Hybrid quantum-classical computing system for processing artificial intelligence applications

Abstract

A hybrid quantum-classical computing system designed to enhance AI applications by integrating classical computational methods with quantum computing power. The system incorporates a classical computing unit and multiple quantum computing units, each specialized for distinct algorithmic tasks. A control module within the classical unit, using predetermined decision logic, manages the distribution of tasks, selectively engaging quantum units for their unique capabilities to optimize efficiency. This scalable architecture allows for significant advances in AI, tackling complex, resource-intensive tasks by effectively harnessing the strengths of both quantum superposition and classical processing.

Description

[0001] Hybrid Quantum-Classical Computing System for Processing Artificial Intelligence Applications

[0002] The present invention relates to a hybrid quantum-classical computing system for artificial intelligence (Al) applications, particularly to a modular hybrid quantum-classical computing system for artificial intelligence (Al) applications.

[0003] Background of the Invention

[0004] As society moves ever further into the information age, the role of artificial intelligence (Al) in shaping our world cannot be overstated. Al applications are becoming increasingly important across a wide range of industries, from healthcare diagnostics to autonomous driving and from personalized education to supply chain management. At the heart of these Al applications are complex algorithms that require immense computing resources. Tasks such as training large language models, performing inference, and fine-tuning machine learning systems require an extraordinary amount of computing power and often necessitate the use of state-of-the-art supercomputers to achieve the desired level of performance and efficiency

[0005] While current supercomputing resources are adequate to support Al applications, the inexorable expansion in the scale and sophistication of Al models will outpace the growth of classical computing capacity. The potential of quantum computing, with its promise of exponential growth in computing power using the principles of quantum mechanics, has long been seen as a possible solution to this looming bottleneck. However, quantum computing technology is still in its infancy, especially considering the leap required for industrial and commercial applications.

[0006] The crux of the matter is the merging of quantum computing with Al applications - a vision that, if realized, could unleash previously unimaginable levels of computing power and creativity. While there are a growing number of ideas and experimental approaches to accelerate Al processing using quantum computers, the path from theoretical potential to practical implementation is still steep and full of complicated challenges. So, the urgent challenge we face is not only the theoretical development of quantum computers, but also the pragmatic implementation of quantum-accelerated Al applications. Bridging this gap is more than just an academic endeavor. It has the potential to revolutionize the industry and enhance human capabilities beyond our current understanding, catapulting us into a future where Al and quantum computing work hand in hand to solve the most complex problems facing the world today.

[0007] Brief Summary of the Invention

[0008] To address the challenges the following system is presented, a hybrid quantum-classical computing system for processing artificial intelligence (Al) applications, comprising a classical computing unit and at least a first and a second quantum computing unit, each of which is communicatively connected to the classical computing unit. The first quantum computing unit is configured to compute a first algorithmic component, and the second quantum computing unit is configured to compute a second algorithmic component. The classical computing unit includes a control module, which is configured to determine whether an Al application at least in part being processed by the classical computing unit requires computation of the first and or the second algorithmic component, and which is further configured to determine, based on a predetermined decision logic, whether the first and or the second algorithmic component is to be computed by the respective first and or second quantum computing unit if the Al application at least in part being processed by the classical computing unit requires the computation of the first and or the second algorithmic component. The control module is further configured to control the first and or the second quantum computing unit to compute the respective first and or second algorithmic component if the control module has determined that the first and or the second algorithmic component is to be computed by the respective first and or second quantum computing unit.

[0009] The system described is a hybrid quantum-classical computing architecture that specializes in the implementation of artificial intelligence (Al) applications. This system is an integration of classical computational methods with the advanced computational capabilities of quantum computing. It is structured to optimize specific parts of Al algorithms by distributing tasks according to their complexity and suitability for classical or quantum computation.

[0010] The classical computing unit serves as the primary processing and orchestration center for the Al applications. It is equipped with conventional processors that can execute most Al algorithm tasks, often those involving classical data processing and decision-making processes. The classical computing unit also houses a control module that acts as the operational core of the system.

[0011] The control module is a component within the classical unit that is responsible for task management and distribution between the classical and quantum computing units. It is designed to apply a set of predetermined decision logic criteria characterized by efficiency, computational necessity, and the specific requirements of the Al application.

[0012] The decision logic dictates how the control module evaluates the Al application processes to determine which algorithmic components could benefit from quantum acceleration. This evaluation considers aspects such as the complexity of the problem, suitability for quantum computation and potential performance gains.

[0013] The system comprises at least two quantum computing units. Each quantum unit is specialized to compute specific algorithmic components within the Al application domain. Quantum mechanics phenomena such as superposition and entanglement are used to perform tasks that are either infeasible or too time-consuming for classical computers.

[0014] The first quantum computing unit is tailored to perform a specific first algorithmic component of the Al application. For example, it could be dedicated to tasks that can be significantly optimized by quantum algorithms, such as optimization problems, sampling or simulation of quantum systems that are part of the broader Al problem to be solved.

[0015] Similarly, the second quantum computing unit is intended to execute another, stand-alone algorithmic component that may deal with other quantum-based computations, such as quantum learning models or quantum annealing processes required by the Al application.

[0016] One exemplary application of this principle is the Quantum Amplitude Estimation (QAE) algorithm, which is adeptly suited for probabilistic inference and tasks akin to Monte Carlo simulations. By estimating the amplitude of a quantum state, which can be analogized to a probability in Al models, QAE has the potential to provide a quadratic speedup. This is mainly because it diminishes the number of samples required to precisely estimate a probability to within a particular error boundary. To precisely execute QAE, a gate-based quantum computer is normally suggested, needing the ability to create entangled states and suitable oracles for the objective function. However, in such a scenario, a fully equipped set of universal quantum gates is not a necessity. It is therefore, suggested to use an algorithm- specific quantum computer with the focus on specific gate configurations that are instrumental for implementing QAE, such as specific gate sequences tailored to QAE.

[0017] Similar principles apply to the Quantum Support Vector Machine (QSVM), which leverages quantum-enhanced feature mappings in classification problems. Here, the 'quantum kernel trick' has the potential to exponentially expedite the evaluation of inner product within a highdimensional feature space. For the QSVM, a quantum computer would principally necessitate the capability to prepare quantum states that represent data points, and then to precisely measure those states in relation to evaluating a kernel matrix. However, when building an algorithm-specific quantum computer for QSVM, some specific gate sequences not related to state preparation, transformation, and measurement can be skipped, focusing instead on gates that facilitate the quantum kernel.

[0018] Furthermore, in the Quantum Nearest Neighbor (QNN) algorithm, a quantum approach drastically accelerates distance estimations between data points. By leveraging quantum parallelism to calculate inner products or distances expeditiously, QNN opens possibilities for potential quadratic speedups in selected cases. Here, the quantum computer's architecture would need to emphasize state preparation and the specific quantum gates required for distance computation. Full connectivity between qubits might not be necessary. Hence, in a proposed algorithm-specific quantum computer for QNN, only those quantum gates needed for state preparation and distance calculations are provided.

[0019] Likewise, quantum clustering algorithms, which would augment tasks within algorithms like k- means, benefit from quantum speedups in distance computations and efficient centroid updates. In these applications, a gate-based computer would demand dedicated mechanisms for data representation through state preparation, distance calculation circuits, and mechanisms to perform interference operations for updates of centroids. Extensive qubit connectivity is generally not be needed, except among those qubits that represent data points and clusters. This finding can beneficially be used when building an algorithm-specific quantum computer for the quantum clustering algorithm.

[0020] In the construction of an algorithm-specific quantum computer tailored to a single algorithm, several simplifications can be envisaged. The quantum gate set could be meticulously tailored, encompassing only those gates integral to the execution of the algorithm, simplifying quantum control hardware. Such customization would enhance efficiency by reducing qubit connectivity to only that which the algorithm's data flow necessitates, deviating from a universal design to a more specialized one. By optimizing the control scheme specifically for the chosen algorithm, not only is the sequence of operations managed more proficiently, but also simpler calibration and error-mitigation techniques could be employed. To further streamline this process, the creation of bespoke quantum compilers and optimizers would be invaluable for converting high-level descriptions of algorithms into actionable operations on the custom quantum hardware.

[0021] The construction of these algorithm-specific quantum computers considers the intrinsic limitations of quantum hardware, such as error rates and decoherence times, the principles of high-quality qubits and gates remain foundational. Nevertheless, a dedicated quantum computer for specific algorithms could potentially require less stringent error correction solutions, particularly if the executions can be completed rapidly within the qubits' coherence times.

[0022] The process is such that the classical computing unit executes the Al application, and the control module continuously evaluates the computational needs. When the control module identifies an algorithmic component that could processed using quantum computation, it decides whether to outsource this component to the first or second quantum computing unit based on its embedded logic. If the outsourcing of computations is deemed beneficial, the control module sends the required data and instructions to the selected quantum unit.

[0023] The calculations on the quantum units are performed according to the quantum algorithms suitable for the identified algorithmic component. Once the quantum calculations are completed, the results are transmitted back to the classical computing unit, where the control module reintegrates these results into the overall processing flow of the Al application.

[0024] To ensure effective communication between the classical and quantum units, the system could include quantum-classical interfaces that include error correction, translation of quantum results into classically readable formats and vice versa.

[0025] The design of such a hybrid quantum-classical system aims to harness the strengths of both computational paradigms. By selectively using quantum computing for parts of the Al algorithms that are bottlenecks in classical systems, the architecture promises significant advances in computational efficiency and capability, especially for complex and resourceintensive Al tasks. The predetermined decision logic bases its result on one or more of the following: an algorithm's suitability, a task's complexity, a quantum speedup potential, an error rate, a qubit's stability, a resource's availability, a problem's encoding, a hybrid algorithm's compatibility, a data transfer overhead, a model's type, an Al application's domain, a scalability, a quantum readiness, a cost consideration.

[0026] The described system is based on the realization that in next-generation computing, one of the most promising endeavors is the development of hybrid classical quantum systems, especially in the field of artificial intelligence (Al). The heterogeneous nature of these systems requires a robust and sophisticated decision-making process within the classical control module to manage and efficiently distribute tasks between classical and quantum components.

[0027] One factor in the control module's decision matrix is the suitability of an algorithm for quantum computation. Quantum computation, with its inherently different operating principles, is ideally suited for certain types of problems, such as factoring large numbers or simulating quantum systems. When an Al application encounters a subtask that matches these strengths, delegating this task to a quantum processor can lead to significantly improved performance.

[0028] The complexity of a task indicates whether it is more suitable for a quantum or a classical processor. Some problems that require exponential scaling on classical systems can reach polynomial time complexity on quantum systems due to quantum parallelism. Identifying such complexities allows the control module to optimize computational resources.

[0029] The promise of quantum speedup - when a quantum algorithm significantly outperforms its classical counterpart - is a major incentive for the use of quantum computation. Tasks with proven or expected quantum acceleration potential are prime candidates for quantum units in the system to ensure accelerated problem solving for sophisticated Al components.

[0030] Quantum systems, especially those using current technology, are prone to high error rates due to decoherence and noise. The feasibility of executing algorithms on quantum processors is highly dependent on the probability of errors that the control module must evaluate to ensure the integrity of the computations. The stability of qubits, the fundamental units of quantum information, is critical to the success of quantum computations. The control module must consider qubit coherence times and error correction techniques when scheduling tasks to ensure that the quantum computation is completed before the qubits lose their quantum state. The allocation of tasks also depends on the availability of computing resources. This includes the number of qubits, the type of quantum gates available and the connectivity between the qubits. The control module must match the requirements of the tasks with the available quantum resources to enable optimal execution.

[0031] Translating a real-world problem into a quantum framework - quantum coding - is a nuanced process. Some problems can be expressed naturally in the language of quantum algorithms, while others require complex transformations. The control module should evaluate the complexity of the encoding to decide on the processing locations.

[0032] Seamless integration of quantum and classical computations is central to the functionality of a hybrid algorithm. The control module must consider compatibility issues and ensure that the data output by a quantum process is in a form that can be used by classical algorithms and vice versa. In a hybrid system, data transfer between the classical control unit and the quantum processors can be time and resource consuming. The control module must weigh up whether these additional costs outweigh the benefits of quantum computation to optimize the overall performance of the system.

[0033] Al models vary considerably in their structure and outlined processes, from neural networks to decision trees. The decision logic within the control module must consider which types of Al models can utilize the improvements of quantum processing without excessive integration issues. Scalability preferably also plays a role in the control module's decisions. The system should scale economically with the addition of more quantum or classical resources. This consideration ensures that the hybrid system retains its performance advantages as Al applications increase in size and complexity without excessive resource consumption.

[0034] Not all tasks or systems are "quantum-ready". The control module must determine whether the current state of algorithm development and quantum technology is sufficient to effectively accomplish a given task, balancing the innovative potential of quantum computing with the reliable maturity of classical methods. Finally, cost considerations are a non-trivial factor in practice. Quantum computing resources can be more expensive, so cost-efficient computation is an essential part of the equation. To ensure the cost-effectiveness of an Al application, the control module should assess whether the potential quantum acceleration justifies the additional cost.

[0035] The control module includes at least one of: a hybrid quantum-classical scheduler, a decision tree-based controller, a rule-based inference system, a machine learning-based predictor, an adaptive load balancer, an SLA-aware allocator, an API-based integration with quantum providers, a cost-performance analyzer, a priority queue manager, a fallback mechanism.

[0036] A scheduler in a hybrid system must manage the allocation of tasks to both quantum and classical computing resources. It dynamically schedules the Al processes based on the current state of the system and the nature of the tasks. The hybrid scheduler considers the specific requirements of quantum algorithms, such as the need for certain qubit arrangements, together with conventional criteria of resource utilization and throughput optimization. By scheduling tasks where they are most appropriate, the scheduler increases the overall efficiency of the system and ensures that quantum acceleration is utilized whenever possible.

[0037] When using a decision tree-based controller, a comprehensive decision path is designed for each potential scenario that the hybrid system might encounter. The advantages lie in the structured and logical way in which the controller can evaluate and manage tasks. It can break down complex decisions into a series of simpler binary decisions, leading to quick and effective decisions on whether to process tasks classically or quantum based.

[0038] A rule-based inference system in the control module contains a set of predefined rules that control the behavior of the system. The inference engine analyzes the tasks based on these rules to determine the most appropriate computing resource. This system is particularly advantageous in environments with well-understood parameters and where procedural logic can dictate the best course of action for a given task.

[0039] A machine learning-based predictor can analyze historical data and performance metrics to predict the best computational strategies for new tasks entering the system. Over time, such a predictor can optimize its decisions based on learned patterns of when and how quantum processing will provide the greatest benefit. Incorporating machine learning helps the system adapt and improve over time, potentially allowing it to make predictive adjustments in resource allocation before bottlenecks occur.

[0040] An adaptive load balancer responds in real time to fluctuations in system demand and distributes tasks between quantum and classical resources to balance the load and prevent over- or under-utilization of either resource type. The adaptive nature ensures that the load balancer responds to changes in task volume or resource performance, resulting in a continuous and efficient distribution of computing tasks.

[0041] Service Level Agreements (SLAs) define the expected performance and reliability standards for computing systems. An SLA-aware allocator within the control module would ensure that the distribution of tasks is optimized not only in terms of performance or cost, but also in terms of compliance with these agreements. This is critical to ensure customer satisfaction and maintain confidence in the hybrid system's capabilities.

[0042] Given the diverse and emerging quantum computing resources that are often available as cloud-based services, API-based integrations allow the hybrid system to connect with multiple external quantum providers. Such integration ensures that the system can take advantage of the best quantum computing performance offerings without being tied to a single provider, promoting flexibility and scalability.

[0043] A cost-performance analyzer ensures that the economic aspects of task allocation may also be considered. By analyzing the cost impact of using quantum versus classical resources for specific tasks in conjunction with the performance benefits, this analyzer helps to make decisions that are not only technically robust, but also financially sound.

[0044] Al tasks can vary significantly in terms of their urgency and importance. A priority queue manager organizes the tasks according to their defined priority level and ensures that the tasks with higher importance are processed faster. This function is particularly beneficial for time-critical Al applications, where certain tasks can significantly affect the result if they are not processed immediately.

[0045] Finally, a fallback mechanism provides a safety net for the inevitable uncertainties that come with cutting-edge computing technologies. Should a quantum component fail or underperform, the system can fall back on classical processing methods to maintain continuous operation. This mechanism provides resilience and reliability, which are critical for user confidence and supporting mission-critical Al applications.

[0046] The control module in a hybrid quantum computing system as described provides a versatile framework that ensures tasks are processed on the platform that is most appropriate, considering the type of algorithm, complexity, urgency and availability of resources.

[0047] The first and or the second algorithmic component is accessed by using at least one of: a quantum-capable Application Programming Interface (API), Software Development Kit (SDK), a function invocation, data translation and encoding, task submission, a call-back.

[0048] The successful operation of a hybrid quantum-classical computing system for processing artificial intelligence (Al) applications depends on the seamless cooperation and integration between the classical and quantum computing units. This integration is facilitated by the interface mechanisms that enable communication and operation between the two paradigms. In particular, the first and / or second algorithmic components within the system can be accessed through various tools and protocols, such as quantum-enabled application programming interfaces (APIs), software development kits (SDKs), function calls, data translation and encoding mechanisms, task submission modules and callback procedures. Each of these interfaces offers certain advantages and is tailored to specific requirements within the hybrid system.

[0049] A quantum-enabled API is a set of protocols and tools that allow developers to access and control quantum computing resources in a way that abstracts away from the underlying quantum hardware and its complexity. APIs are designed to be user-friendly and allow classical software to seamlessly request and retrieve quantum computation results. The advantage of a quantum API in a hybrid system is that it allows easy access to quantum processors without extensive knowledge of the underlying physics, making quantum computation more accessible for Al applications. It often provides a high-level interface that can be integrated into existing classical workflows, allowing developers to focus on the Al algorithms rather than the intricacies of quantum computation.

[0050] A quantum computing software development kit contains libraries, tools, documentation and sample code that developers can use to create, deploy and run applications that utilize quantum computing. An SDK typically contains higher level abstractions that simplify the complex quantum logic into more readable and maintainable code structures. Using an SDK allows developers to use quantum resources more effectively by providing an organized suite of programming tools. This structured environment speeds up the development process for Al applications that require quantum algorithms, ensures that best practices are followed and reduces the potential for errors.

[0051] Function calls represent a direct method call within the software that requests a specific operation or calculation. In a hybrid system, this could mean calling a quantum function within an otherwise classic codebase. The function takes care of the details of the interface with the quantum processor and abstracts from the lower-level operations. The advantage of the function call is the immediacy and encapsulation of quantum operations in callable functions. This approach enables the integration of quantum computation into Al applications with minimal impact on the overall structure and design of the software.

[0052] Quantum computers process data in a format that differs from that of classical computers and requires special forms of encoding for quantum states and operations. Consequently, a data translation and encoding interface may be provided to integrate quantum and classical computation. The benefits of such interfaces include compatibility, which ensures that data flowing between classical and quantum systems can be translated appropriately while preserving the meaning and structure of the data. Proper encoding also maximizes the efficiency of quantum algorithms, which often depend on specific data representations, such as the quantum Fourier transform for certain types of problem solving.

[0053] Task submission interfaces are used to queue tasks for quantum processing. This often involves preparing the task description, encoding the required quantum circuits, and managing the execution queue. The main advantage is the convenience and management of asynchronous quantum tasks. Al applications can submit quantum tasks and proceed with classical processing, knowing that quantum computation will be performed in parallel. This interface is used for optimal resource utilization and scheduling in applications where quantum processing is an integral part, but immediate results are not required.

[0054] A call-back is a programming construct that specifies a piece of code to be executed in response to certain events or conditions. In a hybrid system, a call-back can be used to inform the classical system of the completion of a quantum computation. By incorporating call-backs, Al applications can continue with other tasks while waiting for the results of the quantum computation. This event-driven approach is beneficial for efficiency as it allows the classical system to continue working without constantly polling the quantum result. Callbacks provide a mechanism for synchronization between the classical and quantum computing units and ensure that the continuity of the workflow is maintained.

[0055] The classical computing unit includes one or more of: a Central Processing Unit (CPU), a CPU Core (Core), a Graphics Processing Unit (GPU), a Field Programmable Gate Array (FPGA), an Application-Specific Integrated Circuit (ASIC), a Digital Signal Processor (DSP), a System on a Chip (SoC), an Accelerated Processing Unit (APU), a Tensor Processing Unit (TPU).

[0056] Central Processing Units (CPUs) and their individual CPU cores form the overall computational backbone, providing versatile and robust processing capabilities for a wide range of Al algorithms. Their sequential execution capabilities ensure stable operation for tasks where predictability and accuracy are paramount.

[0057] For data-intensive Al workloads such as deep neural networks, Graphics Processing Units (GPUs) are characterized by their parallel processing power. Accelerated Processing Units (APUs) and Tensor Processing Units (TPUs) are a more specialized subset optimized for advanced graphical and tensor operations that are commonplace in machine learning.

[0058] Field Programmable Gate Arrays (FPGAs) offer a reconfigurable architecture that enables customization for specific Al tasks, increasing both performance and efficiency. In contrast, application-specific integrated circuits (ASICs) are optimized for specific functions and offer maximum efficiency, such as Google's TPUs used for tensor computing.

[0059] Digital signal processors (DSPs) are suitable for real-time processing and are essential for Al applications that involve signal manipulation. Systems on a chip (SoCs), on the other hand, integrate multiple processing units into a compact and energy-efficient package and are ideal for mobile and embedded Al systems.

[0060] The first and the second quantum computing unit, each include one of: a gate-based quantum computer, a quantum annealer, a topological quantum computer, an adiabatic quantum computer, a quantum simulator, quantum dot-based quantum computers, silicon quantum computers, noisy intermediate-scale quantum (NISQ) computer, a quantum neural network (QNN). A gate-based quantum computer uses quantum gates to manipulate qubits and is suitable for a variety of quantum algorithms. Quantum annealers solve optimization problems by finding low-energy states of a system, which is beneficial for specific Al use cases such as pattern recognition. Topological quantum computers promise more stable quantum computations by using topological states that are less error prone. Adiabatic quantum computers are gradually evolving the quantum system to solve optimization problems and can be used in areas where robustness is important. Quantum simulators have been developed to model complex quantum systems that classical computers struggle to cope with, which is instructive for quantum chemistry problems relevant to Al. Quantum dot and silicon quantum computers represent material-based approaches to realizing quantum bits that hold promise for scalability and fabrication with existing semiconductor technologies.

[0061] Noisy Intermediate-Scale Quantum (NISQ) computers are the current reality of quantum computing and offer pragmatic, albeit error-prone, quantum processing capabilities. Finally, quantum neural networks (QNNs) represent a fusion of quantum computers and neural networks and are particularly suitable for Al systems that want to utilize quantum superposition and entanglement for machine learning.

[0062] The gate-based quantum computer is based on one of the following: superconducting qubits, trapped ions, photonic systems, wherein the quantum annealer is based on superconducting qubits, wherein the topological quantum computer is based on Anyons and Majorana fermions.

[0063] Gate-based quantum computers use qubits as fundamental computing units, with several underlying technologies enabling their generation and manipulation. Superconducting qubits utilize the principles of superconductivity to enable low-resistance current flow, allowing fast and coherent quantum gate operations. Trapped ions utilize the quantum states of charged atoms, which can be precisely manipulated using lasers. Photonic systems utilize the quantum properties of light particles (photons) for their qubits, offering stability and the promise of integrated photonic circuits.

[0064] The quantum annealer, traditionally based on superconducting qubits, is optimized for solving optimization problems. This type of quantum computer excels at finding global minima in complex energy landscapes, which is crucial for Al applications in the fields of optimization and pattern recognition. Furthermore, the theoretical framework of topological quantum computers is based on the existence of exotic quasiparticles - Anyons and Majorana fermions. These quasiparticles exhibit non-Abelian statistics, which are believed to be highly resistant to local perturbations, opening avenues towards fault-tolerant quantum computers.

[0065] Preferably, at least one of said first and said second quantum computing unit is an algorithmspecific quantum computer. An algorithm-specific quantum computer comprises a quantum processor that has been developed and optimized for the efficient execution of a specific or a specific set of quantum algorithms. The main advantage of an algorithm-specific quantum computer is the optimization of the physical resources - qubits and quantum gates - which are perfectly matched to the requirements of the algorithm. For example, a quantum computer dedicated to the Grover search algorithm could have a qubit architecture that maximizes superposition and entanglement to improve the algorithm's inherent quadratic speedup when searching through unsorted databases.

[0066] Furthermore, customizing a quantum processor to a specific algorithm or closely related family of algorithms enables a more targeted approach to error correction and mitigation strategies. This can significantly reduce the overhead typically associated with generic quantum error correction techniques, resulting in improved quantum volume - a metric that evaluates the overall capacity and effectiveness of a quantum computer.

[0067] In the field of artificial intelligence, one such approach demonstrates the integration of quantum neural networks (QNNs) with specialized quantum processors that optimally support the entangled layer structures and non-classical activations inherent in these QNN models. In this way, learning processes can be exponentially accelerated and deep insights into data patterns can be gained that simply cannot be recognized by classical means.

[0068] Advantageously, at least one of the first and the second quantum computing unit is optimized for computing the respective algorithmic component.

[0069] By focusing the design and operation of a quantum computing unit on a specific algorithm, the efficiency and capabilities of Al applications is dramatically increased. For example, a quantum processor tuned to the Quantum Approximate Optimization Algorithm (QAOA). This algorithm uses the principles of quantum superposition and entanglement to explore and optimize solution spaces at unprecedented speeds. By adapting the quantum unit to QAOA, Al tasks that involve combinatorial optimization, benefit from accelerated solution finding and higher quality solutions due to quantum acceleration. In addition, the dedicated quantum computing unit can be specifically designed to reduce decoherence and operational errors that are closely related to the chosen quantum algorithm. Such optimization leads to improved accuracy in quantum operations and enables more efficient and practical implementation of theoretically powerful quantum algorithms in Al.

[0070] The first and the second algorithmic component, each, provide at least in part the computation of one of: quantum Fourier transform, Grover's algorithm, amplitude amplification algorithm, quantum phase estimation, quantum clustering, quantum classification, quantum pattern recognition, adiabatic quantum algorithms, quantum approximate optimization algorithm (QAOA), Harrow-Hassidim-Lloyd (HHL) algorithm, quantum-enhanced Monte Carlo method, variational quantum eigensolver (VQE).

[0071] Of advantage is the ability of hybrid systems to integrate algorithmic components such as the quantum Fourier transform (QFT) or the Grover's algorithm into the quantum computing unit, enabling improved problem-solving strategies alongside their classical counterparts. For example, using the QFT as the first algorithmic component facilitates tasks that require the quantum equivalent of Fourier transforms and support algorithms for signal processing and the fast solution of linear equations when integrated with the Harrow-Hassidim-Lloyd (HHL) algorithm. Similarly, the Grover algorithm proves to be an important second component that enables quadratic speedup in database searches, which when applied to Al drastically reduces computation time for unordered data. Algorithmic components such as the Quantum Approximate Optimization Algorithm (QAOA) are central to solving complex optimization problems that permeate machine learning and Al tasks. The amplitude boosting algorithm provides a significant advantage by increasing the probability amplitude of desired outcomes, thus accelerating the convergence of probabilistic algorithms used in Al applications. Specialized quantum routines such as quantum phase estimation, clustering and classification provide accurate quantum pattern recognition results, which are crucial for the analysis of large and complex data sets common in Al big data challenges.

[0072] Another indispensable advantage arises when adiabatic quantum algorithms and the quantum-enhanced Monte Carlo method are used. They are suitable for the development of quantum annealing approaches that support decision and prediction models in Al while providing solutions to problems that would otherwise be unsolvable due to their stochastic nature. The Variational Quantum Eigensolver (VQE) provides an optimal solution for determining the ground state of complex quantum systems, which is crucial in simulations for Al-driven chemical and material discovery.

[0073] Preferably, the Al application at least in part being processed by the classical computing unit implements one or more of: training a model, fine-tuning a model, performing an inference, conducting a validation, running a test, applying transfer learning, utilizing unsupervised learning, engaging in semi-supervised learning, performing reinforcement learning.

[0074] Training a model in this hybrid framework is exponentially faster, as quantum algorithms can explore the vast solution space much more efficiently than their classical counterparts.

[0075] Consequently, fine-tuning, which is often an iterative and resource-intensive task, can benefit from speed increases, allowing Al models to achieve higher accuracy with fewer resources. When performing inference, quantum systems could evaluate multiple possibilities simultaneously, allowing them to make more nuanced decisions that leverage rich representational power.

[0076] Validation and testing of Al models will also be greatly improved, as the quantum paradigm allows for the consideration of many more scenarios within a shorter time frame, ensuring that models are robust and generalize well. Furthermore, methods such as transfer learning become incredibly powerful as quantum systems can more effectively adapt pre-trained models to new tasks by exploring a larger parameter space due to their inherent parallelism.

[0077] In the area of unsupervised and semi-supervised learning, the ability to deal with large unlabeled datasets can be greatly improved, allowing insights to be gleaned from the data that would be difficult to discover by traditional means. Finally, reinforcement learning, where agents learn from interacting with their environment, could lead to improved decision-making processes thanks to the probabilistic nature of quantum computing, which fits well with the stochastic environments often used in this approach.

[0078] Preferably, a model is one or a combination of: a language model, a vision model, a multimodal model, an audio model, a reinforcement learning model, a robotics model.

[0079] A language model is characterized by understanding and generating human language. Its capabilities range from simple tasks such as spell checking to complex tasks such as language translation and the generation of coherent texts. In conjunction with quantum computers, such models could leverage quantum parallelism for faster processing of large corpora and nuanced language patterns. In contrast, an image processing model focuses on interpretation and decision making based on visual data. By using convolutional neural networks, these models are excellent for tasks such as image classification and object recognition. Quantum enhancements can speed up feature extraction and complex pattern recognition through more efficient computation.

[0080] A multimodal model combines input from different types of data - visual, textual, auditory, etc. - to achieve a more holistic understanding. Such integration enables applications such as sentiment analysis from simultaneous image and text evaluation and benefits from the quantum ability to process data fusion at an unprecedented scale. An audio model tailored to the processing of sound waves is suitable for voice recognition, sound classification and audio generation. A quantum-based audio model could improve the speed of signal processing and open new possibilities for noise reduction and signal enhancement.

[0081] A reinforcement learning model learns to make decisions by interacting with an environment and using a system of rewards and punishments to guide its behavior. Quantum computation can refine this exploratory learning process, especially in complex environments with large action spaces or complicated reward dynamics. And, a robotics model works in the area of sensing, motor control, and decision making for physical or simulated robots. Quantum computing can optimize path planning, control algorithms, and coordination tasks for multiple robots by computing optimal solutions more efficiently in the large parameter spaces that are common in robotics.

[0082] Advantageously, the Al application at least in part being processed by said classical computing unit addresses one or more of: healthcare, retail, marketing, manufacturing, climate science, biology, physics, chemistry.

[0083] The integration of Al in key sectors is accelerating progress and increasing efficiency by tackling complicated tasks and analyzing huge amounts of data that exceed human capabilities. In healthcare, Al is improving diagnostic accuracy and enabling early detection of diseases such as cancer through image recognition algorithms. Predictive analytics can forecast disease outbreaks, while Al-powered robotics in surgery helps improve precision and shorten recovery times. Personalized medicine uses Al to tailor treatments to individual genetic profiles. In retail, Al is transforming the customer experience through personalization strategies that recommend products based on shopping behavior and preferences. Inventory management systems predict stock needs, minimize waste, and ensure product availability.

[0084] Chatbots and virtual assistants provide round-the-clock customer service.

[0085] Al tools analyze consumer data to create targeted campaigns and optimize marketing spend by identifying the most effective channels and messages. Sentiment analysis provides insights from social media and enables brands to quickly adapt to consumer trends. In manufacturing, Al improves efficiency through predictive maintenance by anticipating equipment failures before they occur, reducing downtime. Quality control processes are augmented by Al systems that detect defects more accurately than the human eye, ensuring higher product standards. In climate science, Al models simulate complex climate systems, provide more accurate predictions, and contribute to the understanding of long-term climate patterns. In biology, Al is accelerating drug discovery by predicting molecular interactions, significantly shortening research and development times. Al-supported analysis of genetic data helps to understand disease mechanisms, which can lead to breakthroughs in treatment.

[0086] Al algorithms process data from physical experiments, e.g. from particle physics, and uncover patterns and anomalies that could indicate new physical phenomena. Simulations of physical processes, which are often very computationally intensive, are only made possible by the ability of Al to perform complex calculations. In the field of chemistry, Al accelerates the development of new materials by predicting chemical reactions and molecular structures, thus saving time and resources. Al-supported analyses can lead to innovations in the fields of renewable energies, pharmaceuticals, and more environmentally friendly industrial processes.

[0087] Preferred, the first and the second quantum computing unit are used to jointly compute a computation task. This approach allows for remarkable parallelism and quantum entanglement complexity, enabling exponentially faster problem solving than can be achieved by a single quantum unit.

[0088] This collaboration is beneficial in several ways. First, by distributing tasks across multiple quantum computers, larger, more complex problems can be solved that would be prohibitively expensive for a single quantum computing unit due to qubit and coherence time constraints. Moreover, this can lead to more robust quantum algorithms, as multiple quantum computing units can specialize on different parts or aspects of the task, like distributed computing in classical systems. Secondly, interleaving multiple quantum computing units improves error correction capabilities, which is a major hurdle in quantum computations, thus improving the overall fidelity of quantum operations. This redundancy ensures that the calculation can continue correctly even if decoherence or errors occur in one quantum computing unit, as errors are detected and corrected throughout the quantum network.

[0089] The combined efforts of multiple quantum computing units therefore not only increase the computing power, but also the reliability and scale of quantum computing applications. This promises major advances in areas such as cryptography, materials science, and Al itself, where complex, multi-dimensional problems are at the heart of progress.

[0090] Further described is a method performed by a hybrid quantum-classical computing system for processing artificial intelligence (Al) applications. The hybrid quantum-classical computing system comprising a classical computing unit, and at least a first and a second quantum computing unit, each of which is communicatively connected to the classical computing unit, wherein the first quantum computing unit is configured to compute a first algorithmic component, and the second quantum computing unit is configured to compute a second algorithmic component, wherein the classical computing unit includes a control module. The method comprises the following steps performed by the control module. Determining whether an Al application at least in part being processed by the classical computing unit requires computation of the first and or the second algorithmic component. If the Al application at least in part being processed by the classical computing unit requires the computation of the first and or the second algorithmic component, determining, based on a predetermined decision logic, whether the first and or the second algorithmic component is to be computed by the respective first and or second quantum computing unit. If the control module has determined that the first and or the second algorithmic component is to be computed by the respective first and or second quantum computing unit, controlling the first and or the second quantum computing unit to compute the respective first and or second algorithmic component.

[0091] The method preferably starts with the classical computing unit processing the Al application. It plays a central role in managing various tasks such as pre-processing data, scheduling, connecting to external systems and classical algorithmic processing. In the presented method, it also houses a control module. The control module is the element on which the decision-making ability of the hybrid system rests. Its basic task is to determine the processing requirements of each Al application, specifically identifying the need for one or both specialized quantum algorithm components.

[0092] Once the control module determines that such quantum computation elements are needed, it performs a decision-making process driven by a predetermined logic. This decision logic is based on factors such as the computational complexity of the task, the availability of quantum resources or the urgency and priority of the computations. Essentially, this logic ensures that quantum computing resources are used efficiently and effectively, as well as on those factors mentioned above regarding the system.

[0093] If, after consideration, the control module determines that one or both quantum units should process their respective algorithmic components, it initiates the action by sending the appropriate data and instructions to the designated quantum units. Given that quantum resources are currently scarce and expensive, the ability to selectively use quantum units as needed maximizes the efficiency of the system. By selecting which tasks are offloaded to quantum processors, it is ensured that the quantum advantage is used for useful purposes.

[0094] The architecture is inherently scalable - the system can be expanded by adding more quantum computing units as they become available and economically feasible. It also allows flexibility in handling a variety of Al applications, each with different quantum processing requirements.

[0095] By clearly delineating roles and allowing seamless interaction between the different types of computing units, this method exploits the synergy between classical and quantum computing. The classical unit can take on tasks appropriate to its architecture, such as error correction, task queuing and general orchestration, while the quantum units can be reserved for computations that require their unique capabilities. The control module's decision logic can be fine-tuned and adaptive, potentially integrating with machine learning algorithms to optimize decisions over time based on performance metrics and experience.

[0096] Quantum computing can provide acceleration for certain Al components such as optimization problems or sampling tasks that occur in certain algorithms such as Boltzmann machines or quantum-assisted machine learning models. Since each quantum unit is responsible for computing specific components, the system can utilize this quantum speedup more effectively. The hybrid system can solve a larger class of problems than pure classical or pure quantum systems, increasing the problem-solving capacity and potential scope of Al applications.

[0097] Further improvements are described above in connection with the system and should be applied to the method accordingly.

[0098] The hybrid quantum-classical computing system and method as described is a highly adaptable and resource-efficient approach that could significantly improve computational capabilities for Al applications. This method benefits from the power of quantum computing while mitigating its current limitations by supporting classical infrastructure. It thus paves the way for the practical implementation of advanced computing phenomena and their applications in our increasingly data-driven world.

[0099] Brief Description of the Drawings

[0100] Fig. 1 illustrates a schematic of a hybrid quantum-classical computing system according to the present invention.

[0101] Fig. 2 shows a flow-chart describing a method performed by a hybrid quantum-classical computing system for processing artificial intelligence (Al) applications.

[0102] Detailed Description of the Invention

[0103] Fig. 1 illustrates a schematic of a hybrid quantum-classical computing system 100. This system incorporates a classical computing unit 110 and several quantum computing units, specifically a first quantum computing unit 120, a second quantum computing unit 122, and potentially additional units up to an N-th quantum computing unit 124. These units are structured to work together to process artificial intelligence Al applications, with each quantum computing unit optimized for specific algorithmic components.

[0104] The classical computing unit 110 could feasibly be comprised of standard computing hardware and software components, such as CPUs, GPUs, FPGAs, ASICs, or a combination thereof. This unit is responsible for general-purpose processing tasks and for orchestrating the operation of the entire system. The first quantum computing unit 120 is tailor-made to compute a first algorithmic component 126. This component could be optimized for tasks that are well-suited to quantum computation, such as the quantum Fourier transform or Grover's algorithm. The first quantum computing unit 120 could utilize various types of quantum computing technologies, such as superconducting qubits or trapped ion systems. An algorithmic component is a self-contained element with defined interfaces that can be used and reused across different parts of an application and may also be used for different applications.

[0105] The Quantum Fourier Transform operates with a complexity of O(N2), while its classical counterpart, the Fast Fourier Transform (FFT), computes the discrete Fourier transform with a time complexity of O(n2n). This represents an exponential acceleration in processing speed. Grover's algorithm can locate a specific item within an unsorted collection of N items with a time complexity of o(V Q, effectively providing a quadratic improvement over traditional exhaustive search method. To achieve the highest probability of success, only one iteration of the algorithm is typically necessary. In practical applications, after performing the algorithm, measuring the quantum register yields the desired item with a significant likelihood. However, to ascertain the most probable result, the algorithm may be executed multiple times to analyze the outcomes statistically.

[0106] The second quantum computing unit 122 is equipped to compute a second algorithmic component 127, potentially different from the first in terms of its computational approach and suitability for certain Al tasks. This specialization allows both units to be fine-tuned for performance improvements in different aspects of Al processing.

[0107] If the hybrid system supports scalability, the Nthquantum computing unit 124 represents the potential for additional specialized quantum processors within the system, each with a designated algorithmic component 128 that expands the system's computational capabilities.

[0108] The communication between the classical computing unit 110 and the quantum computing units 120, 122, 124 is facilitated by communication interfaces 130, 132, 134. These interfaces could include quantum-classical translators, API integrations, or other interface technologies that bridge the computational models of quantum and classical technologies.

[0109] The classical computing unit 110 includes a control module 140, which is a core component responsible for determining the computational needs 142 of an Al application 144. The control module 140 includes a decision logic 150 that assesses the requirements of the Al application and decides whether to employ quantum computing resources. The decision logic 150 may consider factors like task complexity, available quantum resources, the efficacy of quantum versus classical processing for specific tasks, and cost considerations.

[0110] The decision logic 150 could be implemented through various strategies, such as rule-based systems, machine learning models, or priority schemas. Moreover, it would ensure that the delegation of tasks between classical and quantum units is optimal, making the most efficient use of the system's resources and capabilities to achieve the desired Al processing outcomes.

[0111] Fig. 2 shows a flow chart 200 describing a method performed by a hybrid quantum-classical computing system for processing artificial intelligence (Al) applications. Particularly, the flow chart 200 outlines a method performed by a control module of a classical computing unit in a hybrid quantum-classical computing system. The flow chart describes the decision-making process for distributing computing tasks between classical and quantum units within the system.

[0112] The method commences at the start point 205 and continues to activity 210 where an initial determination is made. In this activity 210, the system assesses whether the ongoing Al application processed by a classical computing unit requires the computational capabilities of the first and / or second algorithmic components. This decision is based on the computing requirements of the Al application, which could involve analyzing the complexity, the efficiency, and the suitability of the task for quantum computation.

[0113] Following this review, the process moves to a decision block 220. If activity 210 yields a positive result, indicating that the Al application indeed necessitates the computation of one or both algorithmic components, the method proceeds to activity 230; otherwise, the method concludes at the endpoint 240.

[0114] In activity 230, the predetermined decision logic of the system is applied to make an informed choice regarding whether the first and / or the second algorithmic component should be computed by their corresponding quantum computing units. This logic may factor in various criteria, such as the potential for a quantum speedup, the availability and capability of quantum resources, the type of algorithmic problem, and other strategic considerations aimed at optimizing the system's overall performance. Subsequently, the method reaches another decision block 250. If the decision logic in activity 230 confirms that leveraging the quantum computing units is beneficial, the method advances to activity 260, where the appropriate quantum computing units are controlled to execute the respective algorithmic components. If the decision logic determines otherwise, the process ends at the same endpoint 240.

[0115] The described flow chart 200 simplifies the complex decision-making in a hybrid quantum- classical computing system, allowing for an organized and efficient approach to managing computational resources and optimizing Al application processing.

Claims

Claims1. A hybrid quantum-classical computing system (100) for processing artificial intelligence (Al) applications, comprising: a classical computing unit (110), at least a first and a second quantum computing unit (120, 122), each of which is communicatively connected to said classical computing unit (110), wherein said first quantum computing unit (120) is configured to compute a first algorithmic component (126), and said second quantum computing unit (122) is configured to compute a second algorithmic component (127), wherein said classical computing unit (110) includes a control module (140), wherein said control module (140) is configured to determine whether an Al application (144) at least in part being processed by said classical computing unit (110) requires computation of said first and or said second algorithmic component (126, 127), wherein said control module (140) is further configured to determine, based on a predetermined decision logic (150), whether said first and or said second algorithmic component (126, 127) is to be computed by the respective first and or second quantum computing unit (120, 122), if said Al application at least in part being processed by said classical computing unit (110) requires the computation of said first and or said second algorithmic component (126, 127), wherein said control module (140) is further configured to control said first and or said second quantum computing unit (120, 122) to compute the respective first and or second algorithmic component (126, 127), if said control module (140) has determined that said first and or said second algorithmic component (126, 127) is to be computed by the respective first and or second quantum computing unit (120, 122).

2. The hybrid quantum-classical computing system according to claim 1, wherein said predetermined decision logic (150) bases its result on one or more of the following: an algorithm's suitability, a task's complexity, a quantum speedup potential, an error rate, a qubit's stability, a resource's availability, a problem's encoding, a hybrid algorithm's compatibility, a data transfer overhead, a model's type, an Al application's domain, a scalability, a quantum readiness, a cost consideration.

3. The hybrid quantum-classical computing system according to one of the preceding claims, wherein said control module (140) includes at least one of: a hybrid quantum- classical scheduler, a decision tree-based controller, a rule-based inference system, a machine learning-based predictor, an adaptive load balancer, an SLA-aware allocator, an API-based integration with quantum providers, a cost-performance analyser, a priority queue manager, a fallback mechanism.

4. The hybrid quantum-classical computing system according to one of the preceding claims, wherein said first and or said second algorithmic component (126, 127) is accessed by using at least one of: a quantum-capable Application Programming Interface, a Software Development Kit, a function invocation, data translation and encoding, task submission, a call-back.

5. The hybrid quantum-classical computing system according to one of the preceding claims, wherein said classical computing unit (110) includes one or more of: a Central Processing Unit, a CPU Core, a Graphics Processing Unit, a Field Programmable Gate Array, an Application-Specific Integrated Circuit, a Digital Signal Processor, a System on a Chip, an Accelerated Processing Unit, a Tensor Processing Unit.

6. The hybrid quantum-classical computing system according to one of the preceding claims, wherein said first and said second quantum computing unit (120, 122), each include one of: a gate-based quantum computer, a quantum annealer, a topological quantum computer, an adiabatic quantum computer, a quantum simulator, quantum dot-based quantum computers, silicon quantum computers, noisy intermediate-scale quantum computer, a quantum neural network.

7. The hybrid quantum-classical computing system according to claim 6, wherein said gate-based quantum computer is based on one of the following: superconducting qubits, trapped ions, photonic systems, wherein said quantum annealer is based on superconducting qubits, wherein said topological quantum computer is based on Anyons and Majorana fermions.

8. The hybrid quantum-classical computing system according to one of the preceding claims, wherein at least one of said first and said second quantum computing unit (120, 122) is an algorithm-specific quantum computer.

9. The hybrid quantum-classical computing system according to one of the preceding claims, wherein at least one of said first and said second quantum computing unit (120, 12) is optimized for computing the respective algorithmic component (126, 127).

10. The hybrid quantum-classical computing system according to one of the preceding claims, wherein said first and said second algorithmic component (126, 127), each, provide at least in part the computation of one of: quantum Fourier transform, Grover's algorithm, amplitude amplification algorithm, quantum phase estimation, quantum clustering, quantum classification, quantum pattern recognition, adiabatic quantum algorithms, quantum approximate optimization algorithm, Harrow-Hassidim-Lloyd algorithm, quantum-enhanced Monte Carlo method, variational quantum eigensolver.11 . The hybrid quantum-classical computing system according to one of the preceding claims, wherein said Al application (144) at least in part being processed by said classical computing unit (110) implements one or more of: training a model, fine-tuning a model, performing an inference, conducting a validation, running a test, applying transfer learning, utilizing unsupervised learning, engaging in semi-supervised learning, performing reinforcement learning.

12. The hybrid quantum-classical computing system according to claim 11 , wherein a model is one or a combination of: a language model, a vision model, a multimodal model, an audio model, a reinforcement learning model, a robotics model.

13. The hybrid quantum-classical computing system according to one of the preceding claims, wherein said Al application (144) at least in part being processed by said classical computing unit (110) addresses one or more of: healthcare, retail, marketing, manufacturing, climate science, biology, physics, chemistry.

14. The hybrid quantum-classical computing system according to one of the preceding claims, wherein said first and said second quantum computing unit (120, 122) are used to jointly compute a computation task.

15. A method performed by a hybrid quantum-classical computing system for processing artificial intelligence (Al) applications, said hybrid quantum-classical computing system (100) comprising a classical computing unit (110), and at least a first and a second quantum computing unit (120, 122), each of which is communicativelyconnected to said classical computing unit (110), wherein said first quantum computing unit (120) is configured to compute a first algorithmic component (126), and said second quantum computing unit (122) is configured to compute a second algorithmic component (127), wherein said classical computing unit (110) includes a control module (140), the method comprising the following steps performed by said control module (140): determining whether an Al application (144) at least in part being processed by said classical computing unit (110) requires computation of said first and or said second algorithmic component (126, 127), if said Al application (144) at least in part being processed by said classical computing unit (110) requires the computation of said first and or said second algorithmic component (126, 127), determining, based on a predetermined decision logic (150), whether said first and or said second algorithmic component (126, 127) is to be computed by the respective first and or second quantum computing unit (120, 122), and if said control module (140) has determined that said first and or said second algorithmic component (126, 127) is to be computed by the respective first and or second quantum computing unit (120, 122), controlling said first and or said second quantum computing unit (120, 122) to compute the respective first and or second algorithmic component (126, 127).

Images