PaperlayerView exact revision on arXiv

Source-linked AI summary

Intelligent Computing: The Latest Advances, Challenges and Future

Shiqiang Zhu, Ting Yu, Tao Xu, Hongyang Chen, Schahram Dustdar, Sylvain Gigan, Deniz Gunduz, Ekram Hossain, Yaochu Jin, Feng Lin, Bo Liu, Zhiguo Wan, Ji Zhang, Zhifeng Zhao, Wentao Zhu, Zuoning Chen, Tariq Durrani, Huaimin Wang, Jiangxing Wu, Tongyi Zhang, Yunhe Pan

arXiv:2211.11281v1cs.AI

TL;DR

Intelligent computing lacks a universally accepted definition while facing challenges in AI capability and escalating computational requirements. This paper proposes a human–physical–information-space framework and surveys the field, including its theories, systems, applications, challenges, and future directions. Its introduction illustrates the potential of collaborative computing with Folding@home’s 2.5 Exaflops achieved by 400,000 volunteers in three weeks.

  • Problem

    Intelligent computing lacks a universally accepted definition, while AI remains limited in interpretability, generality, evolvability, and autonomy.

  • Method

    The paper proposes a definition and unified framework for intelligent computing, then comprehensively surveys its theories, technological fusion, applications, challenges, and future perspectives.

  • Results

    2.5 Exaflops were achieved by Folding@home by combining 400,000 computing volunteers in three weeks, exceeding any supercomputer at that time.

  • Takeaways & Limitations

    Intelligent computing broadens computing beyond data processing toward perceptual, cognitive, autonomous, and human-computer fusion intelligence.

  • Takeaways & Limitations

    Current AI technologies generally work weakly compared with human intelligence and perform well mainly in specific areas or tasks.

Abstract

from arXiv · show

Computing is a critical driving force in the development of human civilization. In recent years, we have witnessed the emergence of intelligent computing, a new computing paradigm that is reshaping traditional computing and promoting digital revolution in the era of big data, artificial intelligence and internet-of-things with new computing theories, architectures, methods, systems, and applications. Intelligent computing has greatly broadened the scope of computing, extending it from traditional computing on data to increasingly diverse computing paradigms such as perceptual intelligence, cognitive intelligence, autonomous intelligence, and human-computer fusion intelligence. Intelligence and computing have undergone paths of different evolution and development for a long time but have become increasingly intertwined in recent years: intelligent computing is not only intelligence-oriented but also intelligence-driven. Such cross-fertilization has prompted the emergence and rapid advancement of intelligent computing. Intelligent computing is still in its infancy and an abundance of innovations in the theories, systems, and applications of intelligent computing are expected to occur soon. We present the first comprehensive survey of literature on intelligent computing, covering its theory fundamentals, the technological fusion of intelligence and computing, important applications, challenges, and future perspectives. We believe that this survey is highly timely and will provide a comprehensive reference and cast valuable insights into intelligent computing for academic and industrial researchers and practitioners.

1 Introduction

Intelligent computing is presented as a new paradigm integrating intelligence with computing to address complex computational needs across human, physical, and information spaces. The paper defines the field, surveys its foundations and applications, and discusses its challenges and future directions.

  • Computing architectures: 2.5 Exaflops were achieved by Folding@home by combining 400,000 computing volunteers in three weeks, exceeding any supercomputer at that time.The paper presents this as an example of horizontal computing collaboration.
  • Challenges: Current challenges include limited interpretability, generality, evolvability, and autonomy in AI, alongside rapidly growing computational demands and hardware–algorithm mismatch.The paper also identifies challenges in moving from data-based intelligence toward perceptual, cognitive, autonomous, and human-machine fusion intelligence.
  • Definition and scope: Intelligent computing addresses complex scientific and societal problems through the fusion of human society, physical, and information spaces.The paper frames this fusion as the basis for its new definition of intelligent computing.
  • Definition and scope: The field encompasses computing theories, architectures, and technical capabilities that support interconnection across the digital civilization.Its stated target is to support computational tasks according to specific needs by matching computing power, algorithms, and desired results.
  • Computing architectures: Intelligent computing dynamically coordinates storage, communication, and computation across edge, cloud, and supercomputing domains.It supports cross-domain systems for cloud collaboration, inter-cloud collaboration, and supercomputing interconnection.
  • Paper contributions: The paper presents the first comprehensive literature survey covering intelligent computing fundamentals, intelligence–computing fusion, applications, challenges, and future perspectives.It aims to provide a reference and insights for academic and industrial researchers and practitioners.

2 Fundamentals of Intelligent Computing

Intelligent computing is presented as a broad paradigm combining human intelligence, machine capabilities, and the physical world to support interconnected systems and complex problems. It distinguishes natural or meta intelligence from machine or generic intelligence and organizes machine intelligence into complementary levels and forms.

  • Concept and theoretical basis: Intelligent computing comprises theoretical methods, architectural systems, and technical capabilities that support interconnection among things and address complex scientific and social problems.Its basic elements are human intelligence, machine capabilities, and the physical world.
  • Meta intelligence: Meta intelligence is natural intelligence carried by carbon-based life, including human comprehension, expression, abstraction, inference, creation, and reflection.The paper treats human wisdom as the source of intelligence in intelligent computing.
  • Forms of intelligence: The paper describes biological embodied, brain, and swarm intelligence as major forms of natural intelligence, with swarm algorithms modeling aggregation, coordination, adaptation, and evolutionary search.Swarm optimization iteratively replaces less feasible solutions with better ones.
  • Generic intelligence: Generic intelligence is machine intelligence carried by silicon-based facilities and expressed through computers’ ability to solve complex problems.Examples include language, image, and speech recognition and target detection.
  • Forms of intelligence: Machine intelligence is organized into data, perceptual, cognitive, and autonomous intelligence, which cooperate on complex tasks.These levels cover data storage and calculation, sensory acquisition, reasoning and explanation, and self-driven consciousness.
  • Machine intelligence: Data intelligence combines symbolic and numerical computation, fuzzy logic, probabilistic methods, neural networks, and evolutionary computation to realize intelligence from data.Perceptual intelligence maps physical-world signals into digital representations through sensors and structured multimodal data.

2.2 Computational Capabilities

The section surveys computational capabilities needed for intelligent computing, emphasizing architectural, chip, storage, biological, and wide-area approaches. It presents these approaches as responses to growing model complexity, physical limits, storage bottlenecks, and geographically distributed tasks.

  • Computational challenges: Intelligent computing faces big scenes, big data, big problems, and ubiquitous requirements, while increasingly complex models demand supercomputing resources.Computing resources can become a barrier to research and may contribute to technological monopolies.
  • Computational challenges: A complete response to computing-power constraints requires optimization across architecture, acceleration modules, integration modes, and software stacks.Algorithm optimization alone cannot fundamentally solve the computing-power problem.
  • Storage and memory: Processing-in-memory addresses the storage wall by reducing processor–memory data movement, which otherwise lowers efficiency and limits bandwidth.The paper identifies memory computing as an effective measure for improving overall computing efficiency.
  • Computing units: Vertical lifting and horizontal expansion provide complementary routes for increasing computing power beyond traditional chip architectures.Examples of new approaches include integrated photonics, quantum computing, and biocomputing.
  • Heterogeneous integration: Heterogeneous integration combines different structures, materials, or computing units so each unit performs suitable tasks within hybrid architectures.It includes chip-level and system-level integration, as well as 2D/3D packaging and Chiplet technologies.
  • Heterogeneous integration: Biological components offer potential advantages in storage capacity, computational parallelism, and ultra-low power consumption when integrated with silicon chips.The paper expects carbon–silicon integration to improve computing power, storage density, and energy efficiency.
  • Wide-area collaboration: Wide-area collaborative computing connects HPC, cloud, fog, and edge resources to support distributed, real-time acquisition, processing, and intelligent analysis.Its construction still involves challenges in cross-domain resource and task matching, scheduling, collaboration, security, and reliability.

2.3 Features of Intelligent Computing

Intelligent computing is characterized by self-learning and evolvability, high capability and energy efficiency, security and reliability, automation and precision, and collaboration and ubiquity. Its two complementary paradigms use intelligence to improve computing and computing to support intelligence.

  • Self-learning and evolvability: Self-learning extracts experience, rules, and knowledge from massive data, while evolvability provides heuristic self-optimization inspired by biological evolution.Neuromorphic and biological computing are cited as techniques for moving beyond traditional von Neumann principles.
  • Capability and efficiency: Intelligent computing pursues high computing capability and energy efficiency through processing-in-memory, heterogeneous integration, and wide-area collaboration.The goal is efficient processing of large-scale, complex, and sparse big data while limiting energy consumption.
  • Security and reliability: Security and reliability require trusted identity, data, processes, and environments together with protection for network, storage, content, and circulation.The approach spans trusted hardware, operating systems, software, networks, and privacy computing.
  • Automation and precision: Automation and precision make computing task-oriented by matching resources, reconstructing systems, and managing services and task lifecycles automatically.The architecture adapts continuously to task execution, with directed coupling across software and hardware.
  • Collaboration and ubiquity: Collaboration and ubiquity integrate physical, information, and social spaces through heterogeneous perception, complementary resources, and human–machine cooperation.Ubiquitous computing combines theoretical methods, architectures, and technical approaches so computing can occur everywhere.
  • Fusion of intelligence and computation: Computing by intelligence uses intelligent approaches to improve computing capability and efficiency, whereas computing for intelligence supplies technologies that advance computer intelligence.The two paradigms jointly motivate new mechanisms, architectures, and systems.
  • Fusion of intelligence and computation: Intelligent computing supports adaptive processing of unstructured scenes and data through task understanding, decomposition, solving, and resource allocation.This process is described as transparent computing and can produce ubiquitous, transparent, automatic, real-time, and secure services.
  • Fusion of intelligence and computation: New computing mechanisms configure hardware resources at different granularities and support autonomous learning and evolutionary iteration during intelligent processes.New hardware and software refactoring and cooperative evolution are among the mechanisms discussed.

3 Computing by Intelligence

The paper organizes intelligent computing around four intelligent abilities and integrated intelligence. For each ability, it reviews significant progress in representative research areas.

  • Integrated intelligence: The section describes four intelligent abilities of computers and the mode through which their intelligence is integrated.It presents progress in typical research areas for each ability.

3.1 Data Intelligence

Data intelligence broadens computing toward greater universality and intellectual capability by supporting diverse data models, intelligent algorithms, and brain-inspired methods.

  • Foundations: Intelligent computing seeks both universal computation across scenarios and higher intellectual capability through methods inspired by intelligent creatures.The section identifies analog and graph problems as requiring varied computation and lists neural networks, fuzzy systems, and evolutionary computing as classical intelligence methods.
  • Analog Computing: Analog computing supports real-time, parallel-valued operation with simple hardware and lower bandwidth consumption, but its results are environment-sensitive and poorly transportable.Analog computers generally solve preset problem types, and environmental factors can make results difficult to obtain.
  • Graph Computing: Graph computing uses graph structures to represent relationships and supports search, traversal, analysis, storage, management, and large-scale processing.Graph processing increasingly combines distributed, parallel, stream, and incremental computing, with node-centric engines designed for parallel graph tasks.
  • Graph Computing: Expanded graph models incorporate attributes, labels, probability, and hierarchy, increasing modeling flexibility while also increasing algorithmic complexity.Graph databases provide flexible expression and broad application scenarios, while richer model semantics make graph-processing results more suitable for natural applications.
  • Graph Computing: Graph neural networks adapt deep-learning frameworks to structured and semi-structured graph data for classification, prediction, and anomaly detection.Examples include graph convolutional, recurrent, attention, and residual networks, with mathematical models improved for graph structure.
  • Artificial Neural Networks: Artificial neural networks model neuron-like signal transmission and support applications including classification, prediction, and natural language processing.The section describes back propagation, multilayer feed-forward networks, convolutional and recurrent models, and pre-training with transfer learning.

3.2 Perceptual Intelligence

Perceptual intelligence enables machines to interpret sensory data and surroundings, but unrestricted real-world perception remains difficult. Smart sensors combine sensing, computation, and communication for monitoring and prediction.

  • Scope and Challenge: Perceptual intelligence begins with sensing and interpreting surroundings, extending beyond constrained industrial settings toward free real-world perception.Unlimited situations and unexpected events make fully autonomous robotic handling challenging.
  • Sensing Modalities: Intelligent sensing spans human-like senses and measurements such as temperature, pressure, humidity, height, speed, and gravity.Its performance may require substantial computation or data training.
  • Recognition: Machines have made major advances in voice, visual, and touch recognition through pattern recognition and deep learning.The passage states that machines have recently become more perceptually intelligent than humans in these capabilities.
  • Smart Sensors: Smart sensors integrate computers, IoT connectivity, signal conditioning, embedded algorithms, and digital interfaces to process live data.They transform live measurements into digital data that can be sent to a gateway and support functions such as process control and quality evaluation.
  • Applications: Sensor data supports process optimization, predictive maintenance, supply management, and resource coordination.Temperature and pressure measurements may be combined to predict the beginning of a mechanical breakdown.

3.3 Cognitive Intelligence

Cognitive intelligence concerns machines’ abilities to understand complex facts, reason over knowledge, and support decisions. The section emphasizes language processing, causal inference, and knowledge reasoning as core areas.

  • Cognitive Intelligence: Cognitive intelligence enables machines to understand, summarize, interpret, plan, and actively apply knowledge to complex real-world situations.It differs from perceptual intelligence by relating data elements, analyzing structured-data logic, and responding from distilled knowledge.
  • Natural Language Processing: Natural language processing comprises natural language understanding and generation, covering lexical, syntactic, semantic, and text-generation tasks.NLU includes word and structure understanding, while NLG produces text from raw text, data, and images.
  • Natural Language Processing: Lexical, syntactic, and semantic analysis address word meanings, sentence structure, and discourse-level interpretation.Named entity recognition, word-sense disambiguation, semantic-role labeling, and coreference resolution are among the described tasks.
  • Causal Inference: Associative machine-learning models have poor interpretability, motivating causal inference to distinguish causal relationships from associations.The text presents causal inference as a way to address machines’ difficulty distinguishing true or false causal associations in data.
  • Causal Inference: Causal inference includes association, intervention, and counterfactual layers, with counterfactual reasoning asking what would differ under alternative past actions.The potential-outcomes framework compares effects for the same subject with and without intervention, while structural causal models represent multiple-variable causality.

3.4 Autonomous Intelligence

Autonomous intelligence aims to move computing from passive task-specific learning toward active interaction, transfer, rapid adaptation, and self-evolution. Transfer learning, meta-learning, and brain-like architectures are presented as possible paths.

  • Autonomous Intelligence: Autonomous intelligence requires generalized models and continuous environmental interaction, progressing from single-task learning toward active learning and self-evolution.The section frames transfer learning, meta-learning, and autonomous learning as feasible development paths.
  • Transfer Learning: Transfer learning maps knowledge from a usually data-rich source domain to a smaller target domain and fine-tunes models with limited target data.Its original intention includes reducing manual labeling by transferring from labeled source data to unlabeled target data.
  • Transfer Learning: Instance-, feature-, model-, and relationship-based transfer learning address transfer through selected examples, aligned features, reused models, or cross-domain relationships.Feature-based methods align source and target distributions, while model-based methods reuse source-trained parameters through fine-tuning or fixed feature extractors.
  • Meta-Learning: Meta-learning targets rapid adaptation to new tasks by learning meta-knowledge and strong initial parameters from a set of tasks.It is mainly applied to few-shot, zero-shot, and other settings with little available data.
  • Meta-Learning: A few iterations can produce desirable results on a new task from meta-learned parameters.The section links meta-learning to reduced model-design cost and improved handling of new tasks with few samples.
  • Limitations: Current transfer-learning approaches are limited to homogeneous tasks with aligned support and query-set sizes.This limitation constrains their transferability across task types and task-set configurations.
  • Autonomous Architectures: LeCun’s autonomous intelligence framework organizes configuration, perception, world-model, cost, and action modules to support prediction, reasoning, and action.The framework is described as learning a model of the world in a self-supervised manner and incorporating common sense and emotion modularly.

3.5 Man-machine Integrated Intelligence

Man-machine integrated intelligence combines human cognitive abilities with machine intelligence through interaction, collaboration, and increasingly natural interfaces. The approach supports complex, dynamic tasks but faces real-time, adaptability, stability, and command-capacity challenges.

  • Human-computer interaction: Multimodal user interfaces use speech, handwriting, posture, sight, expression, touch, smell, taste, and other channels for parallel interaction.This development aims to move interaction toward more natural, human-centered communication.
  • Human-computer interaction: Natural interaction frees people from traditional interaction methods and enables more harmonious human-computer interaction.Graphical user interfaces also reduced typing operations, supported ordinary users, and expanded the user population.
  • Human-machine fusion intelligence: Human-machine fusion intelligence combines sensor-collected objective data, human subjective information, human cognition, and computer computing power.It uses human domain knowledge as learning clues to support computer-based decision-making and complex professional tasks.
  • Human-machine integration: Human-machine integration uses interactive and collaborative learning, intuitive communication, expert feedback, and adaptation to dynamic environments.Its integrated intelligence can continuously evolve as human knowledge, tasks, and data change.
  • Human-machine symbiosis: Human-machine symbiosis embeds hardware such as sensors and wearables into everyday settings while connecting software, immersive environments, and distributed interaction.Proposed applications include remote exploration, complex-system operation, co-driving, and social-problem research and governance.
  • Brain-computer interfaces: Single-modal brain-computer interfaces have limited task capacity, declining classification accuracy with more instructions, poor long-term robustness, and limited stability.Hybrid or multimodal brain-computer interfaces combine a unimodal BCI with another BCI or non-BCI system to support multi-instruction, real-time control.

4 Computing for Intelligence

Intelligent computing is increasingly constrained by the rapidly growing computational requirements of AI models, making scalable and energy-efficient platforms important. The section surveys this growth, its bottleneck implications, and distributed approaches such as federated learning.

  • Computing growth: By 2020, the biggest model required six million times as much computing power as the largest model in 2018.This increase illustrates the scale of computational expansion reported for prominent AI models.
  • Computing growth: The computing power required to develop a breakthrough model grew at about the pace of Moore’s law before 2012.Moore’s law describes computational capacity on a single microchip tending to double every two years.
  • Computing growth: Computational requirements for top AI models doubled every 3.4 months between 2012 and 2018 after AlexNet renewed interest in deep learning.The same 3.4-month doubling period extended from AlexNet through GPT-3 despite algorithmic and architectural advances.
  • Challenges: Computing power is becoming a bottleneck for intelligent computing, while AI/ML platform energy efficiency is increasingly important for reducing cost.The cited trend is expected to slow because its rapid pace cannot be maintained indefinitely, and a slowdown may already be underway.
  • Distributed learning: Distributed machine learning reduces the computational load on a single server, and federated learning supports distributed training while preserving server-side data privacy.Federated learning also avoids transmitting large volumes of data from distributed locations to a central server.

4.1 Large Computing Systems

Large computing systems combine clustered, cloud, HPC, and edge resources to address growing scientific, AI, and IoT workloads. These systems improve scalability and latency, but heterogeneous infrastructure and local-resource limits create operational boundaries.

  • High-performance computing: High-performance computing networks many computers into a cluster, enabling faster processing of massive data volumes and scalable capacity through cloud computing.Cloud computing abstracts computation, storage, and network infrastructure to support application deployment and scalability.
  • Scientific computing: Upgraded scientific instruments are expected to increase sensitivity and resolution 10-100 times in the next decade, requiring corresponding storage and processing capacity.This growth challenges conventional operating models centered primarily on HPC in data centers.
  • HPC and AI: The convergence of HPC and AI has produced new approaches to existing problems and enabled new applications.HPC-based AI platforms encapsulate heterogeneous infrastructure and provide researchers with a consistent environment.
  • AI platforms: AI platforms span commercial cloud services, discipline-specific systems, and supercomputer-based resource aggregation for scientific and industrial workflows.Examples include drug development, thresholdless deep learning, cloud AI resources, and unified interfaces for diverse hardware.
  • Edge, fog, and cloud computing: Cloud computing may not adequately meet applications requiring local computation, while edge servers have less computational power and local rather than global information.Rising endpoint counts can increase the load on edge servers, making edge computing non-optimal for global decision-making in the cited cases.
  • Edge, fog, and cloud computing: Edge computing processes data locally or near devices and sends only crucial data centrally, thereby reducing latency for end-user devices.Fog computing places distributed resources between edge devices and the cloud, supporting local analysis as an intermediary layer.

4.2 Emerging Computing Architectures

Emerging computing architectures address energy, latency, memory-transfer, and performance constraints through specialization, accelerators, and computing-in-memory designs. These approaches offer efficiency gains but require careful hardware-software design, verification, and application-specific choices.

  • Architectural specialization: Architectural innovation targets more efficient energy management, lower power consumption, lower chip cost, and faster error detection and correction.AI accelerators can reduce training and execution time for operations that CPUs cannot perform efficiently.
  • Architectural specialization: Architectural specialization is presented as the only viable hardware option for the next decade absent a practical post-CMOS alternative.The slowing of Moore’s law makes specialization a practical and affordable substitute for full universal computing, with consequences for algorithms.
  • Accelerators: Accelerators are described as effective tools for continued performance gains, but their design should be driven by clearly defined use cases.The surveyed accelerator comparison plots peak power against peak giga-operations per second on logarithmic axes and distinguishes precision, form factor, and workload type.
  • Computing-in-memory: Computing-in-memory allows memory cells to execute elementary logic operations independently of a central processor, addressing the widening memory-processor performance gap.CIM separates neither data movement nor computation into wholly distinct operations within the memory macro.
  • Challenges: CIM performance improvement requires substantial synthesis and creates obstacles for application translation and framework verification.Memristor implementations also face geographical and temporal device variations, performance limitations, and an absence of reliable simulations.
  • Memory technologies: SRAM-CIM offers faster writes, lower write energy, and nearly infinite endurance for low- to medium-capacity systems, while nvCIM retains weights during inactivity.nvCIM requires sufficient memory for all application data because of low durability and high write energy; SRAM-CIM can reduce latency and improve power efficiency.
  • Computing-in-memory: CIM includes near-memory array computing, which stores data in conventional memory cells with an interface connecting circuits to those cells.In-memory array computing and near-memory array computing are the two structures identified in the section.

4.3 Emerging Computing Modes

Emerging computing modes address complexity, scalability, and energy constraints through quantum, neuromorphic, photonic, biological, and graph-computing approaches. These paradigms offer potential advantages, but quantum artificial intelligence remains constrained by limited qubits and environmental noise.

  • Complex AI models remain difficult to deploy on edge devices because power and bandwidth constraints cause long processing times and cumbersome architectures.
  • Quantum computing: Quantum computing uses superposition to enable parallel computation and may increase computing power exponentially compared with classical computers.
  • Quantum artificial intelligence: Quantum artificial intelligence combines quantum data or algorithms with AI, including classical learning for quantum systems and quantum hardware for accelerating learning.
  • Quantum applications: Quantum AI applications include drug synthesis, chemical-reaction treatment, and graph search using Gaussian Boson sampling to identify large fully connected subgraphs.
  • Quantum limitations: Quantum AI is still at an initial stage, with applications limited by the number of quantum bits and error rates caused by environmental noise.
  • Photonic computing: Optical neural networks can perform ten trillion operations per second with energy consumption as low as one photon per operation.
  • Biocomputing: Biological computing offers strong parallel and distributed processing, low power consumption, and more than 1000 DNA calculations per joule versus more than 100 for silicon computers.

5 Applications of Intelligent Computing

Intelligent computing is applied across computational materials, materials informatics, and scientific discovery by using machine learning models, surrogates, and data-driven predictions. Reported applications include faster simulations, improved spectral analysis, transport-property studies, and discovery of candidate materials.

  • Computational materials: AI integration with computational materials spans multiple length and time scales, while machine learning addresses computational demands and accuracy limits in first-principles methods.
  • Materials informatics: The review presents AI integration into computational materials as a new research paradigm supporting materials informatics and data-driven property prediction.
  • Molecular dynamics: DeepMD-based molecular dynamics predicted Li-Si alloy properties 20 times faster than ab initio simulations with similar accuracy.
  • Materials simulation: DeepMD applications computed Raman spectra from classical 2-nanosecond trajectories, enhanced low-frequency spectral resolution, and obtained finite-temperature Na+ diffusion coefficients.
  • Surrogate modeling: Machine-learning surrogate models can be evaluated far faster than physical models with nearly the same accuracy, enabling global searches over materials design and optimization spaces.
  • Materials discovery: Transfer learning screened 21,316 perovskites, predicted 9,238 with the desired bandgap, and identified 11 previously undiscovered materials.

6 Perspectives

The paper identifies theoretical challenges in making intelligent computing more human-like, uncertainty-aware, and knowledge-driven. It calls for new computational theories, architectures, and learning strategies while recognizing limitations in current data-driven models.

  • Theoretical revolution: Axiomatic systems, quantitative descriptions, and measurement criteria are needed to decompose, compare, and formalize multivariate intelligence.
  • Theoretical revolution: Intelligent computing lacks a theory that supports uncertainty and maps theoretical computing space to physical space.
  • Knowledge-driven intelligence: Current deep-learning approaches have weak generalization, poor interpretability, limited knowledge expression, deficient common sense, and catastrophic forgetting.
  • Knowledge-driven intelligence: Knowledge-driven intelligence is proposed as a progression beyond data-driven intelligence through human-like perception, representation, reasoning, learning, and decision-making.
  • Knowledge-driven intelligence: Combining inductive abstraction, deductive reasoning, and constrained optimization of physical theorems remains a key challenge for improving machine intelligence.
  • System architectures: Future systems require flexible architectures and improved knowledge models for creating, storing, retrieving, and describing real-world knowledge.

7 Conclusion

Intelligent computing is presented as a transformative direction for computing during the transition toward an intelligent society. The review synthesizes its foundations, technological integration, applications, challenges, and future directions.

  • The transition toward intelligent society is associated with transformative and potentially disruptive changes in computing technologies.
  • Intelligent computing is described as both intelligence-oriented and intelligence-empowered computing.
  • The paradigm aims to provide universal, efficient, secure, autonomous, reliable, and transparent services for large-scale computational tasks.
  • The paper provides a comprehensive review covering theory fundamentals, intelligence–computing fusion, important applications, challenges, and future directions.
Loading 2211.11281v1…