The 2010s were a great year for artificial intelligence (AI).
With the development of deep learning technology that can learn like humans with the ability to process large amounts of data, humanity has begun to experience artificial intelligence that can replace humans everywhere.
But scientists aren’t seeing deep learning as the final solution. AI experts agreed last week in an online discussion hosted by Montreal.AI that another solution beyond deep learning is needed to bring AI to the human level.
Beyond’deep learning’, artificial intelligence capable of common sense and logical reasoning like humans is being developed through scientists. It is foreshadowing the era of artificial intelligence in which people can think and communicate like people. ⒸGetty Images
The downside of deep learning is human reasoning ability
In this discussion held on the subject of’AI Discussion 2: AI Development: Interdisciplinary Approach’, AI experts pointed out various problems of deep learning.
In particular, cognitive scientist Gary Marcus, an invited debate, explained that, unlike deep learning, which looks splendid from the outside, practitioners are experiencing difficulties due to various problems occurring in the operation process.
The most difficult part is the amount of data that deep learning requires. He said that in the process of collecting, analyzing and processing data, it is demanding too much data compared to other algorithms, which adds to the difficulty in operating artificial intelligence.
Gary Marcus also revealed that deep learning is struggling to apply the knowledge generated from data to other areas. In particular, he pointed out various disadvantages such as opacity and lack of expressive power in expressing reasoning and knowledge.
Humans receive various stimuli through sensory organs composed of five senses.
The information received in this way is transmitted to the brain, where the relationship is analyzed. For example, if you have a book and a ballpoint pen on a desk, you may be reminded that you can read a book in front of the desk and take notes with a ballpoint pen.
This ability is called reasoning. The problem is that artificial intelligence lacks this common sense ability.
As seen in’Alphago’, which beaten Go knight Lee Se-dol, it is a fact that deep learning far exceeds human abilities in some fields. In fact, cases of diagnosing skin diseases more effectively than doctors have been reported in the medical field.
However, if you move to a slightly different area, you cannot exert its ability. This phenomenon suggests that they do not comprehensively understand what they are doing, and that they are not able to process it in a high level.
Interactive AI development with common sense
These shortcomings in deep learning have been frequently pointed out by experts.
As such, experts attending the Montreal AI discussion revealed that they are developing various types of artificial intelligence, such as hybrid AI.
Computer scientist Luis Lamb, who wrote the book “Neural-symbolic Cognitive Reasoning,” added a function of logical formalization to machine learning functions, and “neural-symbolic AI” based on these two axes. Suggested.
The function of logical formalization is to develop logic that fits artificial intelligence just as it does from philosophy to logic. He said that while expanding the logic, it is possible to expand the area of AI participation.
Stanford University professor Fei-fei Li, former Google Cloud’s top AI scientist, said, “Currently, artificial intelligence systems are very lacking and lacking unlike humans in their active interaction with the world around them. .”
Professor Lee said, “In order to compensate for these shortcomings, artificial intelligence needs to be socialized.” “Currently, Stanford Research Institute is building interactive agents that can act on their own while understanding the world around them. ”
Computer scientist Richard Sutton points out that “most AI work lacks the term “computational theory,” coined by the renowned neuroscientist David Marr.”
Computation theory refers to the study of how efficiently a particular algorithm can perform a problem.
Richard Sutton said, “With this computational theory, there is no discovery of this computational theory other than reinforcement learning in defining the goal and the reason for pursuing the goal.” “We need to develop an additional computational theory for this.”
Professor Ye-jin Choi of the University of Washington emphasized the importance of common sense. Artificial intelligence does not have the common sense that people have.
However, Professor Choi said, “We have overlooked the importance of common sense in the development of artificial intelligence. We need to supplement the system so that artificial intelligence can have knowledge about the world around it.”
Professor Choi said, “The space for reasoning is infinite, but we need to expand the area of common sense to support it.”
In addition, Professor Choi said, “In order to reach human-like common sense and reasoning ability, extensive parallel research is required, such as combining symbolic and neural expressions, integrating knowledge into reasoning, and building a benchmark system rather than simple categorization. Insisted.
(15)