Despite ups and downs, artificial intelligence has made significant advancements in the past few decades and is gradually transforming industries such as finance, healthcare, and transportation. At the forefront of the development and implementation of AI, how should we view the technology, and what innovations will it bring in the next decade? This year's AISummit may have uncovered some answers to your questions.

On August 6, 2022, the global technology conference was held online as scheduled. The main session in the morning featured seven senior technology experts in the industry:

Mr. Cui Kang, Vice President and Chief Editor of 51CTO;

Mr. Dou Zhicheng, Associate Dean of Gaoling School of Artificial Intelligence, Renmin University of China;

Mr. Tian Yuandong, Research Scientist and Senior Manager at Meta AI (FAIR);

Mr. Pan Qinghua, Vice President of iFLYTEK Research Institute;

Mr. Zhou Zongwei, Tech Lead and Manager at Google Core ML Performance Team;

Mr. Xiang Liang, Head of the Machine Learning System of Volcengine;

Mr. Liu Yi, Head of the Ad Generation Algorithm Team of Microsoft Bing.

Their presentations offered an in-depth analysis to thousands of viewers worldwide, which included fascinating topics of prevailing trends in AI technology, the latest research findings in academia, and how cutting-edge technologies will influence various industries.

Artificial intelligence requires a comprehensive and in-depth understanding

In the opening speech of the conference, Mr. Cui Kang expressed his understanding of the essence of AI technology and his expectations for the event. He believed that with the rapid development of AI, it is no longer considered a subset of IT but a general technology. It is imperative that we find new solutions to resolve problems among technology, industry, and developers.

As Mr. Cui stated, AI technology and developers are making the world a better place. He hoped the AISummit would be a bridge to link the gap between technology, industry, and developers. The conference should serve as a venue for seeking industrial implementations and breakthroughs, and providing more career advancement opportunities for developers, thus facilitating information flow within the industry and encouraging the integration of industry, technology, and developers.

Towards a new generation of intelligent search

In the over twenty years since the launch of search engines, neither their forms nor architectures have changed significantly. In recent years, with the development of the Internet and the widespread use of smart devices, society has become more dependent on information, especially complex information; thus the traditional method for finding information has become increasingly ineffective.

During the keynote speech of "The Outlook of Intelligent Search in the Next Generation," Mr. Dou Zhicheng, Associate Dean of Gaoling School of Artificial Intelligence, Renmin University of China, stated that search engine input modes based on keywords and document lists are insufficient to meet the current information access needs. In the future, the search will be conversational, personalized, multimodal, knowledge-rich, and de-indexed, and it will be able to handle multimodal information and return knowledge. In the future, we will gradually shift from the existing architecture pattern to a new one centered on models.

Decision making in the real-world scenarios: challenges, opportunities and solutions

In recent years, deep learning seems to have made significant advances in processing natural data inputs. Nevertheless, the industry has not yet solved the core challenges of using deep neural networks(DNN) to process structured data and having neural network alternatives to replace human heuristics for discrete optimization.

Mr. Tian Yuandong, Research Scientist and Senior Manager at Meta AI (FAIR), delivered a keynote speech on "Decision Making in the Real-World Scenarios: Challenges, Opportunities, and Solutions". He presented a detailed description of how to use reinforcement learning and search methods with neural networks to develop heuristic algorithms for complex optimization problems based on actual working cases provided by the Meta/Facebook AI team.

Technology development trend in artificial intelligence

Over the past few years, the core of AI technology has moved toward multimodal sensor fusion and systems of integrated hardware and software. With vast AI applications, many industries have benefited from quality solutions; however, what challenges remain to be overcome in the development of AI?

In the keynote speech of "Technology Development Trend in Artificial Intelligence", Mr. Pan Qinghua, Vice President of iFLYTEK Research Institute, explained the development, implementation, and existing challenges of perceptual and cognitive intelligence in many industries such as education, healthcare, and transportation. In addition, he proposed the new concept of cognitive intelligence 2.0 and five major trends of AI development in the next five to ten years.

As Mr. Pan stated, "When AI is used to empower industries, complex systems with human-machine collaboration and continuous evolution should be constructed to profoundly alter production manners. Future applications of AI in various industries will certainly be realized through the development of different complex intelligent systems tailored to each industry."

Ten Lessons that Shaped Google's TPU

Machine learning and artificial intelligence applications have become increasingly popular due to the failure of Moore's Law for computer processors over the past few years, leading to explosive growth in demand for computing power. Following this trend, Google has been the first to identify and offer solutions for processors based on specific sectors.

In his keynote speech, "Ten Lessions that Shaped Google's TPU", Mr. Zhou Zongwei, tech lead and manager at Google Core ML Performance team, discussed Google's experiences in the past decade in developing AI chips, deploying AI applications like DNN applications, and gave a review of the development of hardware and computer architectures.

According to Mr. Zhou, developers and users of chips should consider the total amount of resources that chips will consume in the long run, rather than focusing only on their costs and prices. "Manufacturing AI chips is not a difficult task; however, the term of co-evolution between hardware and software of neural network compilers is too long."

The technical evolution of computing power of large-scale machine learning

GPU computing power has dramatically increased since the boom of deep learning in 2014. The advancement of computing power has resulted in significant benefits for business, such as the efficiency enhancement of advertising recommendations, CV, and NLP. In addition, it also opens up the possibility of implementing larger models.

In the keynote speech of "Technical Evolution of Computing Power of Large-Scale Machine Learning", Mr. Xiang Liang, head of the Machine Learning System of Volcengine, outlined the development of AI computing power and related models in the last ten years, as well as Volcengine's exploration in this field. According to Mr. Xiang, as AI computing power continues to improve, the physical bottleneck of chip processing is expected to become more and more apparent, and the enlargement in models may also make it more difficult to implement in the real world.

It is necessary to seek new strategies in the supply of computing power to implement large-scale computing power. Additionally, Mr. Xiang discussed recent research findings of his team in scheduling, training networks, and storage, with the machine learning platform of Volcengine as an example.

Diverse search ad generation and real-time customization

Ad generation requires the use of offline and online systems, as well as various generation models that include pre-training, guided text generation, text quality evaluation, and reinforcement learning while ensuring that online texts are matched in real-time.

During the presentation of "Diverse Search Ad Generation and Real-Time Customization", Mr. Liu Yi, head of Microsoft Bing's Ad Generation Algorithm Team, discussed the ad generation technology of DeepGen, including the basic models, and the technology of diversified generation.

As stated by Mr. Liu, DeepGen aims to provide comprehensive support for the entire search advertising generation process. In practice, Microsoft seeks to improve Revenue Per Mille (RPM) by improving Impression Yield (IY) and Click-Through Rate (CTR), and to maintain the stability of Quick Back Rate (QBR) as much as possible. In closing, Mr. Liu said, "DeepGen has significantly enhanced Bing's advertising system, and we will continue to work in this direction in the future."

AISummit provided an in-depth overview of the most recent research findings in AI technology and opened up various opportunities for the industry to explore future evolution and applications. The conference organizer, 51CTO, hopes that AISummit can become a landmark event for advancing artificial intelligence worldwide.

As Mr. Cui said, AI should show that technology is used for good purposes. AISummit aims to set new goals for the future of artificial intelligence and to demonstrate how this technology and its developers are making the world a better place.

Click on the following link to view the replay of AISummit: ??aisummit.51cto.com??