2023.Nov.07

REPORTS

Report on the 12th B’AI Book Club
Gary Marcus and Ernest Davis (2019) Rebooting AI: Building Artificial Intelligence We Can Trust.

On June 28, 2022, the 12th session of the B’AI Book Club, an online book review event for members of the B’AI Global Forum, was held. In this session, Project Professor Takane Ito introduced the book “Rebooting AI: Building Artificial Intelligence We Can Trust” (2019).

The book highlights the disparity between public perceptions and expectations of AI and its actual capabilities. It aims to help readers understand what current AI cannot do and the reasons behind it, proposing a new strategy. The strategy presented in the book suggests that current AI, which employs deep learning, cannot create general intelligence. Instead, it argues that referencing the functioning of the human brain and mind can lead to a solution.

It is worth noting that in 2020, one of the authors, Gary Marcus, proposed the use of a hybrid approach that combines classical knowledge-based learning models and deep learning in a paper titled “The Next Decade in AI: Four Steps Towards Robust Artificial Intelligence” (https://doi.org/10.48550/arXiv.2002.06177). Referring to this paper can provide a deeper understanding.

The main topics discussed during this book review session included the book’s perspective on deliberate manipulation of training data. While this book views manipulation as problematic, some argue that deliberate manipulation is necessary to eliminate bias. Furthermore, there was a discussion about the need to input a wide range of common sense and opinions into AI to instill values and enable self-checking. However, it was pointed out that classical AI faces the challenge of dealing with an overwhelming amount of facts that are too numerous for humans to teach, raising the question of how to formalize this information into a logical format. Regarding the second point, it was noted that the acquisition of common sense among humans varies, with some aspects gradually gained through daily life, while certain cultures employ specific methods, such as allowing children to touch a hot stove to understand that it’s hot. Given that common sense is not universal, there is no one-size-fits-all approach to simply inputting it into AI.

As the use of AI for tasks like translation and summarization becomes increasingly prevalent for graduate students and researchers, it is crucial to consider the challenges associated with the training and learning data behind these applications. We look forward to the future development of research concerning how much we can address challenges through the understanding of the human brain, societal activities, and their application to AI. Additionally, we aim to continue our ethical considerations on this matter.