Data annotation method and system of collaborative computing architecture based on quantum computing

Abstract

The invention discloses a data annotation method and system of a collaborative computing architecture based on quantum computing, and belongs to the field of data annotation. The method comprises the steps that S1, multi-modal data are input and preprocessed; s2, extracting features of each mode after preprocessing; s3, coding the features of each mode into a quantum state, and carrying out mode fusion; s4, performing label reasoning on the quantum state after modal fusion, and performing label constraint optimization by using a quantum approximate optimization algorithm; s5, based on a quantum Bayesian network or an approximate causal graph generation method, generating explanation according to a modal contribution causal path, and deducing marginal contribution of each modal to final label prediction by using a joint probability measurement result; and S6, outputting a labeling result. According to the method, a quantum-classical cooperative computing architecture is designed, the efficiency and accuracy of multi-modal data labeling are remarkably improved, the interpretability, the distributed processing capacity and the high-dimensional feature modeling capacity of the system are enhanced, and a brand new solution thought is provided for development of the multi-modal labeling technology.

Description

Technical Field

[0001] The present invention relates to the interdisciplinary technical field of multimodal data processing, quantum machine learning, and data annotation algorithms, and in particular to a data annotation method and system based on a collaborative computing architecture based on quantum computing, which is suitable for automatic annotation tasks that require processing multi-source heterogeneous data such as images, text, audio, and sensors. Background Art

[0002] With the widespread application of artificial intelligence and big data technologies, multimodal information processing has become an essential component of information systems. Multimodal data refers to information modalities derived from different sensor channels or data types, such as images, text, voice, infrared, radar, depth maps, and sensor signals. In the real world, this type of data often coexists and complements each other. Therefore, the unified modeling and annotation of multimodal data has become a fundamental task for enabling key applications such as cross-modal retrieval, intelligent decision-making, and human-computer interaction.

[0003] The multimodal data labeling task is essentially a comprehensive process involving inter-modal feature extraction, information fusion, semantic alignment, and label prediction. Its goal is to automatically assign unified semantic labels to datasets containing content from different modalities, or to annotate key regions or attributes of each modality at the temporal, spatial, or semantic levels. For example, in autonomous driving systems, it is necessary to simultaneously identify vehicles in images, traffic rules described in text, and obstacle information in lidar signals; in medical systems, doctors need to integrate MRI images, clinical text reports, and patient voice statements to label and diagnose diseases.

[0004] However, current mainstream multimodal annotation technologies generally rely on deep learning architectures, particularly classic models such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), and transformer networks. Although these technologies have achieved certain results, they still face numerous bottlenecks in data annotation efficiency, interpretability, system scalability, and high-dimensional data modeling.

[0005] First, existing technologies often struggle to achieve efficient and accurate feature fusion when processing multimodal data. Large semantic gaps exist between modalities such as images, text, and audio. Existing methods often use feature concatenation or attention mechanisms for fusion. However, this approach has limited fusion effectiveness when faced with complex relationships between modalities and insufficient samples, often relying on large amounts of manually annotated data for optimal performance. As data size and modality types increase, the labor and time costs of annotation tasks rise exponentially, making low annotation efficiency an increasingly prominent issue.

[0006] Secondly, most current multimodal annotation systems are "black-box" deep models that lack clear causal structure modeling. During reasoning, the models cannot track the specific impact of each modal information on the final annotation result, nor can they explain why a certain modality is given a higher weight at a certain moment. This lack of explainability makes it unsuitable for practical applications in scenarios with extremely high security and reliability requirements, such as medical diagnosis, intelligent monitoring, and financial risk control.

[0007] Thirdly, when labeling tasks need to run in a distributed environment or need to process extremely complex semantic relationships and modal interactions, the processing capabilities of traditional architectures are significantly limited. Specifically, when executing large-scale labeling tasks, traditional collaborative computing architectures have low task scheduling efficiency between computing nodes and significant system communication delays, making it difficult to achieve optimal allocation of tasks and resources. Especially in distributed or edge environments, due to the strong coupling between tasks and resource preemption issues, the scheduler faces bottlenecks, resulting in a decrease in the throughput of the entire system and seriously affecting the real-time processing capabilities of tasks.

[0008] Furthermore, as the dimensionality of multimodal features increases, existing computing frameworks face severe resource bottlenecks when processing high-dimensional features. Despite existing GPU-based acceleration methods, these still struggle to cope with the storage and parallel computing pressures associated with large-scale, high-dimensional tensor operations. This results in slow inference speeds and the inability of models to quickly adapt to new samples, limiting their scalability and adaptability in practical deployments.

[0009] In recent years, with the development of quantum computing, its potential in machine learning and optimization has gradually attracted attention. Through mechanisms such as quantum superposition, entanglement, and interference, quantum computing can demonstrate efficiency far exceeding that of classical computing in certain specific computing tasks. It is particularly suitable for tasks such as high-dimensional space modeling, solving complex optimization problems, probabilistic reasoning, and data generation. However, there is no complete, multi-module integrated multimodal annotation algorithm that embeds these quantum technologies into the actual annotation process. Currently, most published work focuses on single-modal tasks (such as image classification and text classification), and most experiments are still based on simulators, lacking an engineering system for large-scale multimodal data processing.

[0010] In summary, existing methods are either limited to classical deep learning models, unable to overcome bottlenecks in computational efficiency and interpretability, or are prototype research in the field of quantum machine learning, lacking a deployable solution for practical multimodal annotation tasks. Therefore, research on how to incorporate quantum computing technology into multimodal data annotation tasks and build a new, efficient, interpretable, and scalable annotation architecture has become a research hotspot in cutting-edge interdisciplinary fields. Summary of the Invention

[0011] The purpose of the present application is to overcome the technical problems existing in the prior art, provide a data labeling method and system based on a quantum computing collaborative computing architecture, which not only integrates quantum feature encoding, quantum optimization, quantum generation and causal reasoning mechanism, but also designs a hybrid computing scheduling architecture that can be actually deployed, and realizes the complete process of data preprocessing, feature fusion, label prediction and causal explanation.

[0012] The purpose of the present application is achieved by the following technical solutions: In a first aspect, a data labeling method based on a quantum computing collaborative computing architecture is provided, comprising the following steps: S1. inputting multi-modal data and performing preprocessing; S2. extracting features of each modality after preprocessing; S3. encoding the features of each modality into quantum states and performing modality fusion; S4. performing label reasoning on the quantum states after modality fusion, and using a quantum approximate optimization algorithm to perform label constraint optimization; S5. generating an explanation based on a quantum Bayesian network or an approximate causal graph generation method according to the causal path of modality contribution, and using the joint probability measurement result to deduce the marginal contribution of each modality to the final label prediction; S6. outputting the labeling result.

[0013] In some embodiments, before step S3 is performed, the following steps are further included: According to the task complexity, dynamically allocate quantum or classical computing resources.

[0014] In some embodiments, the dynamic allocation of quantum or classical computing resources includes: Performing complexity evaluation on the task; When quantum computing is not available or queued, automatically fallback to classical computing; Collecting measurement results of quantum computing in real time and feeding back to the classical optimizer for circuit update; Parallelly manage the modality input stream through a multi-thread scheduling strategy to realize edge input and edge execution.

[0015] In some embodiments, the dynamic allocation of quantum or classical computing resources further includes: Adjust the scheduling priority according to the quality of each modality.

[0016] In some embodiments, the inputting multi-modal data and performing preprocessing includes: Respectively perform intra-modality standardization processing, wherein the image is cropped to a uniform resolution, the text is segmented and vectorized, the audio is converted to a mel spectrum graph, and the time series data is processed into a fixed time window time series segment.

[0017] In some embodiments, the features of each modality after pre-processing are extracted, including: The image extracts an image feature vector through a pre-trained CNN; The text extracts a semantic embedding through a Transformer structure; The audio extracts a spectral feature through a convolution-gating network; The time series data extracts a state representation through a GRU network.

[0018] In some embodiments, the features of each modality are encoded into quantum states through amplitude encoding in step S3; the modality fusion includes: A variational quantum circuit is constructed using a parameterized quantum gate to realize the interaction between modalities; the variational quantum circuit includes a parameter layer, an interference layer, and an output measurement layer, the parameter layer is composed of rotation gates controlled by parameters, the interference layer introduces control gates to form inter-modal interference, and the output measurement layer measures a specific quantum bit to obtain a classification probability.

[0019] In some embodiments, a prediction circuit based on a quantum support vector machine or a quantum variational classifier is used in step S4 to perform label inference on the quantum state after modality fusion.

[0020] In some embodiments, the label constraint optimization using a quantum approximate optimization algorithm includes: A cost function is defined: wherein, represents the cost of selecting label ; represents the conflict penalty of label ; A corresponding Hamiltonian is constructed, and a quantum approximate optimization algorithm circuit is used to perform: The initial state is a uniform superposition; and and are used alternately; the lowest cost label combination is found by measurement.

[0021] In a second aspect, a data labeling system based on a quantum computing collaborative computing architecture is provided, including: A multi-modal data pre-processing module for inputting multi-modal data and pre-processing; A multi-modal feature extraction module for extracting features of each modality after pre-processing; A quantum feature mapping and modality fusion module for encoding the features of each modality into quantum states and performing modality fusion; A label generation and optimization module for performing label inference on the quantum state after modality fusion, and using a quantum approximate optimization algorithm for label constraint optimization; A causal reasoning module, which generates explanations based on modal contribution causal paths using quantum Bayesian networks or approximate causal graph generation methods, and uses joint probability measurements to derive the marginal contribution of each modality to the final label prediction; The annotation output module is used to output the annotation results.

[0022] It should be further explained that the technical features corresponding to the above embodiments can be combined or replaced with each other to form a new technical solution if there is no conflict.

[0023] Compared with the prior art, the present invention has the following beneficial effects: 1. Existing technologies for modal fusion typically rely on deep neural network structures (such as Transformer or attention mechanisms) to model inter-modal correlations. This approach suffers from high computational cost, slow training convergence, and susceptibility to overfitting in high-dimensional feature spaces. This present invention replaces traditional fusion structures with quantum feature coding and quantum variational circuits, leveraging quantum superposition and nonlinear evolution to naturally model high-dimensional correlations between modalities. This reduces the number of model parameters and computational complexity, significantly improving fusion efficiency.

[0024] 2. Existing models are generally "black box models" that make it difficult to explain the modal contribution paths during label prediction, especially in medical and judicial scenarios. This method, by constructing a quantum Bayesian network and a causal graph inference mechanism, can output a clear causal chain and modal influence weights, achieving the explainability and traceability of multimodal predictions and significantly enhancing the credibility of the system.

[0025] 3. Task scheduling and resource management in traditional multimodal systems often rely on static strategies, making it difficult to dynamically adapt to the load conditions of different computing resources. This invention introduces a collaborative scheduling module that dynamically allocates tasks to quantum or classical processing modules based on task complexity and quantum resource status, improving resource utilization efficiency and supporting high-throughput real-time operation in heterogeneous environments.

[0026] 4. For complex scenarios involving label conflicts or label combination optimization problems, traditional approaches typically employ heuristic rules or classical optimization algorithms, but these methods are prone to falling into local optimality. This paper employs the Quantum Approximate Optimization Algorithm (QAOA) to model and solve label constraint problems. This method can more efficiently obtain near-optimal solutions in an exponential search space, improving the global optimality of the labeling system.

[0027] In summary, the present invention not only outperforms the existing technology in algorithm performance, but also achieves substantial innovation in architectural design, interpretability, resource scheduling efficiency, and engineering feasibility, and has obvious technological progress and application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a flow chart of a data annotation method based on a collaborative computing architecture of quantum computing according to an embodiment of the present invention; Figure 2 This is a collaborative scheduling flow chart shown in an embodiment of the present invention; Figure 3 This is a circuit diagram of quantum feature mapping and modal fusion shown in an embodiment of the present invention. DETAILED DESCRIPTION

[0029] The technical solutions of the present invention are described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. The components of the embodiments of the present application generally described and shown in the drawings herein can be arranged and designed in various different configurations. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0030] It should be noted that the defects existing in the solutions in the above-mentioned prior art are the results obtained by the inventor after practice and careful research. Therefore, the discovery process of the above-mentioned problems and the solutions proposed in the embodiments of this application below for the above-mentioned problems should be the contributions made by the inventor to this application in the process of invention and creation, and should not be understood as technical contents known to technical personnel in this field.

[0031] In response to the technical problems pointed out in the background technology, the embodiments provided by the present invention are as follows: In an exemplary embodiment, a data annotation system based on a collaborative computing architecture of quantum computing is provided, comprising: Multimodal data preprocessing module, used to input multimodal data and perform preprocessing; Multimodal feature extraction module, used to extract features of each modality after preprocessing; The quantum feature mapping and modal fusion module is used to encode the features of each modality into quantum states and perform modal fusion; The label generation and optimization module is used to perform label inference on the quantum state after modal fusion and use the quantum approximate optimization algorithm to perform label constraint optimization; A causal reasoning module, which generates explanations based on modal contribution causal paths using quantum Bayesian networks or approximate causal graph generation methods, and uses joint probability measurements to derive the marginal contribution of each modality to the final label prediction; A quantum-classical collaborative scheduling module, which dynamically allocates quantum or classical computing resources based on task complexity; The annotation output module is used to output the annotation results.

[0032] Specifically, the system is based on a heterogeneous computing architecture and applies the advantages of quantum computing to computational bottlenecks such as feature fusion, optimized decision-making, and causal interpretation, forming a task complementarity and collaborative processing mechanism with classical modules. Each module is interconnected through interfaces and control flows to complete the data-driven interpretable label prediction process. The specific implementation method is as follows: 1. Multimodal representation learning: In a multimodal environment, let the input be a set of data samples of different modalities. , where each is the feature vector of the corresponding modality (image, text, speech, etc.). The goal is to learn a cross-modal mapping function , generating a joint representation , and then generate the function through the label Mapped to label space.

[0033] Traditional methods use classical neural networks (such as the attention-based multimodal Transformer) to approximate the functions f and g. However, this approach is unstable when modal differences are large, dimensionality is high, and data is scarce, making the fusion results difficult to interpret. Therefore, the present invention replaces the high-dimensional nonlinear mapping module with a quantum circuit architecture, naturally implementing cross-modal inner products, interference, and optimization, fundamentally overcoming the limitations of existing methods.

[0034] 2. Quantum feature mapping and modal fusion module: The goal of this module is to encode classical feature vectors into quantum states for subsequent fusion and calculation. Input: Classical modal eigenvectors Encoding method: Using Amplitude Encoding: Normalize the eigenvector to unit norm so that it satisfies Then map it to quantum state: This quantum state can be The representation can be performed on qubits, thereby compressing the high-dimensional representation and supporting fast inner product calculations between subsequent quantum states (which can be used for modeling similarities between modalities).

[0035] Encoding circuit: The circuit architecture is implemented using a series of rotation gates (RY, RX, RZ) and control gates (CNOT). Each feature dimension is loaded into the qubit via a corresponding rotation gate angle. For image modalities, block encoding can be used to improve parallelism; for text, sentence vector representations after word segmentation embedding can be used.

[0036] Classic preprocessing stages for audio encoding: The audio data encoding begins with preprocessing and feature extraction in a classical computational pipeline. The raw audio signal is converted into a standardized digital representation through a preprocessing module, which includes steps such as resampling, denoising, and normalization. Specifically, the system uniformly samples the input audio signal to 16kHz to ensure consistency across audio sources. The audio signal is then converted into a Mel-spectrogram representation, which simulates the human auditory perception and better captures the semantic information of the audio.

[0037] The mel-spectrogram conversion process uses a 128-mel filter bank with a time window length of 25 milliseconds and a step size of 10 milliseconds. This parameter configuration achieves a good balance between temporal and frequency resolution. The converted mel-spectrogram has a dimensional structure of 128×T, where T represents the number of time frames and varies with the length of the audio. To process variable-length audio sequences, the system uses a convolutional-gated network architecture for feature extraction. The convolutional layers extract local patterns and texture features from the spectrum, while the gated recurrent unit (GRU) models temporal dependencies. Finally, a global pooling operation compresses the variable-length sequence into a fixed-dimensional feature vector.

[0038] The advantage of this classic preprocessing method is that it fully utilizes mature deep learning technology to effectively extract the semantic features of the audio, while also providing a standardized input format for subsequent quantum encoding. Through this stage of processing, the raw audio is converted into 256-dimensional dense feature vectors. These vectors contain the main semantic information of the audio and lay the foundation for quantum state encoding.

[0039] Quantum amplitude encoding stage: The audio feature vectors obtained through classical preprocessing then enter the quantum encoding stage, which is one of the core innovations of this patent. Quantum amplitude encoding maps the n-dimensional classical feature vector into a quantum state representation of log2(n) qubits, achieving exponential dimensionality compression. Specifically, for a 256-dimensional audio feature vector, only 8 qubits are required to complete the encoding, achieving a compression ratio far exceeding that of any classical compression algorithm.

[0040] The mathematical principle of amplitude encoding is based on the normalization constraint of the quantum state. First, the classical eigenvector is L2-normalized to ensure that the constraint ∑|xᵢ|² = 1 is satisfied. The normalized vector is encoded as the quantum state |ψ> = ∑ᵢ xᵢ|i>, where each component xᵢ corresponds to the amplitude of the quantum state in the ground state |i>. This encoding method not only achieves dimensionality reduction but, more importantly, provides a natural foundation for parallel computing in subsequent quantum computing operations.

[0041] The construction of quantum coding circuit adopts parameterized quantum gate sequence, which mainly includes rotation gates (RY, RX, RZ) and control gates (CNOT). Each characteristic dimension is loaded into the quantum state through the corresponding angle rotation gate, and the angle parameter θ is To improve coding efficiency, the system uses a block coding strategy, which divides the high-dimensional feature vector into multiple sub-blocks. Each sub-block is independently encoded into a different group of quantum bits. This method not only maintains the local structure of the feature but also facilitates parallel quantum operations.

[0042] Time series preprocessing and window encoding: The encoding of time series data (such as sensor signals and physiological signals) faces the challenges of variable sequence length and complex temporal dependencies. This patent uses sliding window technology to segment continuous time series signals into fixed-length time segments, with each segment serving as an independent encoding unit. The window length is determined based on the signal's sampling rate and semantic characteristics, typically set to encompass 1-2 seconds of signal. This ensures sufficient time series information is captured while limiting computational complexity.

[0043] The sliding window overlap ratio is set to 50%. This design ensures the continuity of time series information and avoids the impact of window boundary effects on feature extraction. The signal within each time window is first normalized using a zero-mean unit variance normalization method to eliminate the influence of signal amplitude differences between different sensors or time periods. The normalized time series segments are passed through a gated recurrent unit (GRU) network for feature extraction. The GRU's hidden state dimension is set to 256, which can effectively capture long-term time series dependencies.

[0044] The final hidden state of the GRU network serves as the feature representation of the time series segment. This method compresses variable-length time series signals into a fixed-dimensional vector representation. Compared to simple statistical features (such as mean, variance) or Fourier transforms, GRU encoding can preserve the dynamic characteristics and nonlinear patterns of the time series, providing richer information for subsequent quantum fusion. For multi-channel time series data (such as multi-sensor fusion scenarios), the system adopts a channel-independent encoding and re-fusion strategy. The time series signal of each channel is independently encoded by the GRU and then spliced ​​or weighted fused at the feature level.

[0045] Temporal quantum state encoding and evolution Quantum encoding of time series data requires not only quantum state representation of features but also the preservation of the dynamic evolution of time series information. This patent designs a time series modeling method based on quantum state evolution, encoding time series features into initial quantum states and then simulating the dynamic changes of the time series through parameterized quantum circuits. The innovation of this method lies in leveraging the inherent time reversibility and information retention of unitary quantum state evolution to better model the inherent dynamics of time series data.

[0046] In its implementation, the time series feature vector is first converted into a quantum state representation through amplitude encoding. This encoding process is similar to audio features, but with the addition of encoding of time series tag information. The time series tag is encoded using additional quantum bits to identify the relative positional relationships between different time steps. The encoded quantum state is then evolved using a variational quantum circuit. The circuit parameters are learned through training, enabling the evolved quantum state to predict future states of the time series or extract key features of the time series.

[0047] The advantage of this quantum encoding method for time series lies in its ability to simultaneously process both local features and global patterns in time series. Quantum superposition naturally supports the parallel representation of multiple time states, while quantum entanglement can capture the complex correlations between different time steps. Through quantum measurement operations, the system can extract features of time series at different time scales, enabling multi-resolution time series analysis.

[0048] Modal Fusion: The quantum state after encoding multiple modes As input, a trainable variational quantum circuit (VQC) is constructed for modality fusion and label reasoning, such as Figure 3 shown.

[0049] 3. Label generation and optimization module, which performs label inference through variational quantum circuits. The variational quantum circuits include: Parameter layer: composed of parameters Controlled revolving door composition, e.g. , construct training space; Interference layer: CNOT is introduced to form intermodal interference; Output measurement layer: measures a specific bit (such as the first bit) to obtain the classification probability.

[0050] Output layer: The probability of the measurement result is the confidence of the label: Update circuit parameters using a classic optimizer (Adam) by minimizing the cross entropy loss , complete the training.

[0051] Tag optimization: In multi-label, multi-objective or constraint scenarios, there are label conflicts or redundancy issues. This paper introduces quantum optimization methods to solve such combinatorial optimization tasks.

[0052] Problem Modeling: Define the cost function: in: (For example, if “sleeping” and “driving” appear at the same time, the penalty will be high).

[0053] Construct the corresponding Hamiltonian , using the Quantum Approximate Optimization Algorithm (QAOA) circuit to perform: The initial state is uniform superposition; use and Alternating action; Find the label combination with the lowest cost through measurement.

[0054] 1. Initial state uniform superposition: In the label optimization process of multimodal data annotation, the present invention first constructs a uniform superposition state containing all possible label combinations as the initial state of quantum computing. For a labeling task containing n candidate labels, the system maps each label to a quantum bit, where the quantum state Indicates that the label is not selected, and the quantum state |1> indicates that the label is selected. The mathematical expression of the uniform superposition state is , where each ground state Corresponding to a specific n-bit binary label selection scheme, the coefficient Ensure that the quantum state meets the normalization conditions.

[0055] The uniform superposition state is achieved by applying the Hadamard gate to each qubit in turn. The specific operation is: , where H represents the Hadamard gate, Represents a tensor product operation. A single Hadamard gate converts a certain |0> state into an equally probable superposition state. When n quantum bits are simultaneously applied with Hadamard gates, the entire quantum system is in all This initialization method ensures that the algorithm does not have a priori preference for any specific solution when searching for the optimal label combination, providing the largest search space for the subsequent quantum optimization process.

[0056] In practical quantum circuit implementations, creating a uniform superposition state requires only n parallel Hadamard gate operations, resulting in a circuit depth of constant 1, yet capable of simultaneously representing an exponential number of state combinations. This exponential state representation capability embodies the fundamental advantage of quantum computing over classical computing, enabling the system to simultaneously consider all possible labeling options in polynomial time, laying the computational foundation for solving large-scale multimodal labeling problems.

[0057] 2. Alternating evolution of quantum Hamiltonian: The present invention uses the alternating evolution of two Hamiltonians to achieve quantum optimization of label combinations: the cost Hamiltonian and the mixed Hamiltonian Cost Hamiltonian It encodes the objective function and constraints of the multimodal labeling problem, converting the classic discrete optimization problem into the energy function of the quantum spin system. The objective function for the label selection problem is min ∑ᵢ C(yᵢ) + ∑ᵢ<ⱼ P(yᵢ,yⱼ), where C(yᵢ) represents the cost of selecting label i and P(yᵢ,yⱼ) represents the conflict penalty when labels i and j are selected simultaneously. The system converts this into the corresponding quantum Hamiltonian form: .

[0058] Quantum encoding of single label cost term adopts In the form of, where I is the identity matrix, is the Pauli-Z operator acting on the i-th qubit. This construction ensures that when qubit i is in the |0> state, it contributes a cost C(yᵢ), and when it is in the |1> state, it contributes zero cost, thus realizing the optimization logic of "zero cost for selecting a label, cost for not selecting a label". The inter-label conflict term is adopted The encoding method ensures that the corresponding penalty cost is only incurred when two conflicting labels are selected at the same time.

[0059] Evolution operator of cost Hamiltonian In quantum circuits, this is achieved through a combination of rotation gates and control gates. The cost term for a single quantum bit corresponds to Rotation gate, the two-qubit conflict term is realized by Z gate, and its standard decomposition is: CNOT(i,j) → RZ(γP(yᵢ,yⱼ) / 2, j) → CNOT(i,j). Mixed Hamiltonian Responsible for maintaining the quantum superposition of the system and preventing premature convergence to the local optimal solution. Its evolution operator By applying Revolving door to achieve.

[0060] System Execution Layer QAOA evolution, alternating application of cost and mixed Hamiltonian evolution operators: ,in and For the Adjustable parameters of the layer. Parameters Controls the intensity of the system's evolution toward a lower energy state, parameter Controls the strength of the quantum tunneling effect. The balance between the two determines the trade-off between exploration and exploitation in the algorithm. These parameters are adjusted by classical optimizers (such as COBYLA, Adam, or L-BFGS), and the optimization goal is to minimize the expected energy. .

[0061] 3. Quantum measurement and optimal label combination extraction: After the QAOA evolution is completed, the system is in a complex quantum superposition state that encodes the solution to the optimization problem. In order to obtain the classical label selection result, the system performs a computational basis measurement operation on all quantum bits. According to the Born rule of quantum mechanics, the measurement obtains a specific bit string The probability of Ideally, the quantum state optimized by QAOA should enhance the probability amplitude of low-cost label combinations while suppressing the probability amplitude of high-cost combinations.

[0062] The system estimates this probability distribution by repeatedly performing N measurement operations. The number of measurements N needs to strike a balance between statistical accuracy and computational overhead, and is usually set to 1024 to 8192 times. , No. Bit value Directly express The selection state of each tab: Indicates that this tag is selected. Indicates that the tag is not selected. The system counts all measurement results and calculates the frequency of occurrence of each tag combination as its probability estimate.

[0063] For each measured label combination z, the system calculates its corresponding classical cost: The calculation method of single label cost is: if Contribution ,like The calculation of the conflict penalty requires checking all label pairs: if And label and If there is a conflict, the contribution penalty value The system sorts all unique measurement results by their classical cost values ​​and selects the label combination with the lowest cost as the optimization result.

[0064] To improve the reliability of the results, the system employs a multi-verification strategy rather than simply selecting the most frequently occurring combination. This includes calculating the costs of all measurement results and selecting the optimal solution by cost sorting; maintaining a set of candidate solutions containing all solutions whose costs fall within a set threshold; performing confidence analysis on the optimal solution, assessing its statistical reliability using Wilson confidence intervals; and implementing local search optimization, starting from the optimal solution given by QAOA and searching for further improvements within its Hamming distance neighborhood.

[0065] The system also implements uncertainty quantification, leveraging the probability distribution information inherent in QAOA to assess the confidence of annotation results. By analyzing statistical characteristics such as the probability of optimal solutions, the quality distribution of suboptimal solutions, and the robustness of solutions, the system provides decision makers with a confidence assessment of the annotation results. This uncertainty information is valuable in high-risk application scenarios such as medical diagnosis and content review, facilitating collaborative human-machine decision-making and result verification.

[0066] When anomalies in quantum computation results are detected (e.g., violations of hard constraints or low confidence), the system automatically activates a classical fallback mechanism, switching to a rule-based heuristic algorithm or integer linear programming method to ensure reliable completion of the labeling task. By integrating quantum optimization with classical post-processing, this invention achieves an efficient, accurate, and reliable label selection scheme for multimodal data.

[0067] This process is essentially to find the optimal label assignment strategy through quantum path interference, which has significant advantages in scenarios with multiple targets and multiple time periods in the image.

[0068] 4. Causal reasoning module. To achieve explainable reasoning, the present invention designs a causal network structure based on the quantum probabilistic graphical model.

[0069] Modeling method: Constructing a directed graph ,The node V represents the modal features or intermediate state variables, and the edge E represents the quantum conditional dependency.,Each node corresponds to a qubit group, whose joint probability is calculated through,measurement.

[0070] Using quantum states Measurement operation on , derive: Through the Bayesian chain rule, the causal path of label prediction and the marginal contribution rate of each modality can be derived.

[0071] QBN supports multiple rounds of quantum sampling analysis on the results of a single inference, improving the tolerance for uncertain decisions.

[0072] The Quantum Bayesian Network (QBN) is an inference model that maps the classical Bayesian network structure to quantum state space. Each modal feature is encoded as a quantum bit, and the conditional probability relationships between modalities are constructed using quantum gates (such as CNOT and CRZ) to construct a quantum graphical model. The system performs conditional quantum sampling on a given output label to determine the probability of each input modality's influence on the current inference.

[0073] The QBN module supports multiple rounds of quantum sampling analysis on single inference results, and estimates the inference stability and credibility interval by statistically analyzing sample distribution, thereby effectively improving the fault tolerance and explanation ability of uncertain decisions.

[0074] The relationship between the two modules is as follows: QAOA is mainly used to find the optimal solution in the label search space, while QBN is used to reversely analyze the impact path of features on decision-making after obtaining the label combination results, forming an "optimization-explanation" closed-loop structure.

[0075] 5. Quantum-classical collaborative scheduling module. Considering the limitations of quantum hardware in the NISQ stage, the present invention designs a collaborative controller for dynamic task allocation between classical and quantum computing resources.

[0076] The control logic is as follows: Evaluate the complexity of the task (e.g., number of modalities, feature dimensions, and computing resources); Schedule quantum module queues and running windows; Automatically fall back to an approximate classical simulation path when quantum computing is unavailable or queued; The quantum module measurement results are collected in real time and fed back to the classical optimizer for circuit updates.

[0077] The controller manages the modal input stream in parallel through a multi-threaded scheduling strategy, achieving execution while inputting, meeting the latency and power consumption control requirements in actual deployment.

[0078] The present invention integrates the following innovative elements in principle: Convert high-dimensional modal mapping tasks into quantum state representation and circuit evolution; use variational circuits to replace classical deep neural network modeling; introduce quantum optimization algorithms to deal with label constraints and conflicts; construct quantum Bayesian networks to achieve causal interpretability; and design dynamic collaborative scheduling mechanisms to adapt to quantum hardware resources.

[0079] This principle is not only applicable to approximate quantum devices in the NISQ stage, but also lays the foundation for the basic architectural model of AI systems under future Fault-Tolerant Quantum Computers.

[0080] In another exemplary embodiment, referring to Figure 1 , provides a data annotation method based on a collaborative computing architecture of quantum computing, comprising the following steps: Step 1: Multimodal data input and preprocessing After multimodal data is input, it undergoes intra-modal normalization. Images are cropped to a uniform resolution, text is segmented and vectorized, speech is converted to mel-spectrograms, and time series data (such as sensor data) is processed into time series segments with fixed time windows.

[0081] Step 2: Classical feature extraction and preliminary representation Process each modal data in turn and output a preliminary representation: Image data is trained using a pre-trained CNN (such as ResNet50) to extract image feature vectors. Text data is semantically embedded using a Transformer architecture (such as BERT). The audio is processed through a convolutional-gated network to extract spectral features; Time series data is represented by state extraction through the GRU network.

[0082] Step 3: Coordinated Scheduling The entire system is managed by a collaborative controller, which dynamically allocates quantum or classical computing resources based on information such as task complexity, number of data modes, and quantum node load, and dynamically allocates tasks between classical and quantum computing resources, such as Figure 2 As shown in the figure, whether to transfer to quantum computing is determined based on the task; when executing quantum computing, computing resources are also adjusted according to the characteristics of multimodal data. For example, when the audio modality quality in multimodal data is poor, the controller will reduce the weight of this modality and prioritize images and text for annotation; at the same time, by executing quantum optimization and feature mapping tasks in parallel, resource utilization is maximized.

[0083] Step 4: Quantum feature mapping and modality fusion The extracted modal features are fed into the quantum feature mapping and modal fusion module. This module works through the following steps: Each modal feature is normalized and encoded into a quantum state using an amplitude encoding scheme, i.e. -dimensional vector encoded into qubits; The quantum states of all modes are fed into a multimodal fusion circuit, in which parameterized quantum gates (such as RY, RZ, CNOT) are used to construct variational quantum circuits (such as Figure 3 As shown in ), interaction between modalities is realized; after the modal fusion is completed, the probability distribution of the quantum state finally projected into the label space is obtained through measurement operations.

[0084] Step 5: Label Generation and Quantum Optimization A prediction circuit based on a quantum support vector machine (QSVM) or quantum variational classifier (VQC) is designed to perform label inference on the fused state. Furthermore, to optimize the confidence and consistency of the label output, the QAOA algorithm is used for label constraint optimization. For example, for mutually exclusive labels, quantum Boolean constraints are used to resolve conflicts and select the optimal label combination.

[0085] Step 6: Causal Inference and Explanation Quantum Bayesian networks or approximate causal graph generation methods are introduced to generate explanations based on the modal contribution causal paths; using the joint probability measurement results, the marginal contribution of each modality to the final label prediction is derived, and an explanation report is output.

[0086] Step 7: Output the annotation results.

[0087] The above specific implementation methods are detailed descriptions of the present invention. It cannot be considered that the specific implementation methods of the present invention are limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, they can make several simple deductions and substitutions without departing from the concept of the present invention, which should be regarded as falling within the scope of protection of the present invention.

Claims

1. A data annotation method based on a collaborative computing architecture of quantum computing, characterized in that: The following steps are involved: S1. Input multimodal data and perform preprocessing; S2. Extract the features of each modality after preprocessing; S3. Encode the features of each modality into quantum states and perform modal fusion; S4. Perform label inference on the quantum state after modal fusion and use quantum approximate optimization algorithm to perform label constraint optimization; S5. Generate explanations based on the modal contribution causal paths using quantum Bayesian networks or approximate causal graph generation methods, and use joint probability measurements to derive the marginal contribution of each modality to the final label prediction; S6. Output the annotation results.

2. A data annotation method based on a collaborative computing architecture of quantum computing according to claim 1, characterized in that: Before executing step S3, the method further includes: Dynamically allocate quantum or classical computing resources according to task complexity.

3. The data annotation method based on the collaborative computing architecture of quantum computing according to claim 2 is characterized in that: The dynamic allocation of quantum or classical computing resources includes: Evaluate the complexity of the task; Automatically fall back to classical computing when quantum computing is unavailable or queued; Collect quantum computing measurement results in real time and feed them back to the classical optimizer for circuit updates; The modal input stream is managed in parallel through a multi-threaded scheduling strategy to achieve execution while inputting.

4. The data annotation method based on the collaborative computing architecture of quantum computing according to claim 1 is characterized in that: The dynamic allocation of quantum or classical computing resources further includes: Adjust the scheduling priority according to the quality of each modality.

5. The data annotation method based on the collaborative computing architecture of quantum computing according to claim 1 is characterized in that: The input multimodal data and preprocessing include: Intra-modal normalization is performed separately, where images are cropped to a uniform resolution, text is segmented and vectorized, audio is converted to Mel-spectrograms, and time series data is processed into time series segments with fixed time windows.

6. The data annotation method based on the collaborative computing architecture of quantum computing according to claim 1 is characterized in that: The extraction of features of each modality after preprocessing includes: The image is extracted through pre-trained CNN. The text is semantically embedded through the Transformer structure; The audio is processed through a convolution-gated network to extract spectral features; Time series data is represented by state extraction through the GRU network.

7. The data annotation method based on the collaborative computing architecture of quantum computing according to claim 1 is characterized in that: In step S3, the features of each modality are encoded into a quantum state through amplitude encoding; the modality fusion includes: A variational quantum circuit is constructed using parameterized quantum gates to achieve interaction between modes. The variational quantum circuit includes a parameter layer, an interference layer, and an output measurement layer. The parameter layer is composed of parameter-controlled rotation gates, the interference layer introduces control gates to form inter-modal interference, and the output measurement layer measures specific quantum bits to obtain classification probabilities.

8. The data annotation method based on the collaborative computing architecture of quantum computing according to claim 1 is characterized in that: In step S4, a prediction circuit based on a quantum support vector machine or a quantum variational classifier is used to perform label inference on the quantum state after modal fusion.

9. The data annotation method based on the collaborative computing architecture of quantum computing according to claim 1 is characterized in that: The label constraint optimization using the quantum approximate optimization algorithm includes: Define the cost function: in, Indicates the selection tag the price; Representation Label Conflict penalties; Construct the corresponding Hamiltonian and use the quantum approximate optimization algorithm circuit to execute: The initial state is uniform superposition; use and Alternate; find the label combination with the lowest cost through measurement.

10. A data annotation system based on a collaborative computing architecture of quantum computing, characterized in that: include: Multimodal data preprocessing module, used to input multimodal data and perform preprocessing; Multimodal feature extraction module, used to extract features of each modality after preprocessing; The quantum feature mapping and modal fusion module is used to encode the features of each modality into quantum states and perform modal fusion; The label generation and optimization module is used to perform label inference on the quantum state after modal fusion and use the quantum approximate optimization algorithm to perform label constraint optimization; A causal reasoning module, which generates explanations based on modal contribution causal paths using quantum Bayesian networks or approximate causal graph generation methods, and uses joint probability measurements to derive the marginal contribution of each modality to the final label prediction; The annotation output module is used to output the annotation results.

Images