Berggruen Seminar Series: What is Artificial “Intelligence”? A Comparative Framework for Essential Exploration

Artificial intelligence is a primary focus in today’s world, however neither researchers nor practitioners have reached a consensus on what is meant by intelligence. Artificial intelligence cannot be equated with machine learning, as intelligence has more complex connotations. On the subject of intelligence, Professor Su Yanjie from Peking University’s School of Psychological and Cognitive Sciences, shared her thoughts on the nature of intelligence from a comparative framework of the human mind and animal mind.

Zeng Yi, Berggruen Fellow and research fellow at Chinese Academy of Sciences, introduced the theme of the lecture. He believes that the core question of artificial intelligence is how to realize the human mind in the physical world, and as such, how artificial intelligence research is closely related to the exploration of intelligence in cognitive science. Artificial intelligence acts as a mirror for research of the human mind. The exploration of artificial intelligence is exploration of the human race, that is, the exploration of possible future lives. If our understanding of human intelligence lacks depth, it will be difficult to create future intelligence.

Professor Su began the lecture by emphasizing the importance of the comparative framework in the process of interpreting what we see. Surveys show that when comparing intelligence between various species, most college students tend to compare AlphaGo with chimpanzees, an animal that is considered highly intelligent. That is to say, people usually hold a basic judgment on the intelligence level of things, and based upon this judgment, some robots are very intelligent.

Following this, Professor Su reviewed the historical studies on human intelligence in psychology. Intelligence is a systematic ability, therefore it is difficult to directly define human intelligence. There are many theories on understanding intelligence in the field of psychology and the conceptual system of research is situated in the process of continuous exploration and development. Intelligence testing is closely related to how intelligence is defined. Researchers who adhere to different conceptual systems will develop different measurement systems and design corresponding measurement tools to measure corresponding capabilities. Comparative psychology defines intelligence at different levels of comparison, such as the biological level, the psychometric level, and the outcome level of adaptation to the environment. Psychologists generally understand intelligence as high-level cognitive activity of the brain. However, since machines are not biological systems, it is difficult for them to reflect the cognitive functions of the brain.

There is a difference between wisdom and intelligence, which are both concepts related to intelligence. People’s understanding of wisdom is closely related to their vision on the time axis. Another related concept is cognition at the level of information processing. In contrast, intelligence focuses on external performance while cognition is an internal phenotype similar to a program; intelligence is a general concept while cognition is mainly a specific component of generalization and abstraction.

Professor Su further pointed out that the significance of the comparative framework for artificial intelligence research lies in the fact that we can explore how to apply these concepts to artificial intelligence research by attaining an understanding of which concepts have been used in the study of human intelligence and the comparison of human beings and animals. Human intelligence itself is a complex problem. And it becomes more so when animals are incorporated. Research exploring animal intelligence can be divided into three ways of describing the similarities and differences between human beings and animals: animal psychology and experimental psychology focusing on explicit behaviors, neuroscience focusing on physiological processes, and evolutionary psychology on the remote causes of intelligence. An understanding of intelligence involves various concepts, such as cognition, adaptation, learning, and information processing. As different concepts are introduced, people can discuss intelligence from different perspectives.

In defining animal intelligence, one must pay attention to the fact that different species have different sensory advantages and different availabilities of responses. The same standard cannot be used to measure the capability level of all species. Rather, the measurement method needs to be assessed to make sure it is suitable for that particular species. Additionally, individual differences are also important factors affecting the research results. It is difficult to define the intelligence level of animals through explicit behaviors alone, as there may be other factors at play such as motivation. For example, human beings recognize numbers through language, but animals do it by using their perceptual abilities. Human beings and animals have different representation patterns, so it is difficult to compare their numerical abilities.

From this perspective, Professor Su reflected on the intelligence tests designed for robots and concluded that tests that have been designed to measure human intelligence such as the Wechsler Intelligence Scale are not suitable for measuring robots. Many robots are actually not as smart as their developers claim them to be. For example, researchers often use the mirror test as a standard way of testing for the development of self-awareness, and yet even if similar test results are obtained, it is still worth considering whether the internal mechanisms behind those results are consistent with those of humans’. With this in mind, Yanjie invited everyone to rethink what wisdom is in essence. People often focus on the functional performances of external behaviors, but there may be different internal mechanisms behind the same performance. Ignoring this may confuse our understanding of the comparison results.

During Q & A session, Professor Bai Shunong of Peking University proposed that although wisdom is considered by psychologists as a concept involving the time axis, if the vision and horizon representing “wisdom” are understood as cognition, which is the process of processing information, then the so-called “vision” and “ horizon” actually reflect an information processing pattern in which a larger amount of information can be processed more efficiently. Time is actually an artificial setting. Without memory, there would be no past tense; without imagination, there would be no future tense. Memory and imagination are both specific functions of the brain’s neural network, and as such time is essentially a kind of interpretation of the outcomes of the activities of neural networks.

Professor Su agreed that if information processing is used to explain wisdom, then wisdom can also be understood as a model of information processing and is the same thing as intelligence. She also proposed that we analyze the concept of wisdom purely through description and deduction, however, as the public’s understanding of wisdom is attained from the results of problem-solving, it is difficult to find a standard of judgment. Through information processing, it is possible to quantify and grasp such a difficult concept as wisdom, and provide a basis for scientific research.

Zeng Yi concluded that artificial intelligence has been developing for more than 60 years, and all that can be said of the progress made is that it has promoted human understanding of intelligence. We are far from reaching a consensus on intelligence in this field, and people’s understanding of artificial intelligence is filled with fantasy and imagination. The biggest problem with artificial intelligence lies in our pretense of understanding human intelligence, and people excessively believe that artificial intelligence may end in disaster. Zeng believes that artificial intelligence researchers cannot sway the public, and in fact the unsolved problems of 60 years ago are still unsolved now. As Moravec’s paradox reveals, what appears easy is actually difficult, while what appears difficult is actually easy to achieve.

Zeng gave a simple introduction of his work, namely, artificial intelligence research inspired by the brain structure. He believes that although many so-called robots appear to be highly intelligent, they lack even basic capabilities. Human intelligence levels and artificial intelligence levels are difficult to compare, and the internal mechanisms behind their common behavioral performance are actually different. Zeng reflected that current research on artificial intelligence only focuses on external behavioral performance, which may be the wrong starting point as the interpretation of the mechanism has been ignored.

Zeng believes that if the “intelligence” of artificial intelligence and human performance are widely different to each other, it is necessary to rethink whether that is “intelligence” or not. As Dennett said, if a lion could speak, this wouldn’t help us understand the lion itself at all; by the same token, if the intelligent implementation of artificial intelligence has nothing to do with human intelligence, then whether or not it can still be called intelligence will become a problem. If the mechanisms for achieving intelligence are different, then harmoniously coexisting with such superintelligent agents and treating them as partners of human beings will create huge risk for us. We need to distinguish information processing tools from artificial intelligence. The so-called “artificial intelligence researchers” are rarely the true researchers of artificial intelligence. Human exploration of cognition and neuroscience may be an endless topic; as artificial intelligence mirrors human beings, research on artificial intelligence is also a topic that can be explored for hundreds of years.