Patents
Literature
Patsnap Eureka AI that helps you search prior art, draft patents, and assess FTO risks, powered by patent and scientific literature data.

67 results about "Loss function" patented technology

In mathematical optimization and decision theory, a loss function or cost function is a function that maps an event or values of one or more variables onto a real number intuitively representing some "cost" associated with the event. An optimization problem seeks to minimize a loss function. An objective function is either a loss function or its negative (in specific domains, variously called a reward function, a profit function, a utility function, a fitness function, etc.), in which case it is to be maximized.

Method for improving reasoning ability of large language model based on Monte Carlo tree search

The invention relates to a method for improving the reasoning ability of a large language model based on Monte Carlo tree search. The method comprises the following steps that S1, a mathematical question set and corresponding standard answers are preset; s2, taking each step for solving a mathematical problem as a tree node, and generating a Monte Carlo tree in the whole problem solving process; s3, generating training data based on Monte Carlo tree search; s4, scoring the generated training data; and S5, constructing a loss function according to scores obtained by scoring, optimizing the loss function and updating parameters of the large language model. The reasoning ability of the large language model can be improved.
Owner:GIANT MOBILE TECH CO LTD

Training method, reasoning method and related device of stream matching generative model

The embodiment of the invention provides a training method, a reasoning method and a related device for a stream matching generation model, which are used for improving the accuracy of an action sequence predicted by the trained stream matching generation model. The method provided by the embodiment of the invention comprises the following steps: acquiring noise, a moment t and an environment characteristic of a first action, wherein the environment characteristic of the first action at least comprises an observation value of the first action; inputting the noise, the moment t and the environment characteristics into an initialized flow matching generation model to obtain an output predicted velocity field vector of the conditional probability path at the moment t + 1; calculating the loss between the predicted velocity field vector and the real velocity field vector by using a preset loss function, wherein the preset loss function comprises at least one of a first loss function and a second loss function, and a third loss function; and training the initialized stream matching generation model by using a loss and back propagation algorithm until the stream matching generation model converges to obtain a trained stream matching generation model.
Owner:BEIJING HUMANOID ROBOTICS INNOVATION CENTER CO LTD

Loss function optimization method and system for improving tiny target detection performance

The invention relates to the technical field of target detection, and provides a loss function optimization method and system for improving small target detection performance. In the method, an elliptical bounding box of a target detection model is modeled through two-dimensional Gaussian distribution, a Wasserstein distance between a target box and a prediction box is corrected based on a dynamic gradient gain factor according to the model size of the elliptical bounding box, and normalization in an index form is applied to the obtained dynamic Wasserstein distance, so that the target detection model is obtained. And obtaining a normalized Wasserstein distance as a loss function, and updating model parameters of the target detection model. Therefore, the scale invariance is introduced to enhance the efficiency, the dynamic joint optimization process is promoted, and the attention on a tiny target is enhanced by adjusting the gradient; and the normalized Wasserstein distance is used as a bounding box regression loss function, so that the target detection model has higher convergence speed and detection precision.
Owner:HEBEI FARADAY TECH CO LTD

Lithium carbonate price prediction method and equipment based on large language model, and medium

The invention specifically discloses a lithium carbonate price prediction method and device based on a large language model, and a medium, and relates to the technical field of financial prediction and intelligent decision support. The method comprises the following steps: firstly, collecting and processing numerical type and text type data, constructing a causal graph network by the numerical type data, and extracting factual abstracts by the text type data to construct a news event library; then designing a three-layer progressive prompt template, and driving the large language model to generate a prediction result with explanation; then, implementing a dual-verification self-optimization reflection mechanism, and triggering an reflection generation improvement strategy when misjudgment is carried out or an error exceeds a threshold value; using a KTO algorithm to construct asymmetric loss function fine tuning model parameters; and finally, deploying the optimized model to a production system. According to the method, the dynamic adaptability and robustness of the model are improved through a'prediction-reflection-optimization 'closed-loop optimization architecture and a KTO algorithm driven by a foreground theory, prediction precision and interpretability are considered, and multi-dimensional decision support is provided for lithium carbonate price prediction.
Owner:HEFEI UNIV OF TECH

Dynamic concept cognitive anomaly detection method and system under unbalanced condition

The invention relates to a dynamic concept cognitive anomaly detection method and system under an unbalanced condition, and the method comprises the steps: collecting the related feature data of a sample set in a scene facing a specific extreme event, and dividing an initial data set and a test set; constructing an initial learning model; performing dynamic event type prediction based on the test set, and classifying the event as an extreme event or a non-extreme event; constructing an active three-way concept learning module to form a positive concept space and a negative concept space based on a three-way decision idea and prediction probability values of extreme events and non-extreme events; according to the positive concept space and the negative concept space, introducing a concept center domain strategy for dynamic updating; and adopting a loss function minimization strategy to obtain an optimal learning model. The effectiveness of dynamic detection is improved by utilizing an active three-way concept learning method, a concept center field and a concept dynamic optimization strategy, and the capability of identifying extreme events under the extremely unbalanced scene condition is realized.
Owner:CENT SOUTH UNIV

Method for solving an optimization problem in adiabatic quantum computing

The invention pertains to a computer-implemented method for solving an optimization problem using an analog quantum computer. The method comprises the steps of: (a) training a Physics-informed Neural Network (PINN) using a plurality of training samples to address a Counter-Diabatic driving problem in a quantum system, wherein training the PINN comprises: - defining a temporal domain as the input of the PINN, - constructing the PINN on the basis of the temporal domain, - obtaining outputs from the PINN and derivating them, - updating the temporal domain with derivated outputs; and (b) obtaining a set of counterdiabatic terms as the solution for the counter-diabatic driving problem from the PINN; (c) differentiating at least one partial differential equation characterizing a time-dependent behavior of the system; (d) minimizing a loss function; and (e) solving, using the set of counterdiabatic terms, the optimization problem.
Owner:KIPU QUANTUM GMBH

Radiation therapy treatment planning

A method and apparatus for generating a radiation treatment plan for a volume comprising:receiving a first treatment plan upon which to base a second treatment plan, the first treatment plan indicative of a dose distribution; receiving at least one dose-distribution-derived function configured to provide a value based on at least part of the dose distribution; receiving a target for the respective value of each dose-distribution-derived function comprising receiving a probability distribution; determining an optimization problem, wherein the or each objective function is a function of the dose-distribution-derived function, the respective probability distribution and a respective loss function; performing the optimization process based on said optimization problem; anddetermining said second treatment plan.
Owner:RAYSEARCH LAB

Decoupling adversarial sample generation method

The invention discloses a decoupling adversarial sample generation method. The method comprises the following steps: decoupling an adversarial sample by taking a condition of decoupling the adversarial sample and obtaining a sum of a clean sample and adversarial disturbance as a decoupling principle; constructing a multi-objective optimization function, wherein the multi-objective optimization function comprises a plurality of constraint conditions; constructing a plurality of loss functions to enable the constructed plurality of loss functions to correspond to a plurality of constraint conditions, and training and updating the generative model by using the weighted sum of the plurality of loss functions as a final loss function to obtain a generative model meeting a decoupling principle confrontation sample; evaluating the obtained generative model meeting the decoupling principle confrontation sample by using an evaluation index until the evaluation result meets the decoupling principle and successfully attacks the target model at the same time; and generating a confrontation sample which can simultaneously meet the decoupling principle and aggressiveness by using the trained generative model. According to the invention, the problem that the adversarial sample cannot be decoupled, especially the problem that the adversarial sample cannot be decoupled at the model side, can be solved.
Owner:NAT INNOVATION INST OF DEFENSE TECH PLA ACAD OF MILITARY SCI

Mathematical problem data enhancement method and system based on semantic retrieval

The invention discloses a mathematical problem data enhancement method and system based on semantic retrieval. The method comprises the following steps: loading a mathematical problem data set, calculating features of each mathematical problem by using a preset neural network model encoder, and outputting an original feature set; for one feature in the set, the other feature closest to the feature in the whole feature set is screened out by calculating semantic similarity, and the screened feature is obtained; performing mixed enhancement on the original features and the screened features to obtain enhanced features; inputting the original features and the enhanced features into a preset neural network model decoder to obtain a structured solution; and repeating the above processes, and iteratively training parameters of the model by using a loss function with a gradient stopping strategy to obtain an optimized mathematical problem and a solution thereof. According to the method, the adaptability of the model to different question types can be effectively improved, language barriers can be overcome, and the method is also suitable for mathematical problems of Chinese and English.
Owner:GUANGDONG UNIV OF TECH

Two-stage label noise learning classification method and system

PendingCN120597066AData setOptimality model
The invention discloses a two-stage label noise learning classification method and system, and the method comprises the following steps: constructing a relative similarity filter through the comprehensive measurement of relative similarity and Euclidean distance, inputting an original data set into the relative similarity filter, screening out noise samples through the calculation of the relative similarity of the samples, and carrying out the classification of the noise samples. Outputting the filtered data set; inputting the filtered data set into a robust sorting support vector machine model based on a non-convex slope loss function and a local label weighting strategy for training, and converting the non-convex loss function into a convex optimization problem through a convex-concave process for solving to obtain an optimal model parameter; and updating a model threshold by using the optimal model parameter, and classifying a to-be-classified sample by using the updated model. According to the method, data-level screening and model-level noise suppression are both considered, the problem of unstable performance under a single processing path in a traditional method is solved, and the method is particularly suitable for stable learning tasks under high-dimensional and high-noise scenes.
Owner:GUANGDONG UNIV OF TECH

Large model-based scheme generation and optimization model training method and device

The invention provides a training method and device of a scheme generation and optimization model based on a large model. The training method of the scheme generation and optimization model based on the large model comprises the steps of obtaining a sample experiment scheme and a sample demand thereof; training by taking the sample experiment scheme and the sample demand as training samples and taking the first autoregression loss as a loss function to obtain a scheme generation model; processing and training the second language by taking the historical optimization trajectory and the target scheme as training samples and taking the improved opinion loss as a loss function to obtain a feedback model; taking a sample requirement, a historical optimization track and a sample improvement opinion output by the feedback model in a historical training process as a training sample, and taking a second autoregression loss as a loss function for training to obtain an optimization model; obtaining a scheme generation and optimization model based on the scheme generation model, the feedback model and the optimization model; according to the method provided by the invention, the automatic design and optimization of the scheme for the special equipment or system are realized, and the generation efficiency and accuracy of the experimental scheme are improved.
Owner:INST OF AUTOMATION CHINESE ACAD OF SCI +1

Method and apparatus for training of large model based solution generation and optimization model

The application provides a large model-based scheme generation and optimization model training method and device, the large model-based scheme generation and optimization model training method comprises the following steps: obtaining a sample experiment scheme and sample requirements thereof; taking the sample experiment scheme and the sample requirements as training samples, and training a scheme generation model by taking a first autoregressive loss as a loss function; taking a historical optimization trajectory and a target scheme as training samples, and training a feedback model by taking an improved opinion loss as a loss function; taking the sample requirements, the historical optimization trajectory, and sample improved opinions output by the feedback model in a historical training process as training samples, and training an optimization model by taking a second autoregressive loss as a loss function; and obtaining a scheme generation and optimization model based on the scheme generation model, the feedback model, and the optimization model; the method realizes automatic design and optimization of a scheme for a special device or system, and improves experiment scheme generation efficiency and accuracy.
Owner:INST OF AUTOMATION CHINESE ACAD OF SCI +1

META-SCHOLAR EVOLUTION STRATEGY BLACKBOX OPTIMIZATION CLASSIFIERS

ActiveDE102021204943B4Measurement devicesMachine learningMultivariate normal distributionAlgorithm
A computational method (400) for training a meta-learned evolutionary strategy black-box optimization classifier to learn an actuator control command for actuating a computer-controlled machine, the method comprising: Receiving (402) one or more training functions F and one or more initial metalearning parameters Θ of the metalearning evolutionary strategy blackbox optimization classifier; Sampling (404) a generation of λ samples z1,...,z λ a sampled objective function f ∈ F of the one or more training functions F and an initial mean m (0) of the sampled objective function f ∈ F, where the generation of λ samples z1,...,z λ with a multivariate normal distribution N(m (0) = 0, e (0) C (0) = 1) with initial mean m (0) = 0 and an initial covariance matrix C (0) = 1 in samples x1,...,xλ is shifted and scaled using the equation: xi = m ( t ) + σ ( t ) C ( t ) 1 2 zi , where σ (t) a step size of a number T steps in t = 1,...,T; for the number of T steps in t = 1,...,T, calculating (406) a set of T means m (1) ,...,m (T) by running the meta-scholarly evolutionary strategy black-box optimization classifier on the sampled objective function f ∈ F using the initial mean m (0) ; Calculate (408) a loss function L(f(m 0 ), ..., f(m T )) of the set of T means m (1) ,...,m (T) ; and Updating (410) the one or more initial metalearning parameters Θ of the meta-learned evolutionary strategy black-box optimization classifier in response to a characteristic of the loss function L to obtain an updated meta-learned evolutionary strategy black-box optimization classifier that includes a weighted combination of the generation of λ samples with larger weights for samples with smaller values ​​for the objective function f ∈ F; Sending input signals received from a sensor into the updated meta-scholarly evolutionary strategy black-box optimization classifier to obtain output signals designed to characterize a classification of the input signals; and Sending an actuator control command to an actuator (14) of the computer-controlled machine in response to the output signals; and Actuating the computer-controlled machine in response to the actuator control command.
Owner:ROBERT BOSCH GMBH

Information processing device and information processing method

This improves the accuracy and speed of solving large-scale combinatorial optimization problems by dividing them into subproblems. [Solution] The information processing device 100 for processing combinatorial optimization problems includes: a graph creation unit 112 that creates one or more subgraphs from a main graph; a mathematical optimization unit 115 that solves the combinatorial optimization problem for each subgraph using a mathematical optimization solver; a machine learning unit 117 that trains each subGNN so that the output of the subGNN corresponding to each subgraph is close to the solution of the mathematical optimization solver; a feature vector assignment unit 118 that assigns the feature vectors at each vertex of the subGNN obtained as a result of training to the corresponding vertices of the main GNN as input to the feature vectors of the main GNN corresponding to the graph data of the main graph; and a solution output unit 119 that outputs the solution obtained as a result of training the main GNN by setting a loss function so that the machine learning unit 117 solves the combinatorial optimization problem for the main graph.
Owner:HITACHI LTD +1

Isocenter optimization selection method for radiotherapy multi-target plan

The invention relates to an isocenter optimization selection method for a radiotherapy target plan, and the method comprises the steps: obtaining a CT image, and carrying out the preprocessing of the CT image, and obtaining an initial Mask matrix; based on the initial Mask matrix, calculating to obtain a rotated Mask matrix; establishing and searching an optimal plan center point loss function; and based on a random optimization algorithm, solving to obtain an optimal plan center point. Compared with the prior art, the method has the advantages that potential geometric errors caused by rotation uncertainty in the actual treatment process can be reduced, the optimal isocenter point is determined, and the target coverage rate and the coverage precision are improved.
Owner:ZHONGSHAN HOSPITAL FUDAN UNIV

Method for solving an optimization problem

The invention pertains to a computer-implemented method for solving an optimization problem using an analog quantum computer. The method comprises the steps of: (a) training a Physics-Informed Neural Network (PINN) to obtain a complete decomposition of a Gauge potential using a plurality of training samples to address a Counter-Diabatic driving problem in a quantum system, wherein training the PINN comprises: - defining a temporal domain as the input of the PINN, - constructing the PINN on the basis of the temporal domain, - obtaining outputs from the PINN and derivating them, - updating the temporal domain with derivated outputs; and (b) obtaining a set of counterdiabatic terms as the solution for the counter-diabatic driving problem from the PINN; (c) differentiating at least one partial differential equation characterizing a time-dependent behavior of the system; (d) minimizing a loss function; and (e) solving, using the set of counterdiabatic terms, the optimization problem.
Owner:KIPU QUANTUM GMBH

A method and device for controlling reasoning state based on prior knowledge of large language model

The present invention relates to the field of data processing technology, and in particular to a method and device for controlling the reasoning state based on prior knowledge of a large language model. The method is applied to the problem of autonomous driving discretized navigation, container stacking, cargo stacking, and text numerical calculation under natural language text. The method includes: obtaining a sample data set of decision reasoning of an unmanned system; using the sample data set and a loss function composed of an ATE causal effect estimation algorithm and cross entropy to train the initial large language model; obtaining a long-range action decision problem, constructing a causal probability tree, determining the optimization problem of pruning and expansion directions, and then dividing the long-range action decision problem into multiple short-range action decision problems; using the trained large language decision reasoning model to solve multiple short-range action decision problems, and then completing the action. The present invention can solve the problem of causal hallucination of the large language model in long-range reasoning tasks and improve the accuracy of reasoning.
Owner:UNIV OF SCI & TECH BEIJING

Method and device for determining steady-state parameters

The invention discloses a steady-state parameter determination method and device, which are applied to the field of data processing, and the method comprises the steps: determining iteration variable information and dependent variable information under a target scene configured by a user, the iteration variable information comprises the name of an iteration variable and the initial value of the iteration variable, and the dependent variable information comprises the name of the iteration variable and the initial value of the iteration variable; the dependent variable information comprises a name of a dependent variable and a target value of the dependent variable, constructing a relation equation of the target scene according to the iterative variable information and the dependent variable information, and if a loss function of the relation equation reaches a minimum value, determining a current initial value of the iterative variable in the relation equation as a steady-state parameter, and if the loss function of the relation equation does not reach the minimum value, according to a preset initial value selection rule, re-determining the target initial value to iterate the relation equation until the loss function reaches the minimum value or the number of iterations reaches an iteration number threshold, and determining the corresponding target initial value as a steady-state parameter. Therefore, the steady-state parameter can be automatically determined, and the steady-state problem in the system is solved.
Owner:BEIJING GLOBAL CROWN JINYANG TECH DEV CO LTD

An accurate learning method for Gaussian variational inference of UAV trajectories

The Gaussian variational inference accurate learning method of the UAV trajectory disclosed by the present invention comprises the following steps: step S1: constructing a mathematical model of the UAV according to the aerodynamic principle and kinematic principle of the UAV in actual flight; step S2: based on Gaussian variational inference, linking the true Bayesian posterior and the approximate posterior through KL divergence, converting the state estimation problem into the approximate problem of the posterior, and then converting the calculation of the KL divergence into the optimization problem of the loss function by defining the loss function; step S3: using Newton's method to update the loss function and using the natural gradient descent method to iteratively optimize the mean and the inverse covariance matrix, so that the loss function V(q) is reduced to the minimum to achieve the purpose of effective approximate posterior. Batch state estimation of nonlinear systems is realized. Compared with the maximum a posteriori estimation method, the state estimation under the condition of Gaussian variational inference has higher accuracy and can more accurately estimate the flight trajectory of the UAV.
Owner:NANJING UNIV OF AERONAUTICS & ASTRONAUTICS

Question situation understanding-based mathematical application question intelligent solving method and system

The invention relates to an intelligent mathematical application question solving method and system based on question situation understanding, and the method comprises the following steps: A, carrying out the preprocessing of collected mathematical application question data, and building a training set containing the known conditions, questions and answers of mathematical application questions; b, constructing a deep learning model for intelligent solving of mathematical application questions, determining a loss function, and training the deep learning model on the training set by minimizing the loss function; the deep learning model simulates a dynamic relationship between entities in a question by constructing an interaction graph, and gradually updates an entity state by adopting an edge elimination method so as to improve the ability of understanding the question and answering an application question; and C, inputting known conditions and questions of mathematical application questions into the trained deep learning model, and outputting answers to the questions. According to the method and the system, known condition texts and question descriptions of input mathematical application questions can be understood, and corresponding answers and equation expressions can be given.
Owner:FUZHOU UNIV

Optimization Method of Cosine Optimal Loss Function Based on Global Information

The present invention proposes an optimization method for a cosine optimal loss function based on global information, including: S1. Combining the advantages of existing loss functions with some important new attributes and applying L2 weight normalization; S2. Clearly following the two objectives of minimizing intra-class variation and maximizing inter-class variation, relying on a new algorithm to learn the cosine similarity between class centers and class margins, and respectively proposing two lightweight versions of the cosine optimal loss function; S3. Integrating the above two lightweight versions to create a standard version of the cosine optimal loss function. The present invention mainly aims at the problems that existing loss functions do not apply weight and feature normalization or do not clearly follow the objectives of minimizing intra-class variation and maximizing inter-class variation. By using global information as the feedback for face recognition, a cosine optimal loss function based on global information is proposed. Compared with existing loss functions, this loss function is more effective and achieves more advanced performance.
Owner:HANGZHOU SHANHAIGUAN TECHNOLOGY CO LTD

Model loss optimization method and system for event detection data long tail problem

The invention discloses a model loss optimization method and system for an event detection data long tail problem. The method comprises the steps that a basic model is obtained, and the basic model is a BERT-BiLSTM-CRF model; constructing an optimization loss function of the basic model; taking the optimization loss function as a training parameter of the basic model, training the basic model, and obtaining a trained model; and performing event detection on the input data by using the trained model. According to the scheme provided by the invention, the optimization loss function is adopted for model training, so that the performance of the model on a long-tail data set can be remarkably improved, and the effectiveness of the model in processing a data imbalance problem is displayed.
Owner:NAT UNIV OF DEFENSE TECH

Retriever model training method, birt-hogg-dube syndrome identification method and system based on retrieval enhancement generation

This application discloses a retrieval model training method, a retrieval-enhanced generation-based Bert-Hogg-Dubbs syndrome (BHD) identification method and system, relating to the field of rare disease identification. The training method includes acquiring positive and negative sample pairs; inputting the positive and negative sample pairs into an initial retrieval model; calculating the loss function of the initial retrieval model; and dynamically adjusting the angle margin of the loss function in real time according to a metric variance adaptive mechanism. The initial retrieval model is then optimized based on the loss function calculation results. This application forcibly expands the angle interval between BHD and non-BHD by using the angle margin of the loss function, and uses a metric variance adaptive mechanism to dynamically adjust the angle margin based on the statistical variance of the cosine similarity among all positive sample pairs in the current training batch. This solves the problem of weak image differences and blurred category decision boundaries in DCLDs caused by the highly similar imaging features of various rare diseases, thus improving the recognition accuracy of large models for query information.
Owner:UNIV OF SCI & TECH OF CHINA

A multi-modal model hallucination elimination method based on sharpness perception and target-oriented disturbance

PendingCN122634101AData setAlgorithm
The application discloses a multi-modal model object hallucination elimination method based on sharpness perception and target-oriented disturbance. In view of the hallucination suppression failure and index sharp rebound problem of the existing machine forgetting technology after parameter fine-tuning, the application firstly automatically constructs negative, positive and general sentence data sets based on alignment scores; then a target-oriented min-max optimization model is established, and the sensitive direction of the model on the loss function surface is locked by simulating the parameter disturbance in the worst case; finally, a sharpness perception parameter updating mechanism is used to force the model to converge to the flat extreme value region of the loss space. The application solves the problem that the hallucination rate of the model significantly increases when the model faces parameter disturbance or incremental training interference, realizes deep robust elimination of hallucination, and effectively maintains the general utility of the model on the regular task.
Owner:NANJING UNIV OF AERONAUTICS & ASTRONAUTICS

Student mathematical ability evaluation method based on time series regression and incremental learning model

The application discloses a student mathematical ability evaluation method based on a time series regression and an incremental learning model, relates to the technical field of mathematical ability evaluation, and comprises the following steps: S1, establishing a mathematical question bank and performing ability dimension division, and recording the examination proportion of the ability dimension; S2, setting a dynamic data screening mechanism of a sliding window; S3, collecting student answering sequence data; S4, constructing a mathematical ability evaluation regression model based on the dynamic data screening mechanism of the sliding window and a time decay weight; S5, constructing a loss function of the mathematical ability evaluation regression model, and searching for the mathematical ability evaluation regression model parameters when the loss function reaches a minimum value by adopting a Bayesian optimization algorithm; and S6, performing real-time dynamic evaluation and diagnostic evaluation on the mathematical ability of students.The application has the characteristics of small calculation amount, low complexity, real-time incremental updating, strong interpretability and personalized analysis of learning paths, and can be applied to mobile terminal deployment under resource-limited conditions.
Owner:SOUTHWEST JIAOTONG UNIV

Data model modeling method based on mathematical data analysis operation processing

A data model modeling method based on mathematical data analysis operation processing comprises the steps that (a) dynamic feature screening is conducted on original data, feature importance is calculated based on information entropy and a covariance matrix, redundant features are removed by setting a dynamic threshold value, and the feature weight distribution proportion is adjusted in real time according to data distribution changes; (b) performing tensor decomposition dimension reduction on the screened high-dimensional data: decomposing three-dimensional or more tensors into a product of a low-rank core tensor and a factor matrix by adopting a CP decomposition or Tucker decomposition method, and extracting potential correlation features; (c) constructing a multi-objective optimization model: combining a least square loss function with an L1 / L2 regularization item, and adjusting model parameters in real time through a dynamic feedback mechanism which comprises an error back propagation algorithm and a self-adaptive learning rate strategy; and (d) adjusting and optimizing the hyper-parameters based on a Bayesian optimization algorithm, inspecting and evaluating the distribution consistency of the model by using Kolmogorov-Smirnov, and dynamically adjusting a training strategy according to an inspection result.
Owner:NANJING INST OF MECHATRONIC TECH

ZKML proof efficiency optimization method based on random sampling

The invention discloses a ZKML proving efficiency optimization method based on random sampling. Comprising the following steps: S1, model declaration and key generation; s2, key distribution; s3, the training party carries out training; s4, establishing a credit rating evaluation function D; s5, constructing a total loss function Loss (N); s6, generating a proof; and S7, verifying the correctness of the result by the user. Target data are randomly sampled, so that the generation and verification efficiency of the proof is improved; the loss function determines a sampling proportion; the method is carried out in an untrusted distributed scene, each substructure of a segmented model is independently converted and compiled, and a pair of secret keys are respectively generated, so that the substructures are independently verified; a credit rating evaluation function and a time forgetting factor are introduced to give a lower weight to a longer task, and a higher weight is given to a task closer to the current task; and the basic weight can be randomly adjusted according to specific requirements for the reservation or non-reservation behaviors.
Owner:ZHENGZHOU UNIV

Power marketing language model reasoning optimization method and system based on chain thinking

The invention discloses an electricity marketing language model reasoning optimization method and system based on chain thinking. The optimization method provided by the invention comprises the following steps: encoding initial information of power marketing multi-modal input by using a head module to generate an initial state; iteratively calculating and gradually optimizing the initial state through a recursion module to obtain a potential state, participating in a generation process of soft target distribution by the recursion module in each iteration, and supervising, training and reasoning the potential state in combination with a teacher model; meanwhile, corresponding soft target distribution is generated through a teacher model to serve as a reference standard, selection of the teacher model in each iteration stage and the generation quality of the soft target distribution are optimized by adopting a suitable distillation loss function, and adaptation of a dynamic supervision signal is achieved; and decoding the final potential state by using a tail module to generate prediction information. According to the invention, the reasoning capability and generalization performance of the power marketing language model in a complex power marketing business scene are enhanced.
Owner:STATE GRID ZHEJIANG ELECTRIC POWER CO MARKETING SERVICE CENT

A knowledge distillation method based on inference step disassembly and differentiated supervision

The present application relates to a kind of knowledge distillation method based on inference step deconstruction and differentiating supervision, belong to artificial intelligence field.The structured thinking chain generated by acquiring teacher model is separated according to line, each inference step is deconstructed into independent unit;Each step is type labeled, according to the five-level classification system of basic calculation, basic fact, operation execution, logical reasoning, strategy planning, differentiating weight is distributed, and weight superposition is used to composite step;Weighted loss function is constructed, step weight is introduced into cross-entropy loss, and student model is distilled training.The present application focuses on key inference link in the learning process by step-level type perception and differentiating supervision, overcomes the problem of detail loss and global understanding damage caused by flat sequence supervision in traditional distillation.Experimental results show that, on mathematical reasoning task, compared with equal-weight step distillation method, accuracy is improved by 4.6 percentage points, and efficient migration of large model reasoning ability to lightweight model is realized.
Owner:ZHEJIANG UNIV