Shulou(Shulou.com)11/24 Report--

This article comes from the official account of Wechat: back to Park (ID:fanpu2019), author: gu Fan and

"every generation thinks they can connect neural mechanisms to high-level behavior, but in fact, all we hold on to is straw that will form the basis of skyscrapers. Scientists have to be patient."

Written by Gu Fanhe (School of Life Sciences, Fudan University)

Jeff Hawkins, the creator of the "thousand-brain intelligence" theory, is known as the "legend of the technological generation". He is born in electrical engineering but is curious about how the human brain has intelligence. He once applied for "creating intelligent machines based on brain theory" as a doctoral research subject, but was repeatedly rejected without changing his original mind. Finally, he studied neuroscience at the University of California, Berkeley, then returned to the field of information technology and invented the first PDA, which was a great success. If he continues down this path, he may take the lead in smartphones, but he has resolutely returned to brain theory research, even founded the mahogany Neuroscience Institute (Redwood Neuroscience Institute), and later founded Numenta, which specializes in theoretical research on how the neocortex works.

There is no doubt that Hawkins chose a unique path in the theory of thousand-brain intelligence. There may be two main reasons for this: one is that he read an article by Nobel laureate Francis Crick in Scientific American magazine in 1979, which said: "although detailed knowledge about the brain is accumulating, exactly how the brain works is still quite mysterious."... what brain science clearly lacks is a universal mental framework that can explain these results. " The second was in 1982, when he read an article by Vernon Mountcastle, president of the American Society of Neuroscience, who later recalled: "Moncasser proposed that cortical and microcortical columns throughout the neocortex have the same function and perform a set of basic algorithms responsible for all aspects of perception and intelligence." At the time, he said, "I was so excited that I almost fell out of my chair," which he thought was the "Rosetta stone" that unraveled the mystery of the brain. [note 1] "how I wish you could appreciate the brevity and thoroughness of Moncasser's point of view. The best scientific ideas are often concise, thorough and extraordinary." Therefore, he is determined to do two aspects of work: ① establishes the theoretical framework of the working mechanism of the neocortex and clarifies the nature of intelligence by understanding the working principle of the brain; on this basis, ② creates an intelligent machine similar to the human brain through reverse engineering.

[note 1]: the ancient Egyptian monument found near the town of Rosetta in 1799 was engraved with three characters such as ancient Egyptian hieroglyphics and Latin, thus solving the mystery of ancient Egyptian hieroglyphics. Later, people will use Rosetta stone metaphor to solve the key clues of the problem.

After a long period of painstaking research, Hawkins published the Thousand brain Intelligence (A Thousand Brains) in 2021 after the best-selling book on Intelligence (On Intelligence) published in 2004. As soon as this book was published, it attracted wide attention in the field of neuroscience and artificial intelligence. "I believe we have found the mental framework mentioned in Crick's article, which not only explains the basic principles of how the neocortex works, but also provides a new way to think about intelligence," the book says. " [3] if this is the case, it will naturally be an earth-shattering event in the entire scientific community. The book also says: "to create truly intelligent machines, we first need to reverse engineer the brain." The implication is that Hawkins' work has opened up a new field of artificial intelligence. In 2021, "Thousand brain Intelligence" entered Bill Gates' annual book list, was translated into various languages, and entered the best-selling list promoted by publishing companies and bookstores.

Left: "Thousand brain Intelligence" book marketing materials; right: Hawkins at the 2007 eTech conference. Unfortunately, from the author's point of view, the ways used by nature and engineering technology are completely different, and it is only a myth to reverse engineer the human brain, copy the human brain and create an artificial brain. Although there is no lack of thought-provoking sparks in this book, it is generally more like a pile of colored broken glass mixed with treasures. In this paper, I would like to make a brief introduction and comment on his "thousand-brain intelligence theory". Due to the limitation of space, this paper focuses on analyzing why the theory of "thousand brain intelligence" does not constitute the theoretical framework of brain function, because it is the foundation and core of his whole work. If the foundation is not solid, the magnificent blueprint depicted later will be nothing but mirrors. Of course, the author understands that his own level is limited, and can only hope to arouse the same good interest and seriously think about and discuss these important issues.

If Hawkins' thought framework is to be put in one sentence, it may be that the human brain builds a world model through sensation-movement, and then predicts what will happen next according to this model. "Model-prediction" has also become his definition of intelligence. This core idea is not initiated. For example, the American neuroscientist Walter J. Freeman III once put forward similar ideas [4], but Hawkins clearly summarized these ideas as the basic mechanism of brain function and took it as the definition of intelligence. However, this definition does not answer how the brain models and predicts, and Hawkins'"thousand-brain intelligence theory" tries to answer this question.

Hawkins' answer is mainly based on the following two hypotheses: (1) all cortical columns are almost the same, they are the basic units of brain function and all perform some universal algorithm; (2) each cortical column is attached to multiple reference frames of the same thing, and the properties of things are organized in these reference frames. The upper cell of the cortical column is similar to the positional cell, with the nature of things; the lower cell of the cortical column is similar to the grid cell, with the location information of things. There is a frequent exchange of information between these two types of cells, so if you know the location, you know the content, and when you know the content, you know the location. When the subject moves, the change of position causes a series of changes in content, which is thought. Different cortical columns solve the problem of "binding" by "voting". Hawkins carried out the simulation based on this, and achieved similar results as he imagined. He was thus convinced that his theory was the missing theoretical framework in brain science. It all sounds perfect, but it doesn't stand up to scrutiny in my opinion. Below, we will see whether these two premises are true from the perspective of neuroscience.

The cortical column assumes that the cortical column (cortical column) discovered by Moncasser's study of the somatosensory cortex is the basic source of Hawkins' thousand-brain intelligence theory, so we need to first understand what the cortical column is.

Moncasser's and later David Hubel and Torsten Weisel's studies of the cortical columns in the primary visual cortex have shown the following fact: anatomically, the cell bodies and nerve fibers in the longitudinal sections of the brain show some kind of longitudinal tissue. From the electrophysiological records, the cells perpendicular to the surface of the cortex and along the same line have the same receptive field, and detect the same characteristics (in fact, there are exceptions to this: for example, neurons in the same cortical column in the primary visual cortex are generally thought to detect the orientation of the stimulus segment. But the middle layer of some cortical columns also has a so-called "plaque" structure, which detects the wavelength of the stab laser), and these cells share common inputs and outputs. Except for a few "singularities", the characteristic parameters detected by the adjacent cortical columns change continuously, and this change will also show periodicity. These facts are indeed universally acknowledged.

Cortical column structure? source: wikipedia Huber et al believe that this organization of cortical neurons may bring the following benefits: the shortest distance between neurons with similar functional properties and the sharing of specific input information from different pathways, so that the brain can make the most effective use of space, speed up processing speed, and use the least number of neurons to analyze different attributes. [5] that's all, how can there be him?

But Hawkins claimed: "all cortical columns, even those in low-level sensory areas, can learn and recognize complete objects."

However, almost all neurobiological literatures tell us that the cells in the primary sensory cortical column can only detect the local characteristics of the stimulation object, so there is a "binding problem". In his book, Hawkins emphasizes that "complete objects, not features, are transmitted between levels." It is true that there are cells in the higher cortex that can recognize entire objects, such as faces, but it is not clear whether this is purely the function of a single cell or whether the circuit function appears on a single cell (such as the top of an iceberg exposed above sea level).

"there are many models in the neocortex that target a specific object," Hawkins said. these models are located in different cortical columns. they are not exactly the same, but complement each other. " "knowledge about an object. Stored in thousands of models, that is, thousands of cortical columns, but these still account for only a fraction of all cortical columns in the neocortex. That's why we call it the 'thousand-brain intelligence theory': knowledge about any particular object is distributed in thousands of complementary models." What puzzles the author is whether what he calls a "model" is a model of the whole world or a model of an object. He seems to cross-use the word "model" sometimes to refer to the former, and at other times to refer to the latter.

In addition, at present, the research results of cortical column mainly come from sensory cortex and motor cortex, because for these two kinds of cortex, experiments can detect the function of neurons, so that we can talk about the receptive field or projection field of neurons. For the joint cortex, it is difficult to know what the function of individual neurons is, and there is no way to know whether the joint cortex has a functional cortical column structure.

Question one: does the cortical column really exist? In 2005, celebrating the 50th anniversary of Moncasser's discovery of the cortical column, American neuroscientists Jonathan C Horton and Daniel L Adams summed up half a century of research on the cortical column and concluded that the cortical column may not function. "after half a century, it is still not clear what the term means," they point out. There is no single structure in the cortex that corresponds to it. It is impossible to find a standard microcircuit corresponding to the cortical column. " "in some ways, the idea that the cortical column is the basic functional entity of the cortex must be abandoned. From now on, the exhilarating theories that lead to the conclusion that there is a modular structure in the cerebral cortex seem to be unreliable. the structure of each brain region is different, and in order to fully describe the architecture of the cortex, each cell, each level, each circuit and each projection have to be considered separately." [6]. Of course, not everyone agrees. For example, Nuno Ma ç arico da Costa and Kevan A. C. Martin believe that it is too early to completely deny the function of cortical columns. nevertheless, they admit that "it is clear that there is no recognized single cortical column anatomical entity." [7] they believe that although the cortex is organized not only in the hierarchical structure, but also in the vertical dimension, like the hierarchical structure, "the size and shape of these vertical organizations are also very different" [7]. Moncasser et al regard the longitudinal tissue formed by the apical dendritic bundle of pyramidal cells as the anatomical basis of the cortical column, but the lateral connections between the basal dendrites and axons of pyramidal cells go beyond the range of several "microcolumns". So the so-called cortical column does not have a clear boundary.

After a long and in-depth study of the cortical column problem, Kathleen S. Rockland, an American neuroanatomist, came to the following conclusion: "Anatomical columns are not a structure with a solid foundation, they are part of a system connected to each other in local brain regions, and any particular column mentioned is also involved in a widely distributed network. Columns are not a necessary feature of the cortex. There are also a wide range of columnar structures in the non-cortical structure of the brain. " [8] she believes that "cortical column" is an oversimplified concept that has become too rigid as a term and loses the ability to express complex and dynamic aspects of cortical organization. "the word inevitably has a fixed, homogenous and static meaning."

Many neuroscientists believe that regarding the microcolumn as the basic tissue unit of the cortex is an attractive hypothesis [9, 10], but it is still inconclusive. Daniel P. Buxhoeveden et al analyzed this problem comprehensively, and concluded that although the microcolumns are similar in appearance, they are also highly irregular. Describing the microcolumn as only repetitive, almost like a cloned copy, is very likely to be wrong and a source of some confusion. Although the microcolumns are similar in appearance, they are quite different in internal structure. In the process of development, microcolumns produce many highly alienated synaptic connections, which are different even from their neighbors. Therefore, there may be substantial differences between different cortical regions, or even microcolumns in the same brain region. Not only that, the number of afferent fibers in these microcolumns may be different, but also project slightly different efferent pathway groups. The difference between microcolumns may be due to the effect of environment on the structure of microcolumns, or it may be due to the difference in the distribution of cells and fibers. The microcolumn seems to be strongly affected by the environmental input, which makes each microcolumn different. This difference can explain some of the 'large' morphological differences observed between different microcolumns. These differences include the size of the extra-microcolumn nerve stack space (neuropil space), the cell density in the microcolumn, the degree to which the cell bodies are arranged vertically in a straight line, the width of the microcolumn and the ratio of the microcolumn width to the size of the surrounding nerve stack space. The heterogeneity of microcolumns includes structural shape, response properties, connectedness, immunohistochemistry, metabolism and stimulation preference. "

From the brief review above, it is not difficult to draw the following conclusion: although, as Kandel concluded, the cortex does show a certain form of structure longitudinally, but it is said that this structure is made up of the basic structural unit of the cortex, the "functional column". And all the functional columns in the cortex are exactly the same, so it is suspicious to implement some universal algorithm. If it is understandable and thought-provoking for Moncassel to put forward the above hypothesis when he first discovered this structural form in the 1980s, then after more than half a century of extensive research, Hawkins only adhered to Moncassel's early theory, ignoring the doubts caused by a large number of subsequent experimental results, and regarded this hypothesis as the basic starting point of his thousand-brain intelligence theory. It is tantamount to building a building on top of the sand. No matter how magnificent this theoretical mansion may seem, it is still unreliable. Unfortunately, Hawkins took this dubious assumption as the basis for all his work.

Question number two: does Hawkins really understand spatial cognition? The second starting point of the thousand-brain intelligence theory is called the "reference frame hypothesis". If the cortical column hypothesis has some experimental basis, the reference frame hypothesis is entirely out of a misinterpretation of the results of neuroscience research.

As mentioned earlier, Hawkins' frame of reference hypothesis is that each cortical column is attached to multiple reference frames of the same thing, and the nature of things is organized in these reference frames. He cited the experimental facts as follows: (1) the location cells (place cell) and grid cells (grid cell) discovered in recent years may form a system to let animals know their spatial location [11]; (2) there are two different pathways in the nervous system, "what" and "where", which are responsible for identifying exactly what the object is and where it is located.

What is now known is that there are positional cells in the hippocampus (which is part of the limbic system, not the neocortex, and it does not have a six-layer columnar structure in the neocortex), which is responsible for determining the spatial location of the subject in the familiar environment, rather than the spatial location of external objects [12] (although it is now known that there are so-called social location cells (social place cell) in the hippocampus of bat brain. It can detect the position of other nearby bats relative to themselves [13], but it is not known whether it can detect the spatial location of external objects that generally do not actively signal outward.

As for grid cells, it is now only known that they exist in the entorhinal cortex. In a familiar environment, grid cells are distributed at the vertices of triangles similar to a network of triangles. This triangular network is indeed a bit like a coordinate system within the brain, but it is also relative to the coordinate system of the subject itself. Scientists generally believe that grid cells are the basis for location cells to determine their spatial location in a familiar environment [14]. However, there is little evidence that grid cells can determine the location of external objects in this "coordinate system".

Above: positional cells in the mouse brain (yellow signal activation) and grid cells (blue signal activation) work together to sense "where am I" (mouse location). Below: the two neural connecting pathways in the neocortex of the human brain, the green is the dorsal pathway (dorsal pathway), which is responsible for processing where information, and the purple is the ventral pathway (ventral pathway), which is responsible for processing what information. Map Source: wikipedia in short, the existing consensus in the neuroscientific community is that location cells and grid cells are mainly used to determine their own position in the familiar environment, rather than the position of all external objects in the surrounding environment. In addition, these cells only exist in the hippocampal-entorhinal cortex system, and there is no evidence that there are positional cells and grid cells everywhere in the neocortex.

There is no doubt that the brain can recognize external objects and their position in space. For vision, it is known that these two functions may be relatively independent, which are realized by the ventral visual pathway to the inferior temporal lobe (what pathway) and the dorsal pathway to the parietal lobe (where pathway). To identify external objects and their locations, we need the whole pathway, not just individual cells, let alone location cells and grid cells.

In this way, Hawkins claims that there are identical functional columns everywhere in the neocortex, with the upper cells similar to positional cells to detect the properties of external things, while the lower cells are similar to grid cells to detect the location information of external things. I'm afraid it confuses a pair of cells and a pair of pathways. There are so many grooves that I don't know where to start.

Question 3: does the prediction rely solely on dendritic nerve pulses?

Hawkins used the cortical column hypothesis and the reference frame hypothesis to explain how the brain models, completing the first half of his intelligent "model-prediction" framework (although he did not say whether it was a model of the entire external world or individual external objects). As for how the model in the brain achieves the function of prediction, it is believed that it is due to the fact that some dendrites have been found to emit nerve pulses on neurons in recent years. Hawkins explained that these nerve pulses travel to the Axon hillock, which increases the membrane potential at the axial colliculus, so when an external stimulus comes, it can emit nerve pulses earlier than neurons that do not receive dendritic pulses-this is called prediction.

The Axon hillock is the region of the cell body near the axon in the neuron, which is close to the initial segment of the axon. Is it possible that the advanced cognitive function of "making prediction" depends only on the electrical impulses emitted by certain dendrites? Here Hawkins confuses the two different concepts of prediction and response lag time to stimuli. In addition, even if there is no dendritic pulse, the postsynaptic potential on the dendritic will spread to the axial colliculus as a hierarchical potential, although its intensity is weaker than that transmitted by the pulse, but this is only a quantitative difference, but a qualitative difference. Finally, it is generally believed that predicting external events is the function of the neuron population, not the function of individual neurons. Hawkins once again mistakenly packaged his ideas with neurobiological findings.

Other questions about the contents of the above three questions are more or less related to new discoveries in neuroscience, but there are still some hypotheses in the "thousand-brain intelligence" theory that have no neurobiological basis at all. For example, to solve the binding problem, Hawkins said: "the cortical column will 'vote', meaning that perception is the consensus reached by the cortical column by voting." the voting mechanism in the cortical column solves the 'binding problem', which allows the brain to combine various sensory inputs to form a single representation of perceived things. " At present, one of the sources of ideological confusion caused by artificial intelligence is that it is too anthropomorphic. Hawkins is not only an anthropomorphic problem here, but simply "socialized". He can't bring up the neurobiological facts that bear a slight resemblance to voting, but he still claims that his theory solves the problem of the ideological framework of brain function proposed by Crick, which is not serious in my opinion.

"there is still a lot we don't know about the brain, especially the neocortex," Hawkins wrote in his book Thousand brains. However, I don't think there will be another systematic theoretical framework to fill the boundary parts of the puzzle in a different way. With the passage of time, the theoretical framework will be gradually modified and improved. I expect the same to be true of the thousand-brain intelligence theory, but I believe that the core ideas I put forward in this book will remain largely the same. "

Through the above analysis, the author is deeply skeptical about his words. He stuck to Moncassel's assumption half a century ago, probably because it seemed simple and agreeable to him. This is very similar to what Kakhar described when refuting Gorky at the Nobel Prize ceremony: "Yes, if you look at things only from a logical point of view, so it is convenient and economical to assume that all nerve centers are made up of a continuous network of intermediaries between motor and sensory nerves. Unfortunately, nature seems to ignore our intellectual demands for convenience and unity, and often prefers complexity and diversity. " American writer John M. Barry said in his masterpiece The Great Influenza: "but people who study nature don't always use scientific methods." These researchers believe that if their knowledge is based on a premise they consider reasonable through logical reasoning, then they can understand the thing. This dependence on logic is inseparable from man's ambition to understand the whole world in a broader and deeper way, and this dependence actually adds an obstacle to science, especially medicine. Ironically, pure reason has become the greatest enemy of progress. "" Biological systems are not the product of logic, but the result of evolution, which is a process that pays less attention to precision. Life does not choose the logically best design to cater to the new environment, but only adjusts on the basis of what already exists. " [15] isn't that what happened? Hawkins just emphasized that logical reasoning played a key role in his theory.

At this point, I can't help thinking of a passage in a letter from Professor Nelson Y. S. Kiang: "every generation thinks they can connect neural mechanisms to high-level behavior, but in fact, we still hold on to some straw, but think that it will form the basis of skyscrapers. Scientists have to be patient."

From the author's point of view, intelligence is like complex concepts such as mind, consciousness, complexity, information and so on. It is difficult to give an all-inclusive and universal definition of the meaning that people refer to when they mention these nouns on different occasions. The appropriate thing to do is to state clearly what you mean when doing work in this area. As Professor Wang Pei did when discussing intelligence: "intelligence is the adaptability of an information system when knowledge and resources are relatively insufficient" [16] as the starting point of his work definition for developing his own Nath theory on general artificial intelligence. Different people can use different job definitions, and whether it is useful depends on whether useful conclusions can be drawn from it. Hawkins' model-prediction framework is only one of many working definitions of intelligence (although Hawkins believes that his framework is not only one aspect of intelligence, but intelligence itself). Although this is an interesting definition of work, whether it can be recognized in the end depends on what interesting results can be obtained from such a definition, whether it is to solve engineering problems or to clarify the brain mechanism.

Frankly speaking, the author thinks that Hawkins' thousand-brain intelligence theory is not the theoretical framework of neocortex work at all, and the machine designed according to this is not the product of reverse engineering of the brain. Of course, no matter whether his understanding is correct or not, his ideas can always be inspired by brain research, but whether the machine invented according to this inspiration is really useful or not is still uncertain. Only practice is the only criterion for testing truth. Let's wait and see if Hawkins can really invent a really useful intelligent machine like his Palm PDA.

Thank you for having a useful discussion with Professor Wang Pei, Dr. Karl Schlagenhauf, Professor Liang Peiji and Professor Yu Hongbo during the writing process.

reference

Mountcastle V (1978) An organizing principle for cerebral function: The unit model and the distributed system. In The Mindful Brain, edited by Edelman GM and Mouncastle VB, 7-50. Cambridge, MA: MIT Press, 1978

[2] Hawkins J with Blakeslee S (2004) On Intelligence. Levine Greenberg Literary Agency,Inc.

There have been two Chinese versions: Hawkins and Braks Li, he Junjie et al. (2006) the Future of artificial Intelligence. Shaanxi Science and Technology Press; by Hawkins and Blakeslee, translated by Liao Lu and Lu Yuchen (2022) New Machine Intelligence, Zhejiang Education Press. The author has reservations about the Chinese translation of these two titles, no matter from the original title or the content of the book, it is more appropriate to translate into "on intelligence" or "on intelligence", but of course it is not enough to attract attention.

[3] Hawkins J (2021) A Thousand Brains: A New Theory of Intelligence. Basic Books.

Hawkins, Liao Lu et al. (2022) Thousand brain Intelligence, Zhejiang Education Press.

[4] Freeman WJ (1999) How Brain Make Up Their Minds. Weidenfeld & Nicolson.

[5] Kandel ER et al. (2013) Principles of Neuroscience. The McGraw-Hill Education.

Horton JC, Adams DL (2005) The cortical column: a structure without a function. Philos. Trans. R. Soc. Lond. B Biol. Sci. 1456: 837-862. (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1569491/)

[7] da Costa NM and Martin KAC (2010) Whose cortical column would that be? Frontiers in Neuroanatomy. Doi: 10.3389/fnana.2010.00016

[8] Rockland AS (2010) Five points on columns. Frontiers in Neuroanatomy. Doi: 10.3389/fnana.2010.00022

[9] Buxhoeveden, D. P. and Casanova MF (2002-05-01). "The minicolumn hypothesis in neuroscience". Brain. 125 (5): 935-951. (https://doi.org/10.1093%2Fbrain%2Fawf110)

[10] Jones, Edward G. (2000-05-09). "Microcolumns in the cerebral cortex". Proceedings of the National Academy of Sciences. 97 (10): 5019-5021. (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC33979)

Gu Fan and (2017) three pounds of Universe and Magic mind, Shanghai Science, Technology and Education Press.

Wikipedia (2020) Place cell. "https://en.wikipedia.org/ w / index.php?title=Place_cell&oldid=954316261"

[13] Abbott A (2018) Bat man. Nature 559PAR 165-168

Wikipedia (2020) Grid cell. "https://en.wikipedia.org/ w / index.php?title=Grid_cell&oldid=951093680"

Barry, Zhong Yang et al. (2020) pandemic Influenza: the epic of the deadliest plague. Shanghai Science and Technology Education Press

Wang Pei (2022) outline of Intelligence Theory. Shanghai Science and Technology Education Press

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.