Recently, the theme meeting of the 2024 World Science, Technology and Development Forum "Artificial Intelligence Governance Innovation Builds an International Trust Foundation for Cultivating a Science and Technology Governance Ecosystem" was held in Beijing. Qiao Hong, Chairman of the World Robot Cooperation Organization and Academician of the Chinese Academy of Sciences, released the 2024 Artificial Intelligence (Artificial Intelligence) at the meeting. AI) Top Ten Frontier Technology Trend Outlook.
"They are full of infinite possibilities and potential. They will not only bring a more convenient and efficient lifestyle, but also promote innovation and development in all walks of life." Qiao Hong said, hoping that this release can guide everyone to think about "how to grasp The development direction of artificial intelligence, how to promote technological innovation and industrial upgrading, and how to ensure the sustainable development of artificial intelligence technology."
These ten cutting-edge technology trends are:
AI common technology
1. Small data and high-quality data
A large amount of invalid data not only consumes computing resources, but also brings challenges to reliable training of models. In this context, the value of small data and high-quality data is increasingly important. Small data pays more attention to the accuracy and relevance of data, essentially reducing the dependence and uncertainty of artificial intelligence algorithms on data and enhancing network reliability. Building diverse data sets can not only support the development of AI with different technical routes on a theoretical basis, but also provide new possibilities for solving the bottleneck problem of general artificial intelligence.
2. Human-machine alignment
Only when the output results of AI are consistent with human values can we ensure that the capabilities and behavior of the AI model are consistent with human intentions. Relying on data and algorithms alone is not enough to achieve human-machine alignment, which means that when designing a reward mechanism, you must not only consider the efficiency, effectiveness, and effectiveness of the task, but also whether the behavior complies with human ethical standards.
3. AI usage boundaries and ethical supervision model
At present, the compliance, security and ethical issues of AI systems have become increasingly prominent, and it is particularly necessary to establish an AI supervision model framework. Its main purpose is to ensure that all AI systems follow established principles during development and use by formulating clear standards and specifications, thereby reducing the risk of overuse of AI without a defined system.
4. Interpretability Model
On the premise of ensuring effectiveness, improving explainability will help reduce the consumption of public resources, enhance users' trust in AI systems, and promote its application in key areas. For example, in the medical and health field, a highly interpretable AI diagnostic system can make it easier for doctors to understand the basis for their judgment and reduce unnecessary examinations and treatment procedures.
Large-scale pre-trained models
5.The law of scale
Large-scale pre-training models based on massive parameters and training data can effectively improve human-computer interaction and reasoning capabilities, and enhance the diversity and richness of tasks that can be completed. At present, the law of scale is still valid, not only reflected in language models, but also verified in many fields such as image processing and speech recognition.
6. Full-modal large model
The full-modal large model can process and understand various types of data input such as text, pictures, audio, data tables, etc., and generate various types of output according to task requirements. For example, the introduction of 3D point cloud data modality, which is usually used to capture three-dimensional spatial information, is particularly important for robot navigation and obstacle avoidance.
7. AI-driven scientific research
Use large models, generative technologies, etc. to improve the efficiency and accuracy of hypothesis proposing, experimental design, data analysis and other stages in scientific research. Scientists can use AI technology for real-time experimental monitoring and adjustment, rapid feedback on experimental results, and dynamic optimization of experimental designs and assumptions.
embodied intelligence
8. Embodied cerebellum model
Traditional large models can assist robots in slow-channel response tasks such as decision-making, task dismantling, and common sense understanding, but they are not suitable for fast-channel response tasks such as robot planning and control with strong real-time and high stability. Embodied intelligence (a further extension of artificial intelligence in the physical world, generally refers to an intelligent system that can perceive, understand and interact with the physical world). The cerebellar model can use integrated learning methods such as multi-model voting, combined with the selection of robot ontology structure and environmental characteristics. Reasonable model control algorithms ensure that robots can complete highly dynamic, high-frequency, and robust planned control actions under the premise of understanding their own ontology constraints, making intelligent robots more capable of meeting the precise operation and real-time control needs of the real world.
9. Physical artificial intelligence system
The physical artificial intelligence system empowers embodied intelligence to physical objects in the physical world, enabling traditional equipment to break through its original functional limitations and achieve a higher level of intelligent operation. Humanoid robots are the ultimate form of physical artificial intelligence systems. They not only have multi-modal perception and understanding capabilities, can naturally interact with humans, but can also make decisions and act autonomously in complex environments, and are expected to be applied to more complex tasks in the future. In a work scenario.
generative artificial intelligence
10. World Simulator
The world simulator can provide an immersive high-simulation experience and bring users a richer and more diverse game world. It can be used in education, entertainment and other fields, and can also create more super digital scenes. In the field of robotics, this technology can also be used to build large-scale, standardized multi-modal robot behavior data sets, improving the capabilities of robot ontology design, simulation training, and algorithm migration.