ahead search; and storing, in an exploration history data store, the target network output in association with the observation for use in updating the current values of the network parameters.

Resumen de: US2023080172A1

In some examples, transmitting node instructions can include a logical node of a neural network to: receive a first node instruction, generate an output change value based on the received first node instruction, and transmit the output change value and the first node instruction to a logical edge, and the logical edge of the neural network to: receive the output change value and the first node instruction from the logical node, generate a second node instruction based on the output change value and the first node instruction, and transmit the second node instruction.

Resumen de: US2023083929A1

A modular architecture is provided for denoising Monte Carlo renderings using neural networks. The temporal approach extracts and combines feature representations from neighboring frames rather than building a temporal context using recurrent connections. A multiscale architecture includes separate single-frame or temporal denoising modules for individual scales, and one or more scale compositor neural networks configured to adaptively blend individual scales. An error-predicting module is configured to produce adaptive sampling maps for a renderer to achieve more uniform residual noise distribution. An asymmetric loss function may be used for training the neural networks, which can provide control over the variance-bias trade-off during denoising.

Resumen de: US2023085599A1

The disclosure provides a method for training a tag recommendation model. The method includes: collecting training materials that comprise interest tags in response to receiving an instruction for collecting training materials; obtaining training semantic vectors that comprise the interest tags by representing features of the training materials using a semantic enhanced representation frame; obtaining training encoding vectors by aggregating social networks into the training semantic vectors; and obtaining a tag recommendation model by training a double-layer neural network structure using the training encoding vectors as inputs and the interest tags as outputs. Therefore, the interest tags obtained in the disclosure are more accurate.

Resumen de: GB2610790A

Technology for using a bi-directed graph convolution neural network ("BGCNN") to convert Resource Description Framework (RDF) data into natural language text. Some embodiments perform RDF-to-Text generation by learning graph-augmented structural neural encoders, consisting of: (a) bi-directed graph-based meta-paths encoder; (b) bi-directed graph convolution networks encoder; and (c) separated attention mechanism for combining encoders and decoder to translate RDF triplets to natural language description.

Resumen de: US2023075116A1

Computer-implemented methods, devices, and systems for matching a requester with at least one requestee. An input representing a request from the requester is received. A top-k learning tree-based deep model is applied to at least one node embedding representing at least one requestee successively across at least one time window. Each output from each time window is concatenated with the input into a neural network input. A neural network determines each probability for the input to match each requestee. At least one requestee is presented to the requester for approval based on the probability for the input to match the requestee.

Resumen de: US2023075716A1

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for sequence modeling. One of the methods includes receiving an input sequence having a plurality of input positions; determining a plurality of blocks of consecutive input positions; processing the input sequence using a neural network to generate a latent alignment, comprising, at each of a plurality of input time steps: receiving a partial latent alignment from a previous input time step; selecting an input position in each block, wherein the token at the selected input position of the partial latent alignment in each block is a mask token; and processing the partial latent alignment and the input sequence using the neural network to generate a new latent alignment, wherein the new latent alignment comprises, at the selected input position in each block, an output token or a blank token; and generating, using the latent alignment, an output sequence.

Resumen de: US2023071424A1

Techniques in placement of compute and memory for accelerated deep learning provide improvements in one or more of accuracy, performance, and energy efficiency. An array of processing elements comprising a portion of a neural network accelerator performs flow-based computations on wavelets of data. Each processing element comprises a compute element to execute programmed instructions using the data and a router to route the wavelets. The routing is in accordance with virtual channel specifiers of the wavelets and controlled by routing configuration information of the router. A software stack determines placement of compute resources and memory resources based on a description of a neural network. The determined placement is used to configure the routers including usage of the respective colors. The determined placement is used to configure the compute elements including the respective programmed instructions each is configured to execute.

Resumen de: US2023076562A1

A farming machine including a number of treatment mechanisms treats plants according to a treatment plan as the farming machine moves through the field. The control system of the farming machine executes a plant identification model configured to identify plants in the field for treatment. The control system generates a treatment map identifying which treatment mechanisms to actuate to treat the plants in the field. To generate a treatment map, the farming machine captures an image of plants, processes the image to identify plants, and generates a treatment map. The plant identification model can be a convolutional neural network having an input layer, an identification layer, and an output layer. The input layer has the dimensionality of the image, the identification layer has a greatly reduced dimensionality, and the output layer has the dimensionality of the treatment mechanisms.

Resumen de: US2023071810A1

A system and method for providing spatiotemporal costmap inference for model predictive control that includes receiving dynamic based data and environment based data to determine observations and goal information associated with an ego agent and a traffic environment. The system and method also include training a neural network with the observations and goal information and determining an optimal path of the ego agent based on at least one spatiotemporal costmap. The system and method further include controlling the ego agent to autonomously operate based on the optimal path of the ego agent.

Resumen de: US2023072242A1

A health condition or change in health condition of a person may be determined by processing the person's speech with a neural network. Speech from more than one time period may be processed and, in some implementations, speech from a time period may be associated with a health condition label. For each time period, a feature vector may be computed from the speech and the feature vector may be processed with a neural network to obtain a speech embedding vector. In some implementations, feature vector may include word-piece encodings and the neural network may be a transformer neural network. The speech embedding vectors may be processed with a mathematical model to determine a change in a health condition between two time periods or to determine a health condition label for a specific time period. In some implementations, the mathematical model may be a regression model or a fully-connected neural network.

Resumen de: US2023071730A1

An artificial intelligence framework is disclosed having a spiking neural network configured to extract low-level features from event-based spiking data and provide the low-level features as a spiking neural network output signal. Further included is a hyperdimensional computing block that is configured to map the spiking neural network output into high-dimensional space and classify data from the abstract information.

Resumen de: US2023072438A1

This application provides a method for training a neural network model and an apparatus. The method includes: obtaining annotation data that is of a service and that is generated by a terminal device in a specified period; training a second neural network model by using the annotation data that is of the service and that is generated in the specified period, to obtain a trained second neural network model; and updating a first neural network model based on the trained second neural network model. In the method, training is performed based on the annotation data generated by the terminal device, so that in an updated first neural network model compared with a universal model, an inference result has a higher confidence level, and a personalized requirement of a user can be better met.

Resumen de: WO2023033538A1

Disclosed are an electronic device and a controlling method of the electronic device. Particularly, the electronic device according to the present disclosure comprises: a memory for storing a neural network model; and a processor obtaining a plurality of individual graphs showing the access history of a user with respect to a plurality of content items for each of a plurality of sessions, generating an integrated graph in which the plurality of individual graphs are integrated on the basis of the connection relationship of nodes included in the plurality of individual graphs and the number of repetitions of the connections of the nodes, obtaining a plurality of augmented graphs by augmenting each of the plurality of individual graphs on the basis of the integrated graph, and training the neural network model for providing recommended content, on the basis of the plurality of augmented graphs.

Resumen de: WO2023033280A1

According to an embodiment of the present disclosure, a data sampling method for active learning performed by a computing device comprising at least one processor may comprise the steps of: generating normalized feature vectors for an unlabeled data set on the basis of a neural network model; estimating the density of the normalized feature vectors by grouping the normalized feature vectors on a vector space; and extracting query data for active learning from the unlabeled data set on the basis of the estimated density.

Resumen de: WO2023033282A1

Disclosed is a method for training a multi-task model, performed by a computing device comprising at least one processor, according to some embodiments of the present disclosure. The method for training a multi-task model may comprise the steps of: acquiring a training data set; and on the basis of the training data set, training a neural network model for outputting a result of prediction of an input value and estimating uncertainty of the prediction, wherein a loss function for training the neural network model includes a first loss function for quantifying the prediction result and the uncertainty of the prediction, and a second loss function for improving the prediction accuracy of the neural network model.

Resumen de: WO2023033194A1

Disclosed are a knowledge distillation method and system specialized for pruning-based deep neural network lightening. The knowledge distillation method according to one embodiment may comprise the steps of: pruning a backbone neural network included in a neural network model for lightening; and performing knowledge distillation on the backbone neural network that has been pruned. The step of performing knowledge distillation may include a step for distilling feature information of a teacher backbone neural network, which is the backbone neural network of the original neural network for an input, into a student backbone neural network, which is the backbone neural network that has been pruned.

Resumen de: WO2023031544A1

Said method makes it possible to configure the weights of neural network models of nodes from a set of nodes of a communication network, the neural networks all having a model with the same structure. It comprises: - partitioning the set of nodes into at least one cluster of nodes; - sending, to a node belonging to the at least one cluster, an item of information according to which the node should act as an aggregation node in the cluster and identifiers of the nodes of the cluster, the node subsequently being referred to as the aggregation node of the cluster; - sending, to the aggregation node of the at least one cluster, a request for learning the weights of the node models of the cluster with the weights of a global model for the set of nodes; - a step of receiving, from the aggregation node of the at least one cluster, the weights of an aggregated model of the cluster resulting from the training; and - updating the weights of the global model by aggregating the weights received from the aggregated model of the at least one cluster.

Resumen de: WO2023035009A1

A predictive management system and method for a machine are disclosed. A controller receives a machine data signal and is configured to predict a future event of the machine from a plurality of future events based upon the received data signals. The controller also, upon the predicted future event being a first failure event, sends instructions to a user interface to display an identification and explanation of the first failure event, and at least one data signal during a defined time interval in which the data signal is associated with the first failure event. The controller includes a trained neural network receives the data signal and includes one or more output layers identifying the predicted future event, a segmentation mask which determines the defined time interval, and a determining block which, based upon the defined time interval, determines the explanation associated with the first failure event in human readable form.

Resumen de: US2023062217A1

Hardware neural network processors, are provided. A neural core includes a weight memory, an activation memory, a vector-matrix multiplier, and a vector processor. The vector-matrix multiplier is adapted to receive a weight matrix from the weight memory, receive an activation vector from the activation memory, and compute a vector-matrix multiplication of the weight matrix and the activation vector. The vector processor is adapted to receive one or more input vector from one or more vector source and perform one or more vector functions on the one or more input vector to yield an output vector. In some embodiments a programmable controller is adapted to configure and operate the neural core.