What is AAAI 2023?

AAAI 2023​ ​was held in Washington DC, USA for two weeks from February​ ​7th to 14th, 2023. Due to the influence of COVID-19, this year's event was held on-site and online.

AAAI (Association for the Advancement of Artificial Intelligence) is the premier scientific society in the field of artificial intelligence, established for the purpose of deepening scientific understanding of the mechanisms underlying thought and intellectual behavior.

​AAAI 2023 (AAAI​ ​Conference of Artificial Intelligence) held this time is a conference aimed at promoting theoretical and applied AI research and intellectual exchange among researchers and engineers. This year was the 37th time.

At AAAI2023, major companies such as SONY, IBM, and Amazon science also co-sponsored as sponsors. In addition, research presentations and lectures were given by world-class research institutes such as Harvard University.

This time, we will introduce trends and special lectures in the artificial intelligence field seen at AAAI2023.

AAAI2023 trends

computer vision and language

At AAAI2023, attention was focused on computer vision and natural language processing multimodal (combination of multiple information in different formats) methods and implementations in the field of image recognition.

Examples of the combination of computer vision and language include answering visual questions using natural language processing, generating captions for images, and language-based pre-training for image recognition tasks.

Among them, "Answers using natural language processing for visual questions" is called "VQA (Visual Question Answering)" and has been attracting particular attention in recent years. In VQA, image recognition processing such as convolutional networks for input images and sentence generation using natural language processing are performed as outputs.

​OpenAI 's Chat GPT, which has been making headlines recently, is already using these multimodal technologies. The large-scale language model `` GPT-4'' announced on March​ ​14, 2023 (local time) made it possible to use images as input data and perform language output for them. In addition, since "GPT-4" is already available in Chat GPT 's paid plan "ChatGPT Plus", the experience of general users will progress from now on.

reinforcement learning

This is another area that has received a lot of attention. At AAAI2023, there were many presentations on Multi-Agent Reinforcement Learning (MARL), which has been a hot topic in recent years.

Conventional reinforcement learning is based on one agent finding the optimal policy while maximizing the reward. On the other hand, MARL is said to find the optimal policy through cooperation, competition, or a combination of these relationships among multiple agents. At AAAI2023, cooperative MARL was particularly noted.

Attention was also focused on offline reinforcement learning related to data handling.

Offline reinforcement learning is a method of processing data collected in the past and proceeding with learning while rewarding policies. By the way, online reinforcement learning is a method that continuously acquires data and performs real-time learning.

At first glance, the immediacy of online learning seems superior. However, since offline learning performs simple learning with data that has been acquired in the past, that is, already established data, it has the advantages of stability and ease of model implementation.

From the two elements of MARL and offline reinforcement learning, we can expect further technological development of reinforcement learning and continued efforts to devise ways to implement it in society.

AI security and safety

In the AI security field, not only technical aspects but also various issues surrounding AI are being discussed. For example, issues such as the explainability of AI, the handling of personal information, and legislation regarding AI.

"SimFair: A Unified Framework for Fairness-Aware Multi-Label Classification", published by the research organization of Purdue University, which is famous in the United States, describes the fairness of decision-making derived by machine learning algorithms, and is an excellent paper that deserves the AAAI 2023 paper award. It was selected for an award (AAAI-23 Distinguished Paper).

In addition, presentations were made on the security and safety of AI from various perspectives, such as label diversity and bias in learning data.

Special lecture

​The Learning to See the Human Way talk by Massachusetts Institute of Technology Professor Josh Tenenbaum​ ​presented a presentation on embedding humanity into machines and artificial intelligence.

In recent years, a goal among researchers and scholars in the field of artificial intelligence has been to develop deep learning and machine learning models as if humans were thinking.

Among the artificial intelligence technologies that have been established so far, object recognition models, for example, learn from human-labeled image data. I have a difficult situation.

Machines also do not understand object relationships and concepts, and learn and reason about patterns in data, leading to false positives and misclassifications. On the other hand, humans can recognize objects by judging objects and their situations from vast amounts of data such as experience and knowledge.

This is a work that can be achieved by humans' intuitive senses that connect cognition, logic, language, and planned thinking, and machines have not yet achieved this. The human brain begins to develop from 2 to 3 months after birth, and the above-mentioned cognition and logical thinking begin to work little by little before language comprehension.

For these reasons, Professor Tenenbaum said, "Understanding the basis of human learning from the knowledge of developmental psychology and neuroscience, and working on human learning algorithms (which are more sophisticated than we imagine). Making full use of probabilistic programming, program synthesis, game engine tools, and simulators for this purpose should be a breakthrough for developing machines and artificial intelligence in a way that humans think."

Summary

This time, I introduced AAAI, a world-famous academic society for artificial intelligence.

Every year at this conference, attention is paid to presentations related to new methods and implementations of deep learning and machine learning. In 2023, “computer vision and language,” “reinforcement learning,” and “AI security and safety” were the most prominent topics.

The development of "computer vision and language" seems to be imminent because the possibilities of artificial intelligence are expanding dramatically through multimodal methods. Also, in "reinforcement learning", I found that the learning method and data handling are devised from the viewpoint of implementation.

"Security and safety of AI", which should be taken into account before the utilization of artificial intelligence advances, is one of the topics that will continue to be discussed, but this time the paper on the fairness of algorithmic decision-making was selected as an excellent paper. The high degree of attention from academic societies can also be seen from the fact that

I would like to pay attention to what topics will be discussed at next year's AAAI, and to learn about the "current state of artificial intelligence".