We need operational definitions.
Written by | Gu Fanji (School of Life Sciences, Fudan University)
At the end of 2019, Dr. Karl Schlagenhauf and I co-published a three-volume series of books entitled "Brain and Artificial Intelligence" [1-3]. Our two authors grew up in completely different cultures and had never met: Karl was a German IT engineer and entrepreneur interested in brains, while I was a brain science professor and popular science writer at a Chinese university interested (albeit a layman) in artificial intelligence (AI). In early 2013, we were introduced to a mutual friend, neuroscientist Professor Hans Braun, and since then we have been in constant correspondence, discussing some open issues in brain research and AI, current status assessments and trends, how to view the relationship between the two and the "eye-catching" media coverage of these fields, to discussing scientific methodology and related scientific organizations. This set of books is an important collection of our correspondence records, if not unique, but still unique.
"Brain and Artificial Intelligence" series, Shanghai Education Press, 2019
The friendship between two very different people like us has been able to maintain and deepen because of a common interest in the intersection of brain science and information science. Moreover, we all like the method of rational thinking, and we are always eager to investigate the causes and reasons of things, rather than following the flow or being confined to pedantic thinking. Our views differ markedly due to different experiences, and we seek solutions from the ancient tradition of scientific debate. On some issues, we have reached consensus, while on others we remain divided, and there are even some questions that simply cannot be answered. Some new developments support the views of either or one of us and encourage us to discuss further. Some progress exceeded even our best expectations, or suggested that we or one of them was wrong, so we had to reconsider and learn from our mistakes. All of this fuels our enthusiasm to refocus on the issues we want to discuss and to open up new debates. These arguments are not meant to overwhelm the other side to show their cleverness, but to explore the truth of the matter.
Unfortunately for us, and at the same time fortunate, the timing of the publication of this set of books – the end of 2019. Over the next three years or so, humanity encountered two major events – the coronavirus pandemic sweeping the world and the advent of ChatGPT – that we could not have foreseen when we first talked. These events undoubtedly distract readers from our book, and at the same time serve as a severe test of the main ideas of our book. During this period, there were also some sensational events at the intersection of brain research and information technology—though not on par with the two above—such as Musk's Neuralink brain-computer interface and Jeff Hawkings' theory of "thousand-brain intelligence." And the EU human brain project that triggered our discussion is also about to expire, and it is time to close the coffin.
We are relieved that while we, like most people, failed to foresee specific events like the pandemic and ChatGPT, there are no major missteps in reviewing the overall view of open issues, status quo assessments, and trends in brain research and AI. On May 21, 2023, our set of books was awarded the first prize of the "2022 Shanghai Science Education Innovation Award Achievement Award (Book Category)", which can be seen as a certain degree of affirmation of our views by the public and experts.
At such a special moment, it is interesting to turn to the "few open questions" and "some expectations" that I wrote down when the three books were just undrafted in mid-2018 to see if these understandings at that time can withstand the test of the turbulent waves of the past five years. Of course, this does not mean that our views are necessarily correct, and the purpose of writing that set of books is not to give the reader any conclusion, but only to ask questions, lay out their own opinions and arguments, and importantly, hope to guide readers to join our thinking and discussion. Even if the manuscript was completed five years ago, it is still not obsolete to revisit these issues and think about them today.
The things listed below, whether right or wrong, are the exact words of the time.
A couple of open questions
1. Is the brain an information processing system or a machine for extracting meaning?
2. What are the functional primitives of the brain? (Ion channels?) Synapse? Neuron? Functional column? ... ）
3. Can all research results from non-declarative memory be generalized to declarative memory?
4. What does computing mean in the brain? Can neurons perform calculations in the Turing sense?
5. What is intelligence? What is the relationship between intelligence and skills, intelligence and learning ability?
6. Are consciousness and the content of consciousness the same thing, or two different things?
7. Is Chalmers' "difficulty problem of consciousness" a bottleneck in consciousness research or a pseudo-problem?
8. Is it possible in principle to do mental uploading?
9. Is there free will? How to resolve the contradiction between free will and determinism?
10. Problems that can be solved and problems that cannot be solved by large brain programs.
11. Is it possible to reverse engineer the whole brain to uncover the mysteries of the brain and solve the ultimate problem of artificial intelligence?
12. Should engineering or be able to replicate all strategies of nature's evolution? Is there brain-like computing?
13. What is the exact relationship between AI and brain research?
14. What is the future of neuromorphic chips?
15. What can machine translation do?
16. Is it possible to achieve strong AI in the foreseeable future? If so, should strong artificial intelligence be developed?
17. Is there a "singularity" day? Is the phrase "the singularity is near" credible?
18. Under the von Neumann architecture, can all functions of the brain be achieved by enhancing computer power?
A few points to expect
Technology will advance faster than brain science, but there will be no "singularity" in the foreseeable future. Technology will seek inspiration from brain science research, but it will not replicate the brain. This is because nature and the methods employed by engineers are fundamentally different. Therefore, brain research and engineering technology will still develop in parallel, but they will learn from each other. It is suggested that the Chinese of Brain-inspired be changed to "brain-inspired" instead of "brain-like" to avoid misunderstandings.
"Big Science, Team Science and Open Science" will make significant progress in basic data collection, clinical data collection, atlas, and research tool development in brain research, but it is difficult to hope that a breakthrough will be made in establishing the basic theoretical framework of brain science.
It is rarely possible to achieve "mind upload" and "reverse engineer" to build a whole-brain model of the human brain.
Whether for brain science or artificial intelligence, when the problem involves "inner problems" such as mind, intelligence, and consciousness, the common bottleneck faced is "subjectivity", and the question that should be asked is not "How does subjectivity arise?" Rather, what are the "sufficient and necessary conditions" it produces? There will be steady progress on the latter issue, but there is no hope that it will be resolved in the foreseeable future.
Under conditions where energy savings are extremely important, "neuromorphic chips" may have important applications. However, it is uncertain whether it will be developed into a new generation of computer systems. This will largely depend on how many people are willing to ditch traditional computers and relearn the "ecosystem" of this new generation of computers.
It is expected to make breakthroughs in the study of the activities and mechanisms of many local neural circuits at the mesoscopic level.
Looking back at the "open question" written almost five years ago, it is not systematically asked, and the size and importance of the problem are uneven. Some of them I have already judged (some of them have been written in the "A few expectations" above), while others I am still clueless. However, these issues are still not discussed publicly, and "openness" remains the same. In today's short article, it is impossible to discuss each of these issues in detail, but we can see whether our main expectations can withstand the test of the stormy waves of these five years.
In the book where the two men came to terms after arguing, Carl convinced me more than I convinced Carl. One of the things that struck me the most was Carr's assertion: "Physicists, especially engineers, outperform life scientists in this discipline, and engineering seems to be advancing at an exponential rate, while neuroscience and medicine may only advance linearly." "At first, to me, a brain scientist, it did sound harsh. But then Carl cites a large number of cases in the history of science and technology, which makes me have to admit this reality.
In fact, the new crown epidemic and the advent of ChatPGT can also be regarded as proof of Carr's thesis: on the one hand, a virus with a simple structure that is difficult to say even life is raging around the world, bringing huge loss of life and economic losses, leaving the world's brightest heads in a hurry, and still unable to find a perfect solution; On the other hand, there are those who claim to be about to unravel the world's most complex system, the human brain, which can be copied with computers in a decade or decades. When Musk, wearing a mask, announced at the press conference that he would soon be able to become superhuman by implanting chips that fused the human brain and artificial intelligence, a huge sense of irony came to his face. (I applaud Neuralink's significant advances in technology, but it is not innovative in its thinking, and fusing the human brain and AI to create a superhuman is pure myth)[4]. In stark contrast to the staggering of medicine is the success of computer science, in less than a century, from the invention of the first bulky electronic computer that occupies an entire room to the advent of ChatGPT, it is no longer a fantasy for machines to pass the Turing test, when we talk to chat software, if we do not think about it in advance, deliberately set a trap, it is indeed difficult to distinguish whether the other party is a human or a machine.
Multi-level brain: The various levels of the brain influence each other, forming a complex circular causal relationship.
Carr has long suggested that engineers don't benefit much from the results of brain research, and that going their own way and ignoring biological models would have better results. This is also something that I had a hard time accepting at first. Although there are still many people who say that the bottleneck of artificial intelligence is not understanding the brain, and only copying the brain can make a substantial leap [5, 6], everything that has happened in the past five years has proved otherwise. If neural networks and deep learning were a bit brain-inspired shadows (multi-level projection of receptive fields) at the beginning, then the development of ChatGPT was completely based on big models and big data in information science and technology.
On the contrary, taking the route of copying the human brain, whether in terms of elucidating mental mechanisms or applications, has so far made no impressive achievements. The EU Human Brain Project has long abandoned the goal of copying the human brain, and the "Blue Brain Project" that continues the Markram copy route has achieved some achievements in simulating brain tissue at the neuronal (and perhaps cortical column) level, and has not achieved any cognitive function at a higher level [5]. Hawkins' theory of "thousand-brain intelligence" is not only untenable on a neural basis, but also remains a dead letter in practice [6]. Although the late Edelman's "brain-like machine" (that is, Darwin's machine) imitating the cerebellum can freely traverse curved paths in the laboratory, the unmanned cars that actually drive on the road today are completely the product of machine learning and have nothing to do with the cerebellum (of course, Edelman's work helps to understand the cerebellum mechanism).
Of course, our conclusion – "trying to reverse engineer a biological brain on silicon is not very promising" – remains to be further tested in the future.
Over the past five years, the facts have proved time and again that the large-scale brain project "will make significant progress in basic data collection, clinical data collection, atlas and research tool development in brain research, but it is difficult to hope that a breakthrough will be made in establishing the basic theoretical framework of brain science." [1] Indeed, the greatest achievement of the EU Human Brain Project is the establishment of an information technology platform for neuroscientists to share, the main achievement of the Blue Brain Project and its collaborator Allen Institute for Brain Science in the United States is the publication of the taxonomic map of neurons in various brain regions of the mouse brain, and the progress of many national brain projects on the connectome map, all of which may provide basic data for future breakthroughs in brain mechanism research, but they are not breakthroughs in themselves. As Professor Chen said, "The greatest discoveries in science are not planned." For groundbreaking research and discovery, "we think of passionate individuals and small teams, like captains of small research boats and their crews, especially those young students. [1] However, it is extremely difficult to predict when and where such individuals or small teams will emerge.
A common problem in the current AI R&D community is the confusion between the first-person perspective (the subject's perspective of examining its inner activities) and the third-person perspective (the perspective of a third party observing from the sidelines). Many brain functions, especially higher functions, especially the mind, can be viewed from two different perspectives: first-person and third-person. If the former is to be expressed in one word, it may be called "inner activity"; The latter, if it is also to be expressed in only one word, may be "behavior".
At present, there is no recognized precise definition of "mind", which is often explained by what it covers. For example, Wikipedia's Mind entry describes it this way: "Mind is an umbrella term for a group of cognitive abilities, including consciousness, imagination, perception, thinking, judgment, language, and memory, which come from the brain (and sometimes the central nervous system). It is usually defined as the ability of an entity to think and be conscious. It possesses imagination, recognition and appreciation, and is responsible for processing feelings and emotions that generate attitudes and actions. ”[7]
The inner activity of the first-person perspective is subjective and private, and can only be experienced by the subject himself, and cannot be accurately shared with others. The behavior of the third-person perspective is broad and includes all activities that can be observed and measured. The many aspects of the mind have both dimensions, and people often confuse these two aspects in everyday life. Sometimes one word is used to confuse the two aspects, sometimes different words are used to emphasize different aspects, but there is no clear recognized dividing line, which is also a problem that makes me scratch my head when writing down. For example, emotion is often a general term, while feeling often refers to inner feelings, and expression obviously refers only to the external manifestations of emotions; Similarly, sensation emphasizes the response of the senses to a stimulus and can be measured objectively, while perception is the subjective experience of a stimulus. Of course, when consciousness is involved, subjectivity and privacy are even more prominent - whether there are aspects of consciousness that can be objectively measured is still a matter of opinion. Although the "neural correlation set of consciousness" may sound objectively measurable, this is only "correlation"; And, if there is no subjectivity and privacy, can what remains be called consciousness? This also seems to be a problem.
So far, I believe that all that can be created by humans can only be from a third-person perspective, that is, imitation behavior, and there is nothing to do about the "inner activity" from the first-person perspective. The problem is that people often confuse the two, describing artificial imitation as successfully "achieving artificial inner activity", so some people start talking about "artificial consciousness" and "emotional machines". As for the commonplace "mind control", it is really nothing more than the control of brain signals, which can mislead the public.
Of course, I am not concluding that artificial inner activity will never be possible. Because since the human brain has inner activity, and the human brain is ultimately a material system, there is no reason to rule out the possibility of other material systems emerging from inner activity. What I emphasize here is that mental activity is a property of highly complex material systems that emerge under certain conditions (we just don't know what those conditions are and how complex the system is), rather than being independent of the brain. However, until now we don't know what exactly is needed to emerge from a "first-person perspective". We don't yet understand what people need, let alone artificial systems.
Indeed, from an application perspective, we can leave the first-person perspective aside and only discuss the aspects (i.e., behavior) seen from the third-person perspective, and give some kind of operational definition from that perspective, so as not to be confused. For example, Professor Wang Pei of the Department of Computer Science of Temple University in the United States defined intelligence: the ability to adapt despite insufficient knowledge and resources [8]. This is an important aspect of intelligence described entirely from a third-person perspective, both for human intelligence and artificial intelligence. Wang Pei's definition is broad enough to describe many intelligent behaviors, and from this point of view, he built his NATH AGI system to solve practical application problems, which is enough for his purpose. So is there an aspect of intelligence that is described from a first-person perspective? I think there are also, such as "understanding". Therefore, I strongly disagree with the use of "natural language understanding" in the field of artificial intelligence, and I think that what is being done at this stage is only "natural language processing".
For current research on artificial intelligence, I think that if you use terms similar to psychology (or philosophy of mind), then you should give an operational definition at the beginning, excluding the first-person perspective. And don't apply the conclusions of future work to issues related to the first-person perspective.
In this regard, the fundamental reason why the phenomenon experienced from the first-person perspective cannot be "restored" to the mechanism that can be explained from the third-person perspective is that the former occurs at the upper level of a complex system with many levels, and the existence between these levels is not a linear causal chain in a simple system, but a "circular causal relationship" within and between levels that is mutually causal. For a long time, natural science has been dominated by reductionism based on linear causal chains, and the formal proposal of "circular causation" was only at the end of the last century [9]. Until there is a deeper understanding of this causal relationship, efforts to address artificial "inner activity" will be difficult to achieve if not for the sheep's head selling dog meat.
bibliography
[1] Gu Fanji, Karl Schlagenhauf, translated by Gu Fanji (2019) A New World of Brain Research: A Dialogue Between a German Engineer and a Chinese Scientist, Shanghai Education Press
[2] Gu Fanji, Karl Schlagenhauf, translated (2019) The Mystery of Consciousness and the Myth of the Mind: A Dialogue Between a German Engineer and a Chinese Scientist, Shanghai Education Press.
[3] Gu Fanji, Karl Schlagenhauf, translated by Gu Fanji (2019) The Third Spring of Artificial Intelligence: A Dialogue between a German Engineer and a Chinese Scientist, Shanghai Education Press.
[4] Gu Fanji (2020) If you want to harness AI, first live with AI: Can Musk's "superman" plan succeed? Back to Park, August 12, 2020
[5] Fan X and Markram H (2019) A Brief History of Simulation Neuroscience. Frontiers in Neuroinformatics. 13(Article 32):1-28
[6] Hawkins J (2021) A Thousand Brains: A New Theory of Intelligence. Basic Books.
Chinese translation: by Hawkins, translated by Liao Lu et al. (2022) Thousand Brain Intelligence, Zhejiang Education Press.
[7] https://en.wikipedia.org/w/index.php?title=Mind&oldid=911062349
[8] Wang Pei (2022) Outline of Intelligence Theory, Shanghai Science and Technology Education Press
[9] Haken H (1996) Principles of Brain Functioning: A Synergetic Approach to Brain Activity, Behavior and Cognition. Springer.
Chinese translation: Haken, translated by Guo Zhi'an and Lv Ling (2000) Brain Working Principle: A Collaborative Study of Brain Activity, Behavior and Cognition. Shanghai Science and Technology Education Press
This article is supported by the Popular Science China Starry Sky Project
Produced by: Science Popularization Department of China Association for Science and Technology
Executive producers: China Science and Technology Press Co., Ltd., Beijing Zhongke Galaxy Culture Media Co., Ltd
Special mention
1. Enter the "Boutique Column" at the bottom menu of the WeChat public account to view a series of popular science articles on different topics.
2. "Back to Pu" provides the function of searching articles by month. Follow the official account, reply to the four-digit year + month, such as "1903", you can get the article index in March 2019, and so on.
Copyright note: Welcome personal forwarding, any form of media or institutions without authorization, shall not be reproduced and excerpted. For reprint authorization, please contact the background in the "Huipu" WeChat public account.