Neural networks

Image signal processing

Absstract of: GB2631468A

The present disclosure provides an inference device e.g. a neural network accelerator, configured to implement a command stream representing a differentiable model of an image signal processor as a combination of operations from a set of elementary neural network operations which are available on the inference device. The image signal processor has a pipeline of two or more functional blocks and the differentiable model of the image signal processor comprises at least two differentiable modules, each of the at least two differentiable modules configured to implement a respective single functional block of the pipeline. The disclosure also provides a corresponding method.

COMPUTING RAY PATH BETWEEN SOURCE ANTENNA LOCATION AND DESTINATION ANTENNA LOCATION

Absstract of: EP4488867A1

A computing system (10) including a processor (14) configured to receive a mesh (30) of a three-dimensional geometry (38). The processor is further configured to receive a source antenna location (40) and a destination antenna location (42) on the mesh. The processor is further configured to compute a ray path (60) as an estimated shortest path between the source antenna location and the destination antenna location. The ray path includes a geodesic path (62) over the mesh and a free space path (64) outside the mesh. The ray path is computed at least in part by computing the geodesic path at least in part by performing inferencing at a trained neural network (52). Computing the ray path further includes computing the free space path at least in part by performing raytracing from a launch point (66) located at an endpoint of the geodesic path. The processor is further configured to output the ray path to an additional computing process (70).

MACHINE LEARNING MODEL WITH DEPTH PROCESSING UNITS

Absstract of: US2025005339A1

Representative embodiments disclose machine learning classifiers used in scenarios such as speech recognition, image captioning, machine translation, or other sequence-to-sequence embodiments. The machine learning classifiers have a plurality of time layers, each layer having a time processing block and a depth processing block. The time processing block is a recurrent neural network such as a Long Short Term Memory (LSTM) network. The depth processing blocks can be an LSTM network, a gated Deep Neural Network (DNN) or a maxout DNN. The depth processing blocks account for the hidden states of each time layer and uses summarized layer information for final input signal feature classification. An attention layer can also be used between the top depth processing block and the output layer.

METHOD, APPARATUS, AND COMPUTER-READABLE RECORDING MEDIUM FOR TRAINING GRAPH NEURAL NETWORK MODEL BY USING GRAPH DATA

Absstract of: WO2025005325A1

A method for training a graph neural network (GNN) model by using graph data, according to an embodiment of the present disclosure, comprises the operations of: acquiring graph data including node data regarding nodes constituting a graph and edge data regarding a connection relationship between the nodes; on the basis of the graph data, configuring training data used to train the GNN model; and training the GNN model by using the training data, wherein the training data includes a virtual graph comprising existing nodes and virtual nodes.

DYNAMIC PRUNING OF NEURONS ON-THE-FLY TO ACCELERATE NEURAL NETWORK INFERENCES

Absstract of: US2025005364A1

Systems, apparatuses and methods may provide for technology that aggregates contextual information from a first network layer in a neural network having a second network layer coupled to an output of the first network layer, wherein the context information is to be aggregated in real-time and after a training of the neural network, and wherein the context information is to include channel values. Additionally, the technology may conduct an importance classification of the aggregated context information and selectively exclude one or more channels in the first network layer from consideration by the second network layer based on the importance classification.

DASH CAM WITH ARTIFICIAL INTELLIGENCE SAFETY EVENT DETECTION

Absstract of: US2025002033A1

A vehicle dash cam may be configured to execute one or more neural networks (and/or other artificial intelligence), such as based on input from one or more of the cameras and/or other sensors associated with the dash cam, to intelligently detect safety events in real-time. Detection of a safety event may trigger an in-cab alert to make the driver aware of the safety risk. The dash cam may include logic for determining which asset data to transmit to a backend server in response to detection of a safety event, as well as which asset data to transmit to the backend server in response to analysis of sensor data that did not trigger a safety event. The asset data transmitted to the backend server may be further analyzed to determine if further alerts should be provided to the driver and/or to a safety manager.

SYSTEM AND METHODS FOR MAMMALIAN TRANSFER LEARNING

Absstract of: US2025006362A1

A neural network is trained using transfer learning to analyze medical image data, including 2D, 3D, and 4D images and models. Where the target medical image data is associated with a species or problem class for which there is not sufficient labeled data available for training, the system may create enhanced training datasets by selecting labeled data from other species, and/or labeled data from different problem classes. During training and analysis, image data is chunked into portions that are small enough to obfuscate the species source, while being large enough to preserve meaningful context related to the problem class (e.g., the image portion is small enough that it cannot be determined whether it is from a human or canine, but abnormal liver tissues are still identifiable). A trained checkpoint may then be used to provide automated analysis and heat mapping of input images via a cloud platform or other application.

PERFORMING TASKS USING GENERATIVE NEURAL NETWORKS

Absstract of: WO2024263935A1

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for performing tasks. One of the methods includes obtaining a sequence of input tokens, where each token is selected from a vocabulary of tokens that includes text tokens and audio tokens, and wherein the sequence of input tokens includes tokens that describe a task to be performed and data for performing the task; generating a sequence of embeddings by embedding each token in the sequence of input tokens in an embedding space; and processing the sequence of embeddings using a language model neural network to generate a sequence of output tokens for the task, where each token is selected from the vocabulary.

PERFORMING TASKS USING GENERATIVE NEURAL NETWORKS

Absstract of: US2024428056A1

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for performing tasks. One of the methods includes obtaining a sequence of input tokens, where each token is selected from a vocabulary of tokens that includes text tokens and audio tokens, and wherein the sequence of input tokens includes tokens that describe a task to be performed and data for performing the task; generating a sequence of embeddings by embedding each token in the sequence of input tokens in an embedding space; and processing the sequence of embeddings using a language model neural network to generate a sequence of output tokens for the task, where each token is selected from the vocabulary.

DYNAMIC CULLING OF MATRIX OPERATIONS

Absstract of: US2024426614A1

An output of a first one of a plurality of layers within a neural network is identified. A bitmap is determined from the output, the bitmap including a binary matrix. A particular subset of operations for a second one of the plurality of layers is determined to be skipped based on the bitmap. Operations are performed for the second layer other than the particular subset of operations, while the particular subset of operations are skipped.

SYSTEM AND METHOD FOR OPTIMIZING NON-LINEAR CONSTRAINTS OF AN INDUSTRIAL PROCESS UNIT

Absstract of: US2024419136A1

The present invention provides a robust and effective solution to an entity or an organization by enabling them to implement a system for facilitating creation of a digital twin of a process unit which can perform constrained optimization of control parameters to minimize or maximize an objective function. The system can capture non-linearities of the industrial process while the current Industrial Process models try to approximate non-linear process using linear approximation, which are not as accurate as Neural Networks. The proposed system can further create an end-to-end differentiable digital twin model of a process unit, and uses gradient flows for optimization as compared to other digital twin models that are gradient-free.

INFERENCE METHOD AND DEVICE USING SPIKING NEURAL NETWORK

Absstract of: US2024419969A1

Embodiments relate to an inference method and device using a spiking neural network including parameters determined using an analog-valued neural network (ANN). The spiking neural network used in the inference method and device includes an artificial neuron that may have a negative membrane potential or have a pre-charged membrane potential. Additionally, an inference operation by the inference method and device is performed after a predetermined time from an operating time point of the spiking neural network.

OPTIMIZING NEURAL NETWORK STRUCTURES FOR EMBEDDED SYSTEMS

Absstract of: US2024419968A1

A model training and implementation pipeline trains models for individual embedded systems. The pipeline iterates through multiple models and estimates the performance of the models. During a model generation stage, the pipeline translates the description of the model together with the model parameters into an intermediate representation in a language that is compatible with a virtual machine. The intermediate representation is agnostic or independent to the configuration of the target platform. During a model performance estimation stage, the pipeline evaluates the performance of the models without training the models. Based on the analysis of the performance of the untrained models, a subset of models is selected. The selected models are then trained and the performance of the trained models are analyzed. Based on the analysis of the performance of the trained models, a single model is selected for deployment to the target platform.

JOINT AUTOMATIC SPEECH RECOGNITION AND SPEAKER DIARIZATION

Absstract of: US2024420701A1

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for processing audio data using neural networks.

TWO-PASS END TO END SPEECH RECOGNITION

Absstract of: US2024420687A1

Two-pass automatic speech recognition (ASR) models can be used to perform streaming on-device ASR to generate a text representation of an utterance captured in audio data. Various implementations include a first-pass portion of the ASR model used to generate streaming candidate recognition(s) of an utterance captured in audio data. For example, the first-pass portion can include a recurrent neural network transformer (RNN-T) decoder. Various implementations include a second-pass portion of the ASR model used to revise the streaming candidate recognition(s) of the utterance and generate a text representation of the utterance. For example, the second-pass portion can include a listen attend spell (LAS) decoder. Various implementations include a shared encoder shared between the RNN-T decoder and the LAS decoder.

Neural network architecture for transaction data processing

Absstract of: GB2631068A

A machine learning system 600 for processing transaction data. The machine learning system 600 has a first processing stage 603 with: an interface 622 to receive a vector representation of a previous state for the first processing stage; a time difference encoding 628 to generate a vector representation of a time difference between the current and previous iterations/transactions. Combinatory logic (632, fig 6b) modifies the vector representation of the previous state based on the time difference encoding. Logic (634, fig 6b) combines the modified vector representation and a representation of the current transaction data to generate a vector representation of the current state. The machine learning system 600 also has a second processing stage 604 with a neural network architecture 660, 690 to receive data from the first processing stage and to map said data to a scalar value 602 representative of a likelihood that the proposed transaction presents an anomaly within a sequence of actions. The scalar value is used to determine whether to approve or decline the proposed transaction. The first stage may comprise a recurrent neural network and the second stage may comprise multiple attention heads.

METHOD FOR OPTIMAL VEHICLE DELIVERY OPERATION BASED ON GRAPH ARTIFICIAL NEURAL NETWORK

Nº publicación: KR20240174418A 17/12/2024

Applicant:

한국과학기술원