2.1 The early paths of Intelligence in Neuroscience

Understanding intelligence goes through a path of overcoming. Since the 1970s, when Francis ^{Crick (1979, pp. 219-233)} wrote an article called Thinking About the Brain, considerable data about brain works were already available. For Crick, the brain was like a puzzle where we have a reasonable amount of data, but we don’t know how these data are connected and interconnected among themselves (^{Hawkins, 2021, p. 16}). One of the problems arising from this correlation is how intelligence is made in the brain.

^{Hawkins (2021)} notes that to understand how intelligence is made is necessary to locate the appropriate brain section for such a task. That is: where is intelligence produced? According to Hawkins (2021, p. 25)³, the neocortex, present only in mammals, is the region responsible for the reserve and production of all human intelligence. “Almost all the capabilities we think of as intelligence —such as vision, language, music, math, science, and engineering— are created by the neocortex. When we think about something, it is mostly the neocortex doing the thinking” (^{Hawkins, 2021, p. 25}). Hawkins understands Intelligence as the capacity to understand the world, such as the shape of objects; the behavior of creatures; how doors open and close; and our location and position in the world as human beings.

Interestingly the neocortex alone does not exert direct control over behavior or muscle movements (^{Yokoi & Diedrichsen, 2019}). Breathing and other basic movements will be made by the old brain, which is also responsible for the most primitive functions of human beings, such as survival, sex, and proliferation. But unlike the brainstem, where there is an almost perfect cross-section of its parts, the neocortex has different zones of action, with their respective subdivisions, but there is no possibility of visualizing them clearly as in the brainstem.

For example, ^{Felleman and Van Essen (1991)} show how the neocortex of a primate works, where nerve synapses were detected from stimuli made by the researchers. Each rectangle in the image represents different regions of the neocortex, and the lines represent the direction of the nerve synapses and how information travels from one side to another.

Source. ^{Felleman and Van Essen,1991}.

Figure 1 Connections in a Primate’s Neocortex 

We can have an idea that the working of the neocortex is similar to that of a flowchart whose interpretation would be based on the hierarchy and positioning of each brain zone. However, the image itself shows that the flow can go either horizontally or vertically. Although its external appearance is uniform, ^{Ramón y Cajal (1923)}, a neuroscientist of the 1800s, was responsible for demonstrating how the nerve cells present themselves in the neocortex (^{Gil et al., 2014}). The result of his research, which can be seen below, earned him the Nobel Prize.

Source. ^{Ramón y Cajal (1923)}

Figure 2 A section inside of a little mammal neocortex 

The results brought by Ramón y Cajal showed that the neurons present in the neocortex are arranged in six layers (^{Ogawa et al., 1995}), as well as having axons and dendrites, responsible for the nerve connections called synapses (^{Hawkins, 2021, p. 29}). But this observation can only be made if seen from the inside. The neocortex looks the same and its regions cannot be distinguished from the outside. But what is the reason? ^{Edelman and Mountcastle (1978)} proposed that its similarity is because the whole neocortex is performing the same activity simultaneously, so there would be no reason to differentiate it. In this way, by discovering how a small region works, we could discover how the whole system works as well (Hawkins, 2021). But the authors’ main merit was to verify that intelligence can manifest itself as an algorithm in the cerebral cortex through a cortical column.

Cortical columns are not visible under a microscope. With a few exceptions, there are no visible boundaries between them. Scientists know they exist because all the cells in one column will respond to the same part of the retina, or the same patch of skin, but then cells in the next column will all respond to a different part of the retina or a different patch of skin. (^{Hawkins, 2021, p. 37})

^{Edelman & Mountcastle (1978)} propose that the neocortex has about 150,000 cortical columns. The assumption is made because they cannot be seen even with microscopes. Hawkins says that scientists know that they exist because of the response of other cells to stimuli made between them, but they remain a mystery in the realm of neuroscience.

2.2 A Model of World in Your Head: Cortical Columns and Reference Frames

Hawkins’ assumptions, however, become necessary because it is precisely in this place where the development of intelligence is assumed. Our brain reacts to stimuli made by sensors (hands, eyes, mouth, etc.) to produce expected and unexpected results. These results, however, are not random but have a correction between what you and I expect to see and what you and I expect to feel when we visualize and touch something. Every time neocortex makes these predictions to understand and map the world around it. It is from this context that the brain learns what is normal, what is abnormal, and what can be expected based on experience. As a result, you get what ^{Hawkins (2021)} calls A Model of World in Your Head.

The brain creates a predictive model. This just means that the brain continuously predicts what its inputs will be. Prediction isn’t something that the brain does every now and then; it is an intrinsic property that never stops, and it serves an essential role in learning. (^{Hawkins, 2021, p. 42})

If a prediction is made erroneously, the brain will try to make the correction and conform to the expected or, as a last resort, learn not to replicate the behavior as a consequence (^{Konflanz, 2019}); (^{Byrne & Corp, 2004}); (^{Deaner, Isler, Burkart & Van Schaik, 2007}).

A newborn baby does not have any kind of experience or model of the world in its neocortex. “When you are born, your neocortex knows almost nothing. It doesn’t know any words, what buildings are like, how to use a computer, or what a door is and how it moves on hinges. It has to learn countless things” (^{Hawkins, 2021, p. 44}). It is through its development and the acquisition of experience that a model of the world will develop in the neocortex of each individual. This model will be responsible for actions, predictions, and perceptions. From this observation, Hawkins draws two observations: 1) the world is constantly changing and, consequently, intelligent beings must adapt to this change; and 2) as we move, our conception and model of the world change every time we move around because brain inputs provide us with new data and information to implement the experience so far lived. It is in this way that sensory-motor learning develops.

According to ^{Hawkins (2021)}, the brain learns a model of the world through our observation, a result of how our senses and sensors interact with the world and the information that is perceived by us. From this assumption arise two principles of neuroscience: Principle number one: thoughts, ideas, and perceptions are the activity of neurons, and; Principle number two: everything we know is stored in the connections between neurons. When we learn something new, connections between neurons are reinforced. When we forget something, these connections are weakened, and, interestingly, new synapses are not only (re)forced, but also created and undone with time and experiences acquired by the individual (^{Li et al., 2017}; ^{Kim and Thayer, 2001}).

For these reasons, understanding the brain as a machine or a computer does not seem intelligible. The neocortex responsible for learning the world as a structured model becomes an adequate idea due to its verification. Hawkins demonstrates that the brain can make predictions and forecasts due to the existence of two types of neurons: neurons that fire when the brain is seeing something, and neurons that fire when the brain is predicting that it will see something. The brain needs to keep its predictions separate from reality to avoid hallucinations. Using two sets of neurons does this very well. However, for Hawkins, there are two problems with this idea.

First, given that the neocortex is making a massive number of predictions at every moment, we would expect to find a large number of prediction neurons. So far, that hasn’t been observed. Scientists have found some neurons that become active in advance of an input, but these neurons are not as common as we would expect. The second problem is based on an observation that had long bothered me. If the neocortex is making hundreds or thousands of predictions at any moment in time, why are we not aware of most of these predictions? If I grab a cup with my hand, I am not aware that my brain is predicting what each finger should feel, unless I feel something unusual—say, a crack. We are not consciously aware of most of the predictions made by the brain unless an error occurs. Trying to understand how the neurons in the neocortex make predictions led to the second discovery. (^{Hawkins, 2021, p. 52})

Furthermore, ^{Hawkins (2021)} finds that predictions are made within neurons. Since predictions can appear in two ways, one while the world is changing around us and the other because we are moving with the world, neurons need to figure out how much context is needed to make proper predictions.

A prediction occurs when a neuron recognizes a pattern, creates a dendrite spike, and is primed to spike earlier than other neurons. With thousands of distal synapses, each neuron can recognize hundreds of patterns that predict when the neuron should become active. Prediction is built into the fabric of the neocortex, the neuron. (^{Hawkins, 2021, p. 57})

But prediction is an ambiguous function of the brain, and neuroscientists know this. Yet even in simulated tests, Hawkins was able to prove that neurons can learn diverse sequences and memorize them. Even with the death of about 30 % of them, the sequence was still memorized by the rest (^{Hawkins & Ahmad, 2016}). The results demonstrate that predictions can happen inside neurons.

But how does access to this information happen? As visualized before, it seems to exist, in an intelligible and acceptable way, a map (analogically) in our heads. For each area of the neocortex, we can fraction it until we get to the cortical columns, which must be understood as referential frames that can respond to input through experience and the stimuli that are presented to it. Our brain doesn’t process a picture or an image: its comprehension starts from the reception by sensors behind the eyes, which are then divided into considerable parts. After reaching its destination, we can understand the location, shape, and even texture of the object observed. This is one of the primary functions of the neocortex according to Hawkins: the processing of reference frames.

But why are Hawkins’ reference frames important for understanding intelligence? First, they allow the brain to grasp the structure of something. Second, by defining an object using a reference frame, the brain can manipulate it entirely at once. Third, referential frames are necessary for planning and creating movements. Referential frames are used in many fields. Scientists working in the robotics program, imagine and plan what the movements of a robot’s arm or body should look like. Reference frames are also used in animated movies to render characters as they move.

With this result, Hawkins was able to assume the difficulty of understanding how the apprehension of meaning in the neocortex takes place. However, as hard as the concept may be, its meaning must be gradual and stored in different points of the brain. Thus, when someone wants to understand the concept of democracy, we must review the concept of State and Government, for example. Each of these concepts is located in a reference frame and, when connected, can build other reference frames aimed at understanding the concept of democracy. The same happens with objective good faith, a concept that presupposes the existence of other reference frames, such as principles, norms, laws, contracts, property, and processes. Each of these assumptions is stored in different referential frames that, when activated by the neocortex, its synapses snap together and build individualized concepts from its previous knowledge. In other words, Hawkins demonstrates that a good background is intrinsically linked to data and facts (^{Hawkins, 2021, p. 86}).

The problem is accentuated when language is included as one of the most important cognitive abilities capable of distinguishing human beings from other animals. Hawkins claims that without the ability to store and share knowledge and experience through language, modern society would be impossible to realize. According to Hawkins, ^{Etard et al. (2000)} and ^{Naeser et al. (1987)}, there are two areas in the neocortex of considerable size that are responsible for language⁴. The section responsible for language comprehension is Wernicke’s, while the section responsible for its production is Broca’s. However, there is no consensus on the exact location and extent of each of these regions. Hawkins also points out that these areas are not necessarily differentiated between comprehension and production, since both can act interdependently and simultaneously with similar results, and that language cannot simply be isolated in two small regions of the neocortex.

It is the experience, behavior, and contact with the world that will be able to fill the frames of reference and create new neurological connections to enable the intelligibility of language and, consequently, to disseminate knowledge. Every moment that information enters through observation and body sensors, the world in our brain undergoes a drastic change: new frames of reference are made by new synapses, and new synapses are undone. Although each cortical column doesn’t necessarily know what they are learning and also doesn’t know what these models represent, individually they don’t make sense, when put into a referential frame they begin to make sense and demonstrate how the brain works through them. It is at this point that Hawkins’ Thousand Brain Theory is developed and presents a seemingly acceptable answer to the AI problem.

2.3. Hawkins’ Thousand Brain Theory

An attempt is made to overcome the idealization of the brain as a flowchart and present it as referential frames. The cortical columns, even without less sensitive levels, can grasp and recognize objects. Thus, the neocortex has several models of a particular object. This model can be in different referential frames, which are not identical, but complementary. This complementarity is essential because according to Hawkins a cortical column cannot learn a referential model of every object in the world. Such a situation would be impossible because there is (yet) an unknown limit to how many objects each one could learn. But when they are connected, even if in different regions, they can recognize what is intended by the brain by accessing the previously learned model.

Thus, knowledge for Hawkins is distributed in the brain. Nothing we know is stored in just a single place, a single cell, or a single frame referent. Knowledge is stored everywhere and distributed in multiple columns. Everything in the brain works interdependently. A neuron does not depend on just a single synapse, but on about thirty to recognize a pattern.

Therefore, we should not be surprised that the brain does not rely on one model of anything. Our knowledge of something is distributed among thousands of cortical columns. The columns are not redundant, and they are not exact copies of each other. Most importantly, each column is a complete sensory-motor system, just as each water department worker is able to independently fix some portion of the water infrastructure. (^{Hawkins, 2021, p. 98})

The problem lies in ascertaining how this information is linked and tied together. ^{Hawkins (2021)} proposes the question: how are our sensory inputs tied to a singular perception? There is no randomness, but a voting system elaborated by Hawkins that accesses information according to its sense and intelligibility and brings, at least, un-derstanding to the individual. If a person inserts his or her hand into a black box with an object inside and begins to feel it, the nerve synapses will begin to call each other and “vote” to determine which object among those you have already touched it best represents. It can be a cup, a glass, a pen, or any other physical object. If you have had previous contact, the neocortex will start and finish this vote-counting system to determine which object is the one touched and felt.

Since the connections in each column go up and down using the layers, remaining largely within the boundaries of each column, Hawkins states that only a few cells will be able to vote. Most of them do not represent any kind of information that the others could opine on. It is not a complicated view, but I prefer to keep his words and his original meaning:

Using its long-range connections, a column broadcasts what it thinks it is observing. Often a column will be uncertain, in which case its neurons will send multiple possibilities at the same time. Simultaneously, the column receives projections from other columns representing their guesses. The most common guesses suppress the least common ones until the entire network settles on one answer. Surprisingly, a column doesn’t need to send its vote to every other column. The voting mechanism works well even if the longrange axons connect to a small, ran-domly chosen subset of other columns. Voting also requires a learning phase. In our published papers, we described software simulations that show how learning occurs and how voting happens quickly and reliably. (^{Hawkins, 2021, p. 104})

With this proposal, ^{Hawkins (2021)} solves another mystery of the human brain: how our perception of the world seems to be stable when our brain is constantly changing in the face of incoming information (^{Tehovnik & Slocum, 2004}). As low as the number of active neurons is extremely low (about 2 % only), they are responsible for understanding the world and maintaining stability in our brain. But no matter how small your attention is directed to an object, event, or person, the neocortex never stops creating learning models, and it is these learning patterns that are ephemeral and long-lived.

Hawkins’ theory seems to be correct and, as much as it is not, it is testable and verifiable (and this is important). Notice that when we deal with the brain and intelligence in its neurobiological terms there is a substantial differentiation from the Neural Network of the previous chapter. Even in a skeptical view, it seems that the computational approach is even utopian and fictitious.

The intelligence assumptions developed by Hawkins are tied to frames that are grasped in a much more complex way than artificial neural networks. Thus, if his theory is correct, I suggest that the future of artificial intelligence will be substantially different from what we have seen before, and what most computer scientists predict. The stance is skeptical, but with a view based on the science developed so far. And it is for this reason that we must analyze whether an AI is intelligent.

3.1 Consciousness and Artificial Intelligence

Another aspect that Hawkins’ theory elucidates and clarifies is the relationship between consciousness, machines, and human beings. Although this is an avoided theme in neuroscience because of its abstraction and because we don’t know what consciousness means, Hawkins still manages to physically explain its aspects.

Consciousness is a process and action (awareness) that can be seen as follows. Hawkins suggests imagining if we could reset our brain to the exact moment we woke up this morning. However, before this reset happened, we would normally do all the things and activities that we were already programmed to do. Then all memories about what we did that day would be erased. Later, after that reset, we would believe that we just woke up. If someone said that we had a snack, breakfast, worked or any other task we would deny it vehemently. Of course, we were conscious while doing these activities. Only the memories were deleted to induce us that we were not. Howe-ver, this demonstrates your sense of consciousness as action (awareness), which many people would call being conscious and requires the moment-by-moment formation of memories about our actions.

Consciousness also requires that we form moment-to-moment memories of our thoughts. Recall that thinking is just a sequential activation of neurons in the brain. We can remember a sequence of thoughts just as we can remember the sequence of notes in a melody. If we didn’t remember our thoughts, we would be unaware of why we were doing anything. For example, we have all experienced going to a room in our house to do something but, upon entering the room, forgetting what we went there for. When this happens, we often ask ourselves, “Where was I just before I got here and what was I thinking?” We try to recall the memory of our recent thoughts so we know why we are now standing in the kit-chen. (^{Hawkins, 2021, p. 133})

For Hawkins, to feel conscious (awareness) is to feel present, to feel that we are an agent acting in the world, and this is the heart and the main idea of what it means to be conscious. If an artificial intelligence could contemplate models of the world equivalent to the states of the nervous synapses of neurons in brains, and it could remember states and events through memorization, it could be considered conscious, in a way. But this is a step that has not been taken so far. Thus, nothing that we call AI today is intelligent.

No machine has the flexibility and ability to shape itself according to its experiences in the world. However, there is no denying its ability to be built in the near or distant future. One of the main obstacles is precisely understanding what intelligence is, which ends up being left aside by computer scientists.

But we must consider that intelligent machines if properly made, will not be equal to human beings. Intelligence is the ability of the system to learn a model of the world. However, our brain is not only made up of the neocortex but also of the brainstem responsible for emotions and other more primitive behaviors. When truly intelligent machines are created, they should follow the unique arrangements of the neocortex, ignoring the emotional part by choice: since we can choose which parts go in and which parts don’t, emotion will possibly be left out due to its complexity. Thus, an AI, des-pite needing the flexibility of human behavior and understanding different models of the world, they do not need the basic instincts of survival or procreation, for example.

The components that Hawkins assumes for the elaboration of an AI, therefore, are the following: embodiment; old brain parts; and the neocortex. Since there is a need for interaction and contact with the world through sensors, embodiment becomes necessary precisely to develop the sense of a world in an AI. However, most ML network does not have an embodiment. There are no sensors attached to the world that can capture inputs and transmit them to the reference frames to develop the world model of AI. Without this embodiment, what can be grasped is limited.

Furthermore, Hawkins assumes the existence of a system equivalent to the brainstem. The purpose would not be to replicate emotions but to move the AI’s body. The neocortex does not directly control the movements. It needs a command from the brainstem to act. The same should happen in an AI. The author questions whether it should be this way or whether we could build intelligent machines that where the system equivalent to the neocortex and would directly control the movements. Hawkins does not believe that because the neocortex needs to be connected to something that already has sensors and already has behaviors. In other words, it is not meant to create new behaviors, but to learn how to connect them in new and useful ways.

Hawkins also posits that an AI must possess goals and motivations. Human beings have complex goals and wills. Some are tied to our genes, such as sexual desires, food, and shelter. But it is important to remember that the neocortex itself does not create these goals, motivations, or emotions. It is only actively involved in how these motivations and life goals influence behavior, but it does not lead and determine it. The brainstem is responsible for this task.

And indeed, there is a sense in your statement. Since we are talking about beings that are remarkably intelligent for their respective activities in the social order, to develop machines without any objective or pretension and implement them would be to echo an electronic void without any correspondence. I cannot visualize any sense in an intelligent being that does not act on its own with the means to pursue and build itself.

Finally, for Hawkins AI must have something equivalent to the neocortex. This is the third assumption that a machine must have because it is considered the storage place of intelligence. But learning should not be equated with copying. As every human being is a blank sheet of paper and the world conceptions are built throughout his life through reference frames, the same should happen with an AI. It is not enough that it is inserted into a platform to replicate data and knowledge previously acquired.