52 results about "Log probability" patented technology

The local intrinsic dimensionality (LID) for a diffusion model with respect to a particular data sample is determined by using the diffusion model's diffusion process to apply noise to a data sample and evaluate how the estimated log probability of the data sample changes at different levels of noise. Particularly, the differential of change in noise to change in log probability can be used to determine the local intrinsic dimensionality. This may be determined by evaluating the log probability at several noise levels and determining a slope of the difference. In additional examples, the differential is evaluated directly at a selected noise level. The selected noise level can be optimized by calculating the estimated LID for various data samples at a variety of noise levels and selecting the LID that corresponds to a “knee” where the estimated LID sharply changes.

The invention belongs to the technical field of power system optimization scheduling, and provides a source network load storage intelligent collaborative optimization method, which comprises the following steps of: deploying sensors at four ends of a source network load storage respectively, collecting in real time by utilizing a cloud data center, enabling data of the four ends to be consistent in time sequence through a PTP protocol, constructing a topological graph according to parameters and data, and establishing a source network load storage intelligent collaborative optimization system. Selecting a model in a digital twinning environment for simulation; dividing independent agents at four ends of a source network load storage, setting observation data, an execution space and excitation feedback, forming an excitation item by economy, stability and environmental protection, interactively circulating actual data, a prediction instruction and an excitation value, recording into a sequence, inputting the sequence into a strategy network, and calculating and outputting logarithmic probability gradient to update the parameters of the strategy network; and the intelligent agent completes interactive circulation according to the strategy network, generates a local scheduling instruction, aggregates the instruction to perform weighted calculation, generates a global scheduling scheme, issues the global scheduling scheme to execution equipment, updates parameters by using an average deviation calculated by a deviation vector, resolves the global scheduling scheme and issues the global scheduling scheme to form a closed-loop mechanism.

The invention discloses a multi-modal visual language model illusion relieving method based on illusion induction optimization, and relates to the field of artificial intelligence and computer vision. According to the method, when a generative large language model generates new word marks, firstly, the probability of each word mark in a vocabulary is generated according to an input mark sequence, the contrast ratio of the current input mark sequence is calculated through a contrast model, the contrast ratio is mapped into a preset vector space, and the corrected logarithmic probability is obtained. And selecting a plurality of illusion tokens with the maximum contrast from the illusion tokens, and adjusting the probability of the illusion tokens according to a preset optimization strategy. If the illusion token belongs to the selected illusion token, amplifying the probability of the illusion token according to preset optimization intensity; otherwise, the current probability is the corrected probability. And selecting a new word mark according to the corrected probability, and circulating the process to generate a final text. According to the method, by introducing the illusion induction optimization strategy, the illusion phenomenon generated when the large visual language model generates the text is remarkably reduced.

Techniques for autocontrastive decoding of a machine learning model are provided. In one aspect, a system for machine learning includes: a multi-layer machine learning model; and an autocontrastive decoding module configured to obtain prediction probabilities from multiple, different layers of the multi-layer machine learning model as data propagates through the multi-layer machine learning model, and aggregate in a contrastive manner the prediction probabilities from the multiple, different layers of the multi-layer machine learning model to provide a final output from the multi-layer machine learning model. The multi-layer machine learning model can be a transformer-based machine learning model. The autocontrastive decoding module can be configured to redistribute a prediction probability distribution of the transformer-based machine learning model by maximizing a difference between log-probabilities of a final layer and one or more intermediate layers of the transformer-based machine learning model. A machine learning method using the present system is also provided.

The invention discloses a log anomaly detection method and device, equipment, a medium and a product, and relates to the field of log anomaly detection.The method comprises the steps that a to-be-detected log sequence is received; the to-be-detected log sequences are input into a log anomaly detection model to obtain the probability of abnormal logs in all the to-be-detected log sequences, the log anomaly detection model is obtained after an original anomaly detection model is trained, and a loss function in a weak supervision form is adopted when the original anomaly detection model is trained; determining that the corresponding to-be-detected log sequence is an abnormal log sequence when the abnormal log probability is greater than an abnormal threshold value; and outputting the abnormal log sequence. The accuracy of the abnormal log detection result can be improved.

The invention discloses a text protection method for resisting abuse analysis of a large language model, which comprises the following steps of: firstly, taking a negative logarithmic probability of an output termination condition'end token 'of the large language model as a loss function, and carrying out back propagation to obtain a gradient; constructing a replacement set for each token in the protected text by utilizing the obtained gradient; and constructing a disturbance text through the replacement set, inputting the constructed disturbance text into the large language model, and selecting the disturbance text with the optimal result as a starting point of the next round, so that the large language model outputs an end character when analyzing a protected text, and text protection is realized. According to the method, large language model-oriented protection is provided for the internet text uploaded by the user by utilizing an adversarial sample technology, and effectiveness and concealment are both considered.

The invention provides a zero-sample large model text detection method and system based on comparison prompt differential analysis, and the method comprises the steps: S1, coding a to-be-detected text into a lexical element sequence through a word segmentation device, and defining a human style prompt and a large model style prompt; s2, conditional probability calculation: respectively calculating the conditional probability of each lexical element position under two prompt contexts; s3, correction probability difference calculation: calculating a corrected logarithmic probability difference value of the lexical element at each position; s4, summarizing a difference sequence, and recording the difference values of all the lexical elements in the difference sequence in sequence; s5, parameter fitting and score generation: fitting negative exponential function parameters through a difference sequence; and S6, threshold classification judgment: comparing the fitted parameters with a preset threshold, and judging that the text is a text generated by a large language model or written by human. Based on the technical scheme of the invention, the detection can be completed without any training or fine tuning; the interpretability is high, and the word-level visualization capability is achieved; and the robustness is good and the fault tolerance is high.

The invention provides an AI text recognition method and device based on ensemble learning and advanced semantic statistical feature analysis, and the method comprises the steps: 1, respectively sending a to-be-recognized text into a Bert detector and a high-order natural language statistical feature detector for recognition, the high-order natural language statistical feature detector comprises a word logarithm probability detector, a word ranking logarithm detector, an Entropy detector and a confusion degree detector; and 2, performing election on detection results output by the Bert detector and the high-order natural language statistical feature detector by using an election module to obtain an AI text recognition result. According to the method, an integrated learning strategy is adopted, and a pre-training language model subjected to fine tuning is combined with high-order natural language statistical characteristics, so that when the model detects a large language model to generate a text, the strong expression ability of the pre-training language model can be fully utilized, and a deep rule of the text can be captured through the high-order statistical characteristics; and the detection accuracy is improved.

The invention discloses a deep reinforcement learning-driven intelligent real-time optimization method for drilling parameters, which belongs to the technical field of drilling optimization, and is technically characterized by comprising the following steps of: 1, constructing a drilling speed prediction model; step 2, modeling in a decision-making process: selecting a group of parameter combinations # imgabs0 # to obtain the optimal drilling speed, only selecting one parameter # imgabs2 # at each moment # imgabs1 # needing decision-making, and representing the MDP as a quintuple # imgabs3 #; each drilling parameter is decided one by one, and the previous decision parameter and drilling data are input; and step 3, obtaining a state # imgabs5 # from the environment at each time step # imgabs4 #, inputting the state # imgabs5 # into an Agent, obtaining an action # imgabs6 # and the logarithmic probability # imgabs7 # of the current strategy for updating the subsequent strategy of the algorithm, and achieving the optimization of the bit pressure, the torque and the rotating speed parameters.

The invention discloses a context poisoning attack defense method oriented to an RAG application, and relates to the technical field of RAG. the method comprises the following steps: inputting a target query statement, and retrieving the target query statement to obtain multiple pieces of context information; taking representative sentences in the retrieved context information, and identifying and filtering potential malicious template clusters; the big language model gives all candidate answers according to existing context information, the logarithmic probability of all contexts to different candidate answers is calculated, and after the influence of parameter knowledge of the big language model is removed from the logarithmic probability, the support degree of all contexts to different candidate answers is obtained; the whole logarithmic probability vector is used as a support degree distribution condition of the context to the candidate answers; identifying a single piece of harmful information from the support degree distribution condition of the context to the candidate answers through a logistic regression model so as to filter wrong answers; according to the attack defense method provided by the invention, centralized injection of multiple malicious texts and sparse injection of a small number of malicious texts can be defended.

The invention provides a speculative prefix multiplexing method for verifiable award reinforcement learning training, which comprises the following steps: acquiring a prompt set of a current training batch and an old response generated in a previous training iteration, the old response comprising a plurality of old lexical element sequences, each old lexical element sequence carries a first logarithmic probability and length information corresponding to each lexical element under the old strategy; calculating a second logarithmic probability corresponding to each lexical element under the current strategy according to each prompt in the prompt set and the corresponding old lexical element sequence; according to the first logarithmic probability and the second logarithmic probability, judging each lexical element according to an old lexical element sequence; if the judgment result of each lexical element is accepted, directly multiplexing the old lexical element sequence; if rejected lexical elements appear in the judgment process, the judgment is stopped to obtain reusable prefixes, and the model is called to generate suffixes corresponding to the reusable prefixes to obtain new responses, so that the expenditure in the sampling stage is reduced, and the reply efficiency is improved.

The invention discloses an equipment anomaly monitoring method based on audio signal analysis, which comprises the following steps of: S1, generating a logarithmic Mel spectrogram through time-frequency transformation, generating a gating vector by utilizing a self-adaptive gating module based on median, root mean square and variance, and after weighting fusion with an original spectrogram, splicing supplementary audio features to form a multi-channel fusion feature; and S2, the multi-channel fusion features are partitioned along a time axis and a frequency axis, parallel time domain and frequency domain branches are respectively input, a selective state space model is adopted, the time domain branches capture long-time-history steady-state features and trend anomalies, and the frequency domain branches analyze cross-band energy distribution and a harmonic structure. And S3, performing linear alignment on the two branch features, adding the two branch features element by element to obtain a fusion feature, training a model through a self-supervised ID classification task in combination with an angle interval loss function, and taking a negative logarithmic probability of a target ID category as an abnormal score during reasoning, thereby improving the robustness, generalization ability and accuracy of detection.

The present invention discloses a few-shot modulation recognition method based on a diffusion model and an attention mechanism, belonging to the field of modulation carrier systems, which solves the problem of low modulation recognition rate under few-shot conditions, and includes: inputting the original data into DDIM for progressive denoising training and adopting an inverse denoising process for progressive denoising to generate restored data; adjusting the generation trajectory based on the logarithmic probability gradient of the classifier to obtain synthetic data and a trained DDIM; inputting the original data, the restored data and the synthetic data into FATT, mapping them to the frequency domain through Fourier basis functions, explicitly capturing periodic and aperiodic features, and extracting comprehensive features; through the attention layer, splitting the comprehensive features into frequency features and non-linear features, splicing the weighted frequency features and the non-linear features and splicing them again with the output of the residual convolution block to obtain a predicted modulation type vector; visualizing and displaying the predicted modulation type results through a confusion matrix to obtain a trained FATT; the present invention improves the modulation recognition accuracy.

This invention discloses a text protection method for resisting abuse analysis by large language models. First, the negative logarithmic probability of the large language model's output termination condition, "end token," is used as a loss function, and backpropagation is performed to obtain gradients. The obtained gradients are used to construct a replacement set for each token in the protected text. Perturbed text is constructed from the replacement set and fed into the large language model. The best perturbed text is selected as the starting point for the next round of analysis, causing the large language model to output an end token when analyzing the protected text, thereby achieving text protection. This method utilizes adversarial sample technology to provide large language model-oriented protection for user-uploaded internet text, balancing effectiveness and confidentiality.

The invention relates to the technical field of label noise of table data, a distance measurement-based method and a neural network-based method are currently common label noise screening methods, and the methods are difficult to distinguish wrongly labeled samples and difficult samples with fuzzy categories near a decision boundary; according to the table data noise recognition and correction method based on Stein scoring, the difference of the attribution degree between sample and data feature distribution and label export distribution is analyzed, and the logarithmic probability density gradient of a sample is estimated through a diffusion model; moving the sample along a gradient represented by a scoring function by using a gradient method until the sample is converged to a centroid, forming a moving track moving from an original position of a sample space to a high-density centroid of data of a domain, and calculating the data feature distribution by comparing the differences of the directions and lengths of the moving track of the sample in the data feature distribution and the distribution exported by the label. Samples of potential tag errors in the tabular data are identified and tags thereof are corrected.

The invention discloses a natural language interaction intelligent customer service system, and relates to the technical field of intelligent interaction. The system comprises a standardization module used for collecting original data from a database of an intelligent customer service system and performing text standardization processing on the original data to obtain standard data, the original data including text data, voice data and image data; the word vector module is used for converting the standard data into word vectors through a word embedding method; the method is technically characterized in that the relations of sentences can be classified more accurately according to sentence meanings, probability distribution is output by using a Softmax function, a final classification result is given, the probability of each category is provided, the prediction result of the model is more interpretable and reliable, and the prediction efficiency of the model is improved. The combination of the cross entropy loss function and Softmax provides a very clear and efficient optimization direction for the model, the logarithmic probability of a correct category can be directly maximized, the gradient is clear, and the convergence speed is high.

The application discloses a natural language interaction intelligent customer service system and relates to the technical field of intelligent interaction; the system comprises a specification module, which is used for collecting original data from a database of an intelligent customer service system and performing text normalization processing on the original data to obtain standard data, wherein the original data comprises text data, voice data and image data; and a word vector module, which is used for converting the standard data into word vectors through a word embedding method; the technical key points are as follows: the relationship of a sentence can be classified more accurately according to the meaning of the sentence, a probability distribution is output by using a Softmax function, the final classification result is given, and the probability of each category is provided, so that the prediction result of the model is more interpretable and reliable, a cross-entropy loss function and the Softmax are combined to provide a very clear and efficient optimization direction for the model, the logarithmic probability of a correct category can be directly maximized, the gradient is clear, and the convergence speed is fast.

The application discloses a kind of information retrieval method and system based on large language model topological semantic reordering, including Euclidean rough screening, topological gravity calculation, power law sorting and result output step.This method uses vector embedding model to carry out rough screening to query and candidate document, then calls large language model to obtain the logarithmic probability value of each candidate document, and is re-sequenced according to power law decay law.The application also discloses a recall boundary determination method and a model alignment degree detection method based on semantic decay index.In polysemy disambiguation task, retrieval accuracy can be improved from 69% to 99%, with the advantages of high retrieval accuracy and self-adaptive model alignment state.

The application relates to the technical field of natural language processing, in particular to an AI generated text detection method and system based on a deep learning model. The method comprises the following steps: performing character normalization and sentence segmentation on a to-be-detected text, segmenting the text by a sliding window to generate a segmented sequence; constructing deep semantic discrimination features, logarithmic probability difference trace features, style structure features and entity consistency features; inputting a gate fusion discrimination network, outputting a generated probability and confidence degree after probability calibration, and outputting a segmented contribution degree; performing threshold adaptive determination according to a risk level, a text length and the confidence degree, outputting a review state and generating a traceable record when the confidence degree is insufficient; and triggering incremental updating based on drift monitoring and a feedback sample pool. The application improves cross-domain robustness and interpretability, reduces false positives and false negatives, and supports long-term stable operation.

This invention discloses a method for generating pseudo-labeled data for a large-scale visual model based on the next-point prediction paradigm. The method includes: using a self-consistent multi-path heuristic sampling module to drive the large-scale visual model to perform multiple decoding operations, generating multiple inference paths and obtaining the log probabilities of discrete tokens; using a micro-fine-grained uncertainty calibration module to calculate the joint uncertainty score of the category and bounding box tokens in the inference paths, eliminating high-entropy illusions and obtaining surviving candidate boxes; using a macro-multi-path consistency fusion module to spatially cluster the candidate boxes, and performing consensus voting on the clustered boxes; and performing weighted box fusion based on the confidence score of the inverse mapping to generate the final pseudo-labeled data. This invention can effectively reduce the illusion phenomenon of large-scale visual models in complex scenes, generate high-quality pseudo-labeled data, and provide reliable long-tail data support for downstream student models without sacrificing real-time performance.

The application provides a wind speed abnormal fluctuation detection method based on SSA and an LSTM network, and relates to the technical field of mine ventilation; sensor monitoring data is decomposed and reconstructed to remove data noise generated by turbulent pulsation, then parameters of the LSTM are optimized, the pretreated data is predicted by using the optimized LSTM model, and reconstructed wind speed is obtained; finally, the abnormal score of the monitoring wind speed and the reconstructed wind speed is calculated by using a logarithmic probability density function, the threshold value is set by calculating the abnormal score of the training set data sample, and the abnormality of the monitoring wind speed is detected; the application provides technical support for identifying abnormal fluctuation in monitoring data, and has great significance for improving the reliability and applicability of sensor monitoring data.

The invention relates to the technical field of natural language processing, in particular to an AI generated character detection method and system based on a deep learning model. The method comprises the following steps: performing character standardization, sentence segmentation and sliding window segmentation on a to-be-detected text to generate a segmentation sequence; constructing a deep semantic discrimination feature, a logarithmic probability difference generation trace feature, a style structure feature and an entity consistency feature; inputting a gating fusion discrimination network, outputting a generation probability and a confidence coefficient which are subjected to probability calibration, and outputting a segmentation contribution degree; executing threshold adaptive judgment according to the risk level, the text length and the confidence coefficient, and when the confidence coefficient is insufficient, outputting a to-be-rechecked state and generating a traceable record; and triggering incremental updating based on drift monitoring and the feedback sample pool. According to the method, the cross-domain robustness and interpretability are improved, false alarms and missing alarms are reduced, and long-term stable operation is supported.

The application discloses a kind of inhibiting inference generation control methods of teaching intelligent agent followability error, it is related to artificial intelligence and educational technology field, in this method: input text sequence is input to parameter frozen model and obtains original probability distribution, each word element gradient influence quantity is calculated, and high influence position set is screened;High influence position word element embedding vector is contracted interpolation to average embedding vector, and reference embedding tensor is generated;Original and reference embedding tensor is input to parameter frozen model respectively, and original and reference logarithm probability is calculated;Original and reference aggregated probability distribution is calculated in dimensionality reduction probability space, and error follow risk coefficient is determined;According to logarithm probability residual, each word element correction direction and penalty weight are determined;In discrete rectification intensity candidate set, the intensity that makes the minimum comprehensive loss function is selected, and the original logarithm probability is corrected and decoded output.The application reduces the follow probability of teaching intelligent agent to error preset, improves the fact accuracy and logical rigor of teaching reply.

The application discloses a metasurface holographic optimization method based on a limited super atom response library selection, belongs to the metasurface holographic field, and has the steps that a three-dimensional logarithmic probability matrix P (an element represents the logarithmic probability of a structure in a response library selected by a super atom at a grid position) is randomly initialized; a Gumbel-Softmax sampling strategy is used to generate a discrete selection matrix, an index corresponding complex amplitude response is constructed, and a complex amplitude distribution matrix of the whole metasurface is constructed; then, a differentiable wave front propagation algorithm is used to calculate a target complex amplitude distribution; multi-polarization multiplexing is supported, image superposition is realized through time multiplexing; a structural similarity constraint is introduced; a total loss value gradient relative to the probability matrix is calculated; the probability matrix is updated through a stochastic gradient descent; and finally, iteration is performed until convergence, so that a selection matrix and a directly manufacturable metasurface structure layout are obtained. The application can realize high-precision and high-efficiency polarization multiplexing metasurface holographic design.

The invention discloses an information processing method and device based on a large language model, equipment and a medium, which are used for more accurately identifying information needing to be shielded and avoiding information leakage caused by malicious confusion. The method comprises the steps of obtaining a to-be-processed original text; determining a lexical token sequence corresponding to the original text through a large language model, and determining the logarithmic probability of each token in the token sequence; on the basis of the logarithmic probability of each token in the token sequence and a preset logarithmic probability threshold value used for judging the disclosure degree of information, the tokens to be shielded in the token sequence are marked; based on the to-be-shielded token, determining to-be-shielded information in the original text; and according to the to-be-shielded information, determining an output text corresponding to the original text, the output text being a text obtained by shielding the to-be-shielded information in the original text.

The invention discloses a multi-pedestrian non-autoregressive trajectory prediction method in a dynamic environment. The method comprises the following steps: extracting coding features of historical trajectories of multiple pedestrians by using a Transform structure of an encoder module; processing coding features by using a contrast distribution potential coding generator CDLCG; potential coding and historical trajectory coding features are input through an MRG decoder, a plurality of candidate future trajectories are generated through decoding, and a better prediction trajectory is screened out according to the logarithmic probability of potential coding. According to the embodiment of the invention, a CDLCG (Comparative Distribution Potential Code Generator) is designed, in actual deployment, a guide model learns how to deduce proper Gaussian distribution parameters from limited historical observation information, potential codes are sampled from the distribution, and the rationality of the potential codes is evaluated by calculating the logarithmic probability of the potential codes, so that the real-time performance of the system is improved. And finally, screening out an optimal prediction result from the plurality of candidate prediction trajectories. The trajectory prediction precision is remarkably improved, and excellent reliability and precision are shown.

This invention relates to the field of generative artificial intelligence and discloses a semantic processing method for 5G message parsing, comprising: inputting a 5G message text sequence into a generative language model; calculating the information entropy of the intermediate network layer; terminating the calculation and extracting the hidden layer state tensor when the information entropy is lower than a threshold; determining the set of legitimate intentions based on the 5G protocol state and transition topology; modulating the initial logarithmic probability tensor using a mask tensor to mask semantic outputs outside the protocol range; calculating the normalized probability distribution sequence and determining the target parsing intention. This invention maps protocol rules to the boundary of the logarithmic probability space, blocks signaling routing failures caused by probability divergence, resolves the conflict between semantic generation and protocol rigor, and ensures the determinism of service parsing under telecommunications conditions.

This invention discloses a deep reinforcement learning-driven intelligent real-time optimization method for drilling parameters, belonging to the field of drilling optimization technology. Its key technical points include the following steps: Step 1: Construction of a drilling rate prediction model; Step 2: Modeling of the decision-making process, where the decision-making process involves selecting a set of parameters to achieve the optimal drilling rate. At each decision-making moment, only one parameter is selected, and the MDP is represented as a quintuple; each drilling parameter is decided sequentially, with the input being the previous decision parameter and drilling data; Step 3: Overall framework structure, at each time step, the logarithmic probability of the action and the current policy is obtained from the state input from the environment in the Agent and used for subsequent policy updates in the algorithm. This method has the advantage of optimizing drilling pressure, torque, and rotational speed parameters.

The invention discloses a multi-modal pre-training data filtering method based on a multi-modal large model, and the method comprises the following steps: 1, obtaining multi-modal pre-training data of a training model, and classifying the multi-modal pre-training data; 2, adjusting a training reference model on external high-quality multi-modal data by using a standard visual instruction; 3, calculating the reference loss of all text marks according to the logarithmic probability obtained by the reference model RM; and step 4, sorting and screening the multi-modal pre-training data obtained in the step 1 through the reference loss calculated in the step 3 to obtain final high-quality multi-modal pre-training data. According to the multi-modal pre-training data filtering method based on the multi-modal large model, the data set can be carefully checked so as to find out the data sample with the highest representativeness and training value.

This invention proposes a speculative prefix reuse method for verifiable reward reinforcement learning training, comprising: obtaining the cue set of the current training batch and the old responses generated in the previous training iteration, wherein the old responses include multiple old word sequences, each old word sequence carrying the first logarithmic probability and length information of each word under the old policy; calculating the second logarithmic probability of each word under the current policy based on each cue in the cue set and the corresponding old word sequence; judging each word according to the first logarithmic probability and the second logarithmic probability in the order of the old word sequence; if the judgment result for each word is acceptance, the old word sequence is directly reused; if a rejection word appears during the judgment process, the judgment is stopped to obtain a reusable prefix, and the model is called to generate the suffix corresponding to each reusable prefix to obtain a new response, thereby reducing the overhead of the generation sampling stage and improving the response efficiency.