14results about How to "Valid encoding" patented technology

The invention discloses a medical field large model pre-training and fine tuning method and device based on a knowledge graph, and the method comprises the steps: training a reward model through employing collected medical related text data and a quality marking result, and screening out high-quality data from the newly obtained medical related text data through employing the trained reward model; coding the medical knowledge graph into a text format processed by a large language model, and alternately inputting the high-quality data and the coded medical knowledge graph into the large language model in different batches for model pre-training; and designing an instruction fine tuning data format comprising an instruction, a knowledge graph reasoning path and a generated answer, inputting a training sample constructed in the instruction fine tuning data format into the pre-trained large language model for model fine tuning, and outputting a final answer and reasoning path according to the instruction newly input by the user by utilizing the fine-tuned large language model. According to the method, the professionality of the large language model in the medical field can be effectively improved, and AI application in high-risk fields such as medical treatment is promoted.

The invention discloses a model construction method, device and equipment for generating positive and negative sample nodes, and relates to the technical field of graph data processing. And constructing a neural network comprising positive and negative sample node generation branches, a self-information completion branch and a global information completion branch, and collecting a data set to train the neural network to obtain a model for learning graph data representation. According to the method, the node feature matrix generated by the encoder is constrained by setting the additional global information completion branch and the node self-information completion branch, so that the node feature matrix which enables the loss of the global information of the graph data and the node self-information to be minimum is obtained; and the representation of the positive and negative sample nodes with the minimum information loss can be obtained based on the node feature matrix.

The invention provides a double-view fair recommendation method. The fairness and accuracy of recommendation are improved by dissociating sensitive information. The method starts from two perspectives: one perspective is based on a real sensitive attribute of a user, and the other perspective is an anti-fact hypothesis perspective based on the change of the sensitive attribute of the user. The method comprises the following steps: firstly, constructing four representation vectors which are respectively used for performance orientation, true sensitivity, anti-fact sensitivity and fairness; these representations are generated and optimized by different algorithms to ensure that sensitive information is removed and user preferences are preserved. In particular, the fair representation enables dissociation of information by minimizing mutual information and cosine similarity differences between the fair representation and the sensitive representation. And meanwhile, mutual information between the fair representation and the performance oriented representation is maximized, so that the preference of the user is effectively captured. Finally, according to the method, the interaction score of the user and the project is calculated through fair representation, a recommendation result is generated, and the fairness of recommendation is evaluated with the assistance of fairness measurement.

The invention discloses an information source channel joint coding method for joint digital communication and analog communication. The method comprises the following steps: step 1, in a digital branch, inputting a picture, and firstly performing down-sampling and image coding to obtain a code stream; 2, in the analog branch, subtracting the predicted value from the input picture to obtain a residual error, and performing learning-based transformation on the residual error to obtain a characteristic of the residual error; step 3, bandwidth allocation is realized; 4, sending the characteristics subjected to bandwidth allocation to a receiving end through a wireless channel; 5, at a receiving end, the signals received by the digital branches are restored into predicted values through demodulation, channel decoding, image decoding and up-sampling in sequence; and step 6, performing demodulation and learning-based inverse transformation on the signals of the analog branches, reconstructing a residual error, and adding the residual error and a predicted value to obtain a reconstructed picture. According to the invention, source and channel coding is carried out by using a traditional separated coding method through a digital branch, and transmission is carried out by using a residual error through an analog branch, so that better performance is realized.

The invention discloses a power waveform zero crossing point detection method and system under wide-frequency-domain harmonic waves. The weighted sum of a data unit vector, a quantized embedded vector of environment temperature, a quantized embedded vector of environment relative humidity and a position code is used as an input sequence; according to a set mask processing proportion and a noise processing proportion, replacing a unit segment containing a power waveform zero crossing point in the input sequence with different types of noise, inputting the noise into the trained Transform model, and taking a starting point position of the unit segment containing the power waveform zero crossing point as a label value of the input sequence; the model output sequence comprises the prediction data of the unit segment containing the power waveform zero crossing point in the input sequence, and the label value of the output sequence is the distance prediction value from the starting point position of the unit segment containing the power waveform zero crossing point to the power waveform zero crossing point; the sum of the label value of the input sequence and the label value of the output sequence is the predicted value of the power waveform zero crossing point, the data labeling amount is reduced, and the generalization ability is improved.

The invention relates to a database high-performance storage system for an electric power intelligent terminal, which is characterized by comprising a database function unit, a data cleaning unit based on a lightweight rule engine and an asymmetric data compression unit based on bit back coding, the database function unit comprises a data access engine and an OpenAPI interface; the data cleaning unit based on the lightweight rule engine comprises a data deduplication module, a rule configuration module, a rule loading module, a regularization verification module and a mode matching module adopting a mode matching algorithm mainly based on a regular expression. The asymmetric data compression unit based on bit-back coding introduces an asymmetric data compression technology based on bit-back coding to perform data compression on the basis of Huffman coding and arithmetic coding. According to the invention, the storage and access performance of the data is improved in an intelligent terminal environment with limited resources.

This invention proposes a brain-inspired navigation method based on DSI decoupling representation and composite potential field path optimization. The method includes: acquiring a state sequence and calculating a successor representation matrix; obtaining a DSI representation from the successor representation matrix; generating a navigation path sequence based on the DSI representation; constructing a total potential function from the navigation path sequence; generating an optimal path sequence based on the total potential function and simultaneously introducing obstacle information; using the optimal state sequence extracted from the optimal path sequence as a supervision signal to update the DSI representation using gradient descent, and storing the obtained corrected DSI representation in a calibration experience buffer; repeatedly executing online navigation, path optimization, and calibration steps using the corrected DSI representation in the calibration experience buffer, and obtaining the optimal brain-inspired navigation sequence after a preset number of iterations. This invention applies composite potential energy optimization and obstacle constraints to the initial path, resulting in a geometrically more compact and safer path.

The invention relates to the technical field of laser communication microwave coding, in particular to an interstellar laser communication microwave coding method, which comprises the following steps: acquiring a microwave signal to be transmitted, converting the microwave signal into a digital signal, arranging the acquired digital signal according to a magic square construction rule, generating a magic square code, loading the magic square code onto a laser beam, and transmitting the magic square code to the laser beam. And transmitting the laser beam loaded with the microwave signal to a destination. A mathematical principle of a magic square is utilized, microwave signals to be transmitted are converted into digits, then the digits are arranged in the magic square according to a certain rule, the digits are loaded to light waves through laser beams, transmission of the microwave signals is achieved, the signals are effectively coded and transmitted, the anti-interference capacity of the signals is improved, the microwave signals are loaded to the laser beams, and the transmission efficiency is improved. According to the invention, high-speed and high-bandwidth transmission can be realized, different coding modes can be realized through different magic square construction methods, and the flexibility is very high.

The invention discloses a small sample radar signal intelligent identification method and electronic equipment, and the method comprises the steps: carrying out the preprocessing of a radar signal through CWD, and obtaining real sample data; inputting the real sample data into a DCGAN network, and expanding the real sample data to obtain corresponding generated sample data; and inputting the real sample data and the generated sample data into a VIT network to form a small sample radar signal intelligent identification network model. According to the method, the CWD is adopted to preprocess the data set of the radar signals, so that the method has relatively good energy aggregation while keeping a certain time-frequency resolution, and subsequent further processing is facilitated. A DCGAN network is combined with a VIT network to establish a small sample radar signal intelligent identification network model, a large amount of generated sample data with distribution closer to that of real sample data is generated to train the identification model, the fitting effect is more stable, and the identification accuracy of small sample radar signals is improved.

The present application relates to the field of computer vision, in particular to the subject matter of video processing and video encoding, and more particularly to a method, decoder, encoder and computer readable medium for video encoding using a machine learning model. According to a first aspect, a method of processing image and / or video data by a decoder is provided. The method includes continuously decoding a plurality of segments of an image or video. For one or more of the plurality of segments, the decoding comprises: decoding a corresponding encoded residual segment to obtain a decoded residual segment; generating predicted segments from respective reference segments using a machine learning model, wherein the respective reference segments are selected from one or more reference segments stored in a set of reference segments; generating a reconstructed fragment based on the prediction fragment and the decoded residual fragment; storing the reconstructed fragment in the reference fragment set; and updating parameters of the machine learning model based on the corresponding reference segments and the reconstructed segments.

The invention relates to an eye movement recognition method based on Transform global semantic coding, and relates to the field of eye movement recognition. Comprising the steps of collecting corresponding eye movement data according to different task requirements, and constructing an eye movement recognition data set; the eye movement data set comprises an image and eye movement data corresponding to the watched image; according to the fixation point position recorded in the eye movement data, performing fixation point mask processing on the related image, and taking the image subjected to mask processing and the image not subjected to mask processing as the input of a Transform model; the method comprises the following steps: carrying out pre-training on a Transform model; extracting deep features from all fixation point positions of each image by using a pre-trained Transform model, and predicting an image-based category probability based on the deep features; and integrating the category probabilities based on the images to obtain a comprehensive classification result of the subjects. The problems that an existing eye movement recognition method is insufficient in deep semantic information mining and cannot capture the global dependency relationship are effectively solved.

The application discloses a printed circuit board defect detection method and system based on deep learning. The application uses an industrial camera to collect a printed circuit board image and pre-processes the same; the pre-processed printed circuit board image is input into a pre-trained improved YOLOv12 printed circuit board defect detection model to obtain a detection result. The NAMHSA module designed by the application can effectively capture global context and local dependency, thereby effectively capturing fine-grained features of printed circuit board defects. The C3K2-SSMSCA module designed by the application can not only effectively reduce the loss of local details and spatial position information, effectively enhance useful channels and suppress irrelevant channels, but also emphasize meaningful features and suppress meaningless features in the spatial dimension. The application can effectively reduce the interference of non-defect targets and background noise, effectively and adaptively extract printed circuit board defect features, and thereby improve the detection precision of printed circuit board defects.

This invention provides a method and system for enhancing the cross-linguistic capabilities of a large language model, applicable to the fields of natural language processing and large language model technology. The method includes: constructing a probe corpus; based on the probe corpus, obtaining neuron sets corresponding to the source language and neuron sets corresponding to the target language through a large language model; obtaining overlapping neuron sets and activation matrices of the overlapping neuron sets for the source and target language sets; calculating the dependency score between each candidate bridging language and the activation matrix of the overlapping neuron sets based on the activation matrices of the overlapping neuron sets, and determining the optimal bridging language from the candidate bridging languages; and providing a Cross-lingual In Context Learning (X-ICL) construction strategy for cross-linguistic tasks based on the optimal bridging language. Thus, a significant performance improvement in X-ICL can be achieved with zero samples and without fine-tuning on cross-linguistic tasks and low-resource languages ​​covering multiple language families.