243 results about "Artificial neuronal network" patented technology

The application discloses a kind of high-pressure fuel filter prediction model construction method and flow prediction method, including two prediction models, method specifically includes: using computational fluid dynamics method carries out fluid dynamics simulation calculation to high-pressure fuel filter, constructs working condition sample data set;The size of filter hole, quantity, filter working pressure, fuel viscosity and filter pressure drop are used as input variable, and filter flow is used as output variable, two new high-pressure filter general nonlinear flow prediction models are established using artificial neural network and regression prediction method respectively, the flow characteristics of different models of filter in working interval can be accurately predicted using the neural network prediction method proposed in the application at any working condition point, and the error is low;The flow of different models of filter under different conditions can be quickly and accurately obtained, the accuracy meets the engineering needs, while greatly shortening the experimental time and calculation time, saving design cost.

Methods and systems for disambiguating musical artist names are disclosed. Musical-artist-release records (MARRs) may be input to a multi-modal artificial neural network (ANN). Each MARR may be associated with a musical release of an artist, and may include a release ID and an artist ID, and release data in categories including music media content and metadata categories including sub-definitive musician name of the artist and release subcategories. All n-tuples of MARRs may be formed, and for each n-tuple, the ANN may be applied concurrently to each MARR to generate a release feature vector (RFV) that includes a set of sub-feature vectors, each characterizing a different category of release data. For each n-tuple, the ANN may be trained to cluster in a multi-dimensional RFV space RFVs of the same artist ID, and to separate RFVs of different artist IDs. The MARRs and their RFVs may be stored in a release database.

There is provided a compression method implemented by at least one hardware processor, including: acquiring attention data generated in a layer of a deep learning model based on an attention mechanism; converting a data format of the attention data into a converted data format, wherein a bit pattern of the attention data is maintained in converting the data format of the attention data into the converted data format; realigning, based on data characteristics of input data of the deep learning model, the attention data in the converted data format; and generating first compressed data by inputting the realigned attention data to an encoder model based on an artificial neural network.

A neural processing unit for performing inference operations of a large-scale language model based on an artificial neural network is disclosed. The neural processing unit according to the present disclosure includes a processing element core configured to perform an attention mechanism-based operation based on input data in vector format to output an operation result, a special function unit comprising a plurality of arithmetic circuits including at least one vector-dedicated arithmetic circuit that exclusively performs vector operations and at least one mixed arithmetic circuit capable of performing both vector and scalar operations, and configured to perform a special function operation on the operation result, and a controller configured to, upon receiving an RMS normalization operation execution command, activate at least one of the plurality of arithmetic circuits to control the special function unit to perform an operation of converting at least one of the operation result or the input data into a normalized vector whose magnitude is adjusted based on a root mean square (RMS), wherein the operation result may include an attention score for the input data.

The present invention relates to a method for detecting and removing a motion artifact of a functional near-infrared spectroscopy signal. Disclosed are a method and a device for detecting and removing a motion artifact of a near-infrared spectroscopy signal in real time on the basis of an artificial neural network model, the method comprising: a conversion step of converting time series data of functional near-infrared spectroscopy signals measured through a plurality of channels from a target into image data; a detection step of detecting whether noise is present in the converted image data using a pretrained detection model; and a removal step of removing noise from image data where noise has been detected using a pretrained noise removal model.

This disclosure provides a message delivery method, etc., that is executed by at least one processor. [Solution] In one embodiment of the present disclosure, a message delivery method performed by at least one processor includes: identifying an event relating to a message request in an instant messaging application; using an artificial neural network model to obtain second information relating to at least one of a context relating to the time of event identification or a prompt relating to the event from first information relating to a user account connected to the instant messaging application; using an artificial neural network model to generate a first message based on the second information; and providing the first message through a first message room of the instant messaging application relating to the artificial neural network model.

A method for training an artificial neural network, particularly a Bayesian neural network, using a training data set includes a step of matching the parameters of the artificial neural network according to a loss function. This loss function includes a first term representing an estimate of the lower bound of the distance between the classification of the training data set by the artificial neural network and the desired classification of the training data set. Furthermore, the loss function includes a second term configured to adjust for differences in random uncertainty in the training data set by different samples passed through the artificial neural network.

A method of determining deformation of a battery includes obtaining a first image by scanning a cross-section of a battery cell in one direction, inputting the first image into an artificial neural network model trained to distinguish a plurality of parts of the battery cell in the first image and obtaining coordinates corresponding to each of the parts, and generating a second image in which at least some of the coordinates are aligned according to a winding sequence related to the battery cell. Deformation of the battery cell is determined based on the second image.

According to various aspects, an electronic circuit (100) is provided, wherein the electronic circuit (100) comprises: an input interface (104) and an output interface (106) for coupling to an external circuit (160); and a trained artificial neural network (102) designed to: receive one or more input signals (110, 112) which represent one or more operating parameters (120) of the external circuit (160) at a predefined point in time within a first cycle of a duty cycle signal; determine a duty cycle (122) to be used in a second cycle of the duty cycle signal after the first cycle; and provide an output signal (114, 116) to the external circuit (160) which represents the determined duty cycle.

A computer-implemented method for performing a code generation for calculating a Softmax function of a neural network includes (i) providing a displacement s and a multiplier for the quantized representation of the input tensors of the Softmax function of the neural network, (ii) creating a second lookup table to replace a nested function to calculate the EXP_ON_NEG (MUL_SAT ( )) function from a CMSIS NN library depending on the displacement and the multiplier, wherein element values of an input tensor normalized to a negative value range between 0 and a minimum value are used as arguments, wherein 0 is assigned to a maximum possible element value of the input tensor and the minimum value is assigned to the smallest possible element value of the input tensor, and (iii) implementing an access to the second lookup table in a code generated to calculate the Softmax function, so that it replaces the function call of the EXP_ON_NEG (MUL_SAT ( )) function.

A method and system for training an artificial neural network (ANN) is disclosed, specifically addressing the technical problem of training instability caused by randomly initialized weights. The invention introduces a novel, multi-phase training protocol executed by a processor. During an initial stabilization phase, the processor modifies the backpropagation of a weighted loss function by deprioritizing training on data from classes for which the ANN has a confidence level below a predetermined low threshold (e.g., less than or equal to 25 percent). This initial, focused training prevents erratic gradient updates to the weights of the network's layers, thereby forming stable and rigorous foundational neural pathways. Subsequent to the stabilization phase, a dynamic training phase commences where the network trains on the full dataset, with the training focus adjusted in real-time to reinforce the stable pathways. This protocol results in faster training convergence, a more robust final model, and improved accuracy, particularly for complex datasets, representing a significant technical improvement to the functioning of machine learning systems.

A set of unordered points associated with road markings is received. The unordered points are inputted to a trained artificial neural network. Using the artificial neural network, an output is generated by applying non-linear regression to the unordered points. Based on the output, a pose of a vehicle is determined.

A method for upscaling an image includes executing a first subnetwork of an artificial neural network (ANN), with pixels of an image patch of the image as an input layer thereto, to yield (i) a first predicted image patch as an output layer of the first subnetwork and (ii) a first vector output by a source hidden layer of the first subnetwork. The method also includes executing a second subnetwork of the ANN, with pixels of the image patch as input thereto, to yield a second predicted image patch as an output of the second subnetwork, wherein executing the second subnetwork includes concatenating the first vector with the input to a receiving hidden layer of the second subnetwork to yield a concatenated vector. The method also includes determining an upscaled image patch as one of the first predicted image patch and the second predicted image patch.

Described herein are systems, methods, and apparatuses associated with using deep learning techniques to denoise medical images, such as fluoroscopy images. A first artificial neural network (ANN) is trained to denoise input medical images according to a provided target noise level. The training of the first ANN is performed by pairing noisy input images with target denoised images that include different noise levels. During different training iterations, these target denoised images are generated using a second ANN as an intermediate output of the second ANN. As such, the first ANN can learn to perform the denoising task in an unsupervised manner without requiring noise-free training images as a gold standard.

The application relates to a long-short memory neural network control circuit and method, the circuit comprising: an input gate circuit, a forget gate circuit, a unit state configuration control circuit and a reconfigurable unit, the input gate circuit processes an input signal and outputs a current signal to the reconfigurable unit, and the reconfigurable unit converts the current signal into a voltage signal; after a first switch between the forget gate circuit and the reconfigurable unit is closed, the current signal is converted into the voltage signal and then loaded to a transistor, and the current flowing through the transistor is adjusted to discharge the capacitor of the reconfigurable unit; the unit state configuration control circuit determines reconfiguration parameters of the reconfigurable unit module according to task requirements; and the reconfigurable unit obtains current time output parameters and unit state parameters after being reconfigured based on the task requirements, the voltage signal and the reconfiguration parameters. Thus, the problems of long delay and high power consumption of a current general hardware platform running a long-short artificial neural network are solved, and the processing speed of information is effectively accelerated and the system power consumption is reduced.

The invention provides a tunnel blasting equivalent load prediction method and system based on an artificial neural network for tunnel blasting dynamic response rapid evaluation. The method comprises the following steps: firstly, establishing engineering parameters for describing a single-hole blasting working condition of a tunnel, and converting the engineering parameters into dimensionless input parameter vectors; aiming at continuity parameters and grading parameters, generating sample working conditions by adopting an orthogonal test and Latin hypercube combined sampling mode, carrying out numerical simulation on each working condition, extracting a blasting triangular wave speed-time history curve of a representative measuring point in the model, and establishing a blasting triangular wave speed-time history curve with engineering parameters of a single-hole blasting working condition as input; the database takes blasting triangular wave parameters as output; an artificial neural network is trained under the loss function, rapid prediction of triangular wave parameters is achieved, and triangular waves can serve as blasting equivalent load time history to be applied to the normal direction of the plane where representative measuring points are located and used for tunnel blasting dynamic response calculation. According to the method, the calculation cost and the modeling complexity can be remarkably reduced, and the numerical calculation stability and the engineering applicability are improved.

The present application relates to a kind of shale oil carbon dioxide fracturing production dynamic intelligent prediction method, the method is applied to the improved AHP-GRA-EWM method quantitatively calculate the influence weight of different categories factors to different stage carbon dioxide fracturing-throughput shale oil well production dynamic parameter, obtain the main control influence factor of different production stage production dynamic parameter;The main control influence factor of different production parameters of different production stages is respectively introduced into different artificial neural network system, nonlinear polynomial regression model system and different percolation mathematical model, establishes prediction model, selects the model with highest prediction accuracy to correct improvement, combination generates the production dynamic intelligent prediction model of the entire production cycle of target shale reservoir block carbon dioxide fracturing-throughput shale oil well;Based on the intelligent prediction model, the different production parameter values of different production stages of new well are calculated, and the production dynamic curve is drawn.The method designed by the present application can realize the fine and efficient prediction of carbon dioxide fracturing-throughput shale oil well production dynamics, provide basis for the development of development plan and optimization, improve the single-well production and recovery ratio of target shale reservoir block.

Numerous embodiments of analog neural memory arrays are disclosed. In one embodiment, an analog neural memory system includes an array of non-volatile memory cells arranged in rows and columns, with the columns arranged in physically adjacent column pairs. Within each adjacent pair, one column includes a cell storing a W+ value, and another column includes a cell storing a W- value. The adjacent cells in the adjacent pair store a differential weight W obtained according to the formula W = (W+) – (W-). In another embodiment, an analog neural memory system includes a first array of non-volatile memory cells storing W+ values ​​and a second array of non-volatile memory cells storing W- values.

A specialized integrated circuit for artificial neural networks is integrated with high bandwidth memory. The neural network includes a systolic array of interconnected processing elements, including upstream processing elements and downstream processing elements. Each processing element includes a pair of input / output ports for concurrent forward and backward propagation. The processing elements can be used for convolution, in which case the pair of input / output ports can support fast and efficient scanning of a kernel over activations.

The present specification provides a training method of an artificial neural network model for correcting image color and an image correction method using a trained artificial neural network model. The training method of the artificial neural network model according to the present specification uses, as training data, a pair of images before and after color correction. In the training method, the image before color correction is input into the artificial neural network model, N parameter values (N is a natural number) are obtained from the artificial neural network model, and at least one color correction program is executed according to the obtained parameter values, whereby a converted image, the color of which has been converted from the image before color correction contained in the training data, may be obtained. In addition, the artificial neural network model may be trained such that a difference value between the image after color correction contained in the training data and the converted image is reduced.

The invention relates to a device and a method for machine learning with an artificial neural network, characterized in that a layer description is provided for a computing hardware (110) for an artificial neural network, which layer description defines at least a part of a layer of the artificial neural network, wherein the layer description defines a tensor of input values of the at least a part of the layer, a tensor of weights of the at least a part of the layer and a tensor of output values of the at least a part of the layer, in particular initial addresses of these tensors, wherein the computing hardware (110) sends a message (216) for the transmission of the input values, which comprises the initial address of the tensor of the input values in a memory (102), wherein the computing hardware (110) sends a message (224) for the transmission of the weights, which comprises the initial address of the tensor of the weights in the memory (102), or wherein the computing hardware (110) sends a message (232) for the transmission of the output values, which comprises the initial address of the tensor of the output values in the memory (102).

Aspects related to how to predict the vehicle's destination include: Process a graph of the vehicle user's local settings, showing nodes corresponding to locations the user has previously visited. First, using the first graph neural network, at least one spatial graph processing is performed. This displays information about the geographical proximity of a location. The temporary graph shows information about... The locations visited by users one by one, the time between visits, and a preference graph displaying the data. Regarding the locations visited one by one and the frequency of visits, as analyzed by the neural network. The second graph combines the results of processing by the first neural network graph and the results of processing by the second graph. By constructing a second neural network graph using at least one layer of artificial neural networks and applying the results of... At least one neural network layer to predict the destination position;

The present application relates to a kind of three valve combined intelligent control method based on fuel cell membrane moisture content, it relates to fuel cell control technical field, for the unstable operation phenomenon caused by hydrogen pressure fluctuation, flow mismatch and membrane moisture content imbalance of fuel cell, specifically provide a kind of three valve combined intelligent control method based on fuel cell membrane moisture content, method includes: based on artificial neural network model, membrane moisture content is predicted and valve opening is corrected;Current hydrogen pressure fluctuation value in hydrogen discharge process of hydrogen system is obtained, and valve is again fine-tuned using PID control mode, realizes the self-adaptive regulation of system under all operating conditions and decoupling control under multiple outputs, finally achieves the purpose of hydrogen accurate supply and stack stable operation.

Aspects of the present invention are directed to computer-implemented techniques for improving the training of artificial neural networks using a reduced precision (e.g., float16) data format. Embodiments of the present invention rescale tensor values prior to performing matrix operations (such as matrix multiplication or matrix addition) to prevent overflow and underflow. To preserve accuracy throughout the performance of the matrix operations, the scale factors are defined using a novel data format to represent tensors, wherein a matrix is represented by the tuple X, where X=(a, v[.]), wherein a is a float scale factor and v[.] are scaled values stored in the float16 format. The value of any element X[i] according to this data format would be equal to a*v[i].

A sensor system is proposed in one example for analysing electroencephalography sensor signals obtained by a plurality of sensors connected to an object to be stimulated with a sequence of stimulation signals. The sensor system comprises means for: acquiring a set of sensor signals in response to stimulating a plurality of objects with a plurality of sequences of stimulation signals comprising a set of standard stimulation signals and a set of deviant stimulation signals; pre-processing the acquired set of sensor signals to obtain a set of pre-processed sensor signals; preparing an artificial neural network for data analysis by using a first set of the pre-processed sensor signals, the preparation comprising the steps of training and validating the network by using the first set of the pre-processed sensor signals; feeding a second set of the pre-processed sensor signals to the network as trained and validated; and combining output signals of the network as trained and validated in response to feeding the second set of the pre-processed sensor signals to the network to obtain a single classification result for a respective object and for a respective sequence of stimulation signals.

The invention relates to a computer-implemented method for generating and playing back an individualized, child-friendly, AI-based audio file in a vehicle (110), comprising the steps of: - capturing at least one predetermined value of a generation parameter, which is to be taken into account when generating the individualized audio file; - using the value for the generation parameter as data input for an artificial neural network (104) trained on audio files; - generating the individualized, child-friendly audio file with the trained artificial neural network (104), taking into account the captured value of the generation parameter; and - acoustically playing back the generated audio file using an audio output device (106). The invention also relates to a computer program and a vehicle (110).