The concept of a “self” is central for understanding suffering, but it is highly complex. Some aspects of self were already considered in Chapter 6. In this chapter, I consider another central aspect of self, related to control. Self can be seen as the entity that is in control of actions, including control of cognitive operations inside the agent, or, to put it simply, in control of the mind.
In preceding chapters, we have seen cases where the mind seems to be difficult to control, due to automated interrupts and wandering thoughts. Here, I consider a general cognitive principle that explains why control is limited. The idea is that in the case where the information processing is parallel and distributed, it is difficult for any single part of the agent’s information-processing system to be in charge of the whole system, e.g. the whole brain. This massively parallel and distributed nature of the brain thus creates most of the uncontrollability in the human mind. The lack of control considered here can also be seen as a generalization of dual-process nature of the mind considered in earlier chapters. Here, there are not just two processes competing for control, but a great number of them.
These considerations necessarily lead to the question of free will: can an AI, or even a human, actually have free will—and what does that mean in the first place. From the viewpoint of the theories of perception in the preceding chapter, we can ask if perception of control and free will are simply illusory perceptions, thus providing another link between the uncertainty of perception and uncontrollability. Such considerations have lead some philosophers to propose that there is no self, or no doer of actions, and I will revisit these ideas from a computational viewpoint.
Suppose you just raise your arm—you can physically do it while reading this if you like. You probably think it was you who decided to raise the arm, and it was you who actually executed the action. You felt being able to control the world, or at least your arm in this case.1 This “you” that first controlled your mind by making a decision, and then controlled your arm, is what can be called the self —in one meaning of the word. The self chooses actions, and controls some aspects of the world, including your inner world.2
However, a number of thinkers have proposed that in fact, “you” are not really in control of anything. A case in point is wandering thoughts. It can be claimed—following a strict definition of the term—that we never want to have wandering thoughts: if we want to think what we are actually thinking, the thoughts are not called wandering. Furthermore, wandering thoughts often feel unpleasant, for example in the extreme case of rumination. So, why do we then continue having them?
A well-known experiment on the control of thoughts is to try to not think of a pink elephant. This is another exercise you can do right now: for a minute or so, do not think of a pink elephant. What invariably happens is that you will be thinking of a pink elephant in spite of your trying not to, or perhaps precisely because of that trying. Clearly, our control of thoughts is limited. In addition, interrupts such as fear, anger or desire capture our mind and direct the processing in ways we might not want. Even habitual behaviour can be seen as a lack of control in some cases: if you mindlessly follow habits, you may end up doing something you would not have done if you had actually deliberately planned your actions.
Lack of control increases suffering in our basic framework of suffering as frustration. Lack of control reduces the probability that the agent reaches the goals it has committed to; it cannot get the things it wants, or avoid the things it is averse to. That means there will be more frustration and reward loss. In fact, the very existence of suffering can be seen as a form of uncontrollability, since if you could really control your mind, you would probably just switch off any feelings of suffering.
In philosophy, the idea of lack of control and its connection to the self goes back to, at least, the Buddha’s times. In a famous discourse, he explained why there actually is no such thing as “self”. He started his refutation by considering the human body, saying3
[I]f the body were self, the core of our being, then it would not tend to affliction or distress, and one should be able to say of it, ’Let my body be thus (in the best of conditions); let my body not be thus (in a bad condition).’ It should be possible to influence the body in this manner.
He continued by going through different aspects of the human mind (perception, thinking, etc.), and denying that any of them could be called the self either, since none of them can be properly controlled. For example, “no one can wish for and manage thus: ’Let my perceptions be thus, let my perceptions be not thus’ ”. If you smell something disgusting, you cannot just decide not to smell it.
Thus, originally, the Buddha framed the very concept of self in terms of control: self is what is in control.4 Since, as he argues, there is actually no (or little) possibility of control, there can be no self. Realizing this is thought to be essential to reduce suffering.5
In ancient Greece and Rome, the Stoic philosophers had similar ideas. Perhaps the very core of Epictetus’s philosophy is contained in his attitude towards control:6
Some things are in our control and others not. Things in our control are opinion, pursuit, desire, aversion, and, in a word, whatever are our own actions. Things not in our control are body, property, reputation, command, and, in one word, whatever are not our own actions.
Epictetus’s idea of uncontrollability is more limited: we cannot control what others do or think about us, or, in line with the Buddha, our bodies. But in stark contrast to the Buddha, he thinks we can control our thoughts and feelings, including desires and aversion. Presumably, Epictetus did not practice the same kind of meditation as the Buddha, which might convinced him of the uncontrollability of thoughts and feelings. In any case, both philosophers advocated recognizing how little control we have as a means of reducing suffering—we will discuss such practical implications in Chapter 14.
We seem to actually have two different kinds of uncontrollability here. First, the uncontrollability of the outside world as emphasized by Epictetus; and second, the uncontrollability of the mind as emphasized by the Buddha. The uncontrollability of the outside world is easy to understand, and its causes are rather obvious. The agent has limited strength: it probably cannot lift a mountain. It has limited locomotion: if it is designed to move on wheels, it probably cannot fly. If it lives in a society, it has limited means of influencing other agents.
What is less obvious, and my focus here, is that there seems to be so much uncontrollability regarding the mind. We have already seen examples where control of the mind is lacking, as in the case of interrupts and wandering thoughts; the dual-process structure of the mind creates further conflicts and reduces control. Therefore, the question arises whether there is some general principle behind all of those manifestations of uncontrollability.
The basic idea here is that the lack of control of the human mind is fundamentally based on one property of the brain: parallel and distributed processing. That is, there are many processors, or neurons, processing the information at the same time, and to some extent independently of each other. If there are many processors working independently, each of them cannot be in control of the agent’s actions: there has to be some kind of arbitration, at the very least. Modern AI also uses such parallel and distributed processing, in particular in the form of neural networks. Both the brain and neural networks in AI are in this way fundamentally different from an ordinary computer, which typically uses serial processing in a single processor.7
While these properties have been mentioned in earlier chapters, we have not really considered the question of why parallel and distributed processing happens. From a biological perspective, we need to find some evolutionary justification for why the brain is parallel and distributed, and from a computer design perspective, we need to explain why such processing would be useful. Perhaps we can answer both questions if we simply find some fundamental computational advantage in parallel or distributed computation.
Let’s first consider the question of parallel processing from an AI viewpoint: What is the point in using many processors? If you want to speed up your computations, why not just get a single processor which is, say, a hundred times faster, instead of putting together one hundred more ordinary processors that compute in parallel? Obviously, there is a limit to how fast processors you can buy for an AI. Perhaps you need faster computation than what is given by the fastest single processor available today. That is why all the supercomputers in the world are highly parallel; they are collections of thousands of processors. That is the only way to increase the computational power to record-breaking extremes.
On the other hand, if you’re really lazy, you might be tempted just to wait. We all know that the technology behind the processors has been developing at an enormous speed. The famous Moore’s law states that the computing power of a processor doubles every two years. This may lead to the impression that there is really not that much reason to go through the trouble of parallelization: if the fastest processor is not fast enough, just wait a few years, and it will be. If this logic were true, it would also mean that there may not be any fundamental reason why computation in AI needs to be parallel, since the power of a single processor seems to grow exponentially and without limit.
Yet, there are fundamental reasons why really efficient computation may not be possible at all without parallel computation, and why, in fact, Moore’s law is not true anymore. One reason is that making processors faster is to a large extent driven by making them smaller. A smaller processor means shorter delays in transmitting the information inside the processor. Such miniaturization cannot go on forever because at some point, you get too close to the level of single atoms, and the laws of physics basically change in the sense that quantum phenomena start appearing.8
A more practical problem is that due to complicated physical phenomena, faster single processors use much more energy than a set of slower processors with the same total computational capacity.9 Energy is obviously expensive and cannot be used in unlimited quantities. Moreover, such an increase in energy consumption has another, surprising effect, which is that the processors heat up very quickly, and keeping processors cool is increasingly becoming a problem. If you design a new processor which is ten times faster than your current one, the power consumption and the heat generated are usually much more than ten times larger.
So, these are convincing reasons why it is necessary in AI to use many processors in parallel. In fact, the speed of a single processor (“clock rate”) even in mainstream computers has not been really increasing since around 2005. The overheating problem became so serious that faster processors became impractical to use.10 Since splitting the computations into many processors generates less heat, manufacturers started putting together several processors on a single chip— the processors are now called ”cores”. The number of cores in an ordinary computer is still small, though, so this is very far from the massively parallel case seen in the brain.
Thus, the great promise of parallel processing is that it can be much faster than serial processing, given the same budget of energy, or, indeed, money. However, there is a problem. If you have one hundred processors that process the same information at the same time, the processing could be, in principle, a hundred times faster. But that only happens in an ideal scenario which requires that the computations are such that they can be parallelized, i.e. they can be simultaneously performed on one hundred processors without any problems. Some problems can easily be parallelized, while others are more difficult, perhaps even impossible. Programming parallel systems needs special algorithms, as well as specialized expertise.
Consider a problem of finding a small object, say a single very black pixel, in an input image. (Suppose for simplicity there is only one such object in the image). You could have a single serial processor scanning the image pixel by pixel. That might take, say, 100 microseconds (one microsecond being one-millionth of a second). On the other hand, if you have 100 processors, you could split up the image into 100 regions, and tell each processor to search for the pixel in one of the regions, and then report to a central processor whether it was there or not. That should not take much more than 1 microsecond. This problem is easy to parallelize, and the speed-up (100x) is basically the same as the factor by which you multiplied the number of processors (100x). A neural network, whether in AI or in the brain, can do such computations massively in parallel, and thus incredibly fast. This is one of the reasons for the impressive behaviour of the human visual system, and the success of neural networks computer vision tasks.11
Then there are tasks that are really difficult to parallelize. This is generally the case when you need to compute an intermediate result before proceeding further. As an intuitive example, consider building a house with rather traditional methods. You first have to build a foundation, and let it dry. Then you build the walls, and finally, set the roof. Suppose you had an unlimited number of builders that you can use; telling them what to do is like trying to parallelize computation. Now, the problem is that you cannot meaningfully divide the builders into three teams so that one of them sets the roof at the same time as another group lays the foundation! Also, if you really have a huge number of builders, they would not even fit on the building site. So, parallelization can be tricky.
Optimization by a gradient method is an example of something that is typically considered difficult to parallelize because you need to do it step by step. Yet, a lot of effort has been spent in computer science research to figure out methods that enable parallelization of such algorithms, sometimes quite successfully.12 With a lot of intellectual effort and ingenuity, it is possible to parallelize even seemingly impossible problems. However, such parallel methods are quite complicated, and thus parallel computation in gradient methods is not currently very often.
The fact that some computational problems are hard to do in parallel while others can be parallelized very efficiently is part of the reason why ordinary computers and the brain are good in very different things. The brain is particularly good at vision, for example. Vision can be rather easily parallelized, as was seen in the simple pixel-finding example above, and indeed the best AI solutions to vision have imitated the brain using neural networks. On the other hand, ordinary computers are very good in logic-symbolic processing, as discussed earlier.
But what is the evolutionary import of these considerations—does it make sense to claim that the brain is massively parallel because of the above-mentioned reasons related to the clock-speed of processors? Certainly, the constraints in building an intelligent system with biological hardware are very different, and the logic above may be mainly relevant for AI. What it actually shows is that progress in AI seems to need computers which are more and more similar to the brain. Yet, it is possible that the massive parallelization in the brain might have some relation to the energy-efficiency considerations that we just saw.
The second question is why distributed processing is needed. Distributed processing is different from parallel processing in that the emphasis is on different processors working independently with as little communication as possible. Distributed computing is important, even necessary, simply because communication is often quite expensive. In the brain, most of the volume actually consists of white matter, which is nothing else than “wires” (called “axons”) connecting different neurons. Those wires take up much more space than the actual processing units. So, the sheer space available in the head strongly limits the connectivity of brain neurons.13 In addition, communication consumes energy which is, again, another limiting factor.
What makes achieving full connectivity particularly difficult is that the number of possible connections between processing units grows quadratically as the number of processing units grows. If you have a million processors, and you want to build connections between all the possible pairs, you need almost a trillion (1,000,000,000,000) wires (assuming each wire can transmit information in one direction only, as happens in the brain). So, the amount of connections easily becomes a limiting factor, and it is important to do the computations needed using minimal information transfer between the processing units, by judiciously designing the way the different areas are connected with each other.14
This is the central point about distributed processing: When communication between the processors is expensive, special solutions are needed. In AI, there is a thrust to distribute AI computation to smartphones that collect the data in the first place, so that the amount of data they transmit to each other or any central server would be minimized.15 In the brain, part of the solution is that processing is very clearly distributed on the level of large brain areas. There are areas responsible for processing visual input, areas for processing auditory input, areas responsible for moving the muscles, areas for spatial navigation, and so on. Each of these areas does its computations relatively independently. That is possible partly because they get different input (visual vs. auditory), and partly because they need to solve different computational tasks (object recognition vs. moving muscles). The communication between those areas can then be strongly limited, and less wiring is needed.
Distributed processing will create its design problems, just like parallel processing. Some tasks are easy to distribute over processors, while others are less so. Again, neural networks are an example of processing which is highly, even massively distributed, and clearly works well in applications such as sensory processing of images and sounds. Considering the example of finding a small object in an image described above, it is easy to see that the computation described is also strongly distributed since the 100 processors each get their own input and then do their computations with no communication between them needed.
The logic above suggests that sophisticated intelligent agents may have to be a collection of relatively independent parts or processors–and that is certainly the case in the brain. The resulting computing system is very different from the view we intuitively have of ourselves. We tend to think of ourselves as serial processors because much of our inner speech and conscious thinking is serial. Speech is inherently serial because the words follow one after another in one single “train” of thought. But such introspection, based solely on what we can consciously perceive, is quite misleading.
A simple metaphor for illustrating the counterintuitive properties of a parallel and distributed system is the “society of mind”: the different mental faculties are compared to human individuals that together constitute a society which is precisely the mind.16 One individual (or processor) is monitoring, say, the state of the bowels, and another one is, independently, responsible for recognizing the identities of faces whose images are transmitted by the eye. Those processors are like human workers with well-defined, separate tasks. Each one may be active much of the time, thus working in parallel. In line with the computational arguments we just discussed, it may also be intuitively clear that it is important that the different individuals mind their own business most of the time, focusing on their own part of the work. Therefore, they only interact if it is really necessary, with minimum communication; thus, the operation is distributed. This metaphor is trying to counteract the intuitive impression we tend to have that the mind is a single, serially processing entity which would be difficult to divide into parts.
Now, to see the point regarding control, consider whether it is possible that one of the independent processors is actually in control of all the others. Psychological theories often use the term central executive for that part of the mind which is supposedly in charge, controlling the rest.17 At first sight, having such a central executive sounds like common sense. The brain has many sensory processing systems (vision, audition, etc.), it can send commands to a multitude of muscles to execute actions, and above all, it has complex information-processing capacities in terms of planning and learning. It would seem that such a system must fall into complete chaos unless there is one area which controls the others. That would be the central executive, a brain area that controls all, or at least most, of the other areas. It would integrate information coming from them and, in return, send processed information and commands to each of them. In the society of mind metaphor, this would correspond to a leader of the society that tells all the individuals what they should do.
It could be argued that having a single area to control all the others is to some extent in contradiction with the whole point of distributed and parallel processing. The central executive would need to have particularly great processing power, and it would need to receive a huge amount of information from all the other parts of the whole system. Thus, both the two bottlenecks discussed above, processing speed and communication capacity, would resurface—but we will see below that this is not really the case.
Designing such a system with a central executive is not very different from designing different decision-making systems in a human society or organization. If there is a single leader, she must inevitably delegate a lot of power to others (say, ministers) in order to reduce the processing power needed by herself. Then, the leader is strongly dependent on the information passed on by the ministers; the leader does not have enough time to make decisions on all the details. So, the power of the central executive is limited due to the limitations on the computational power of a single processor.
On the other hand, if there were a central executive, what about wandering thoughts, emotional interrupts, or habits? Is the central executive just watching when the whole system is hijacked by the fear elicited by the sight of, say, a spider? We argued in earlier chapters that emotional interrupts are useful for evolutionary purposes, so the leader might actually not be very unhappy about that. But interrupts, by their very nature, cannot be prevented, not even by the central executive. Is there any point in calling such a leader the central executive if she is not really controlling everything that happens? What if you eat chocolate because you have a habit of doing it every day (in addition to an irresistible desire, perhaps), even though one part of you knows it is bad for you in the long run—who actually made that decision?
This logic has led many to the proposal that in the human mind and brain, there is no central executive, or, metaphorically speaking, the society of mind has no leader. That is, there is no part in the mind that controls the rest, nothing that controls everything else that happens in the society.18 The society is fundamentally a collection of relatively independent actors. This means very concretely that there is no particular part of the mind or the brain that would control our thoughts, feelings, or desires: they just come and go depending on a complex interaction between different brain areas. Each part of the mind can propose its own mental actions. One part of the visual system might tell the motor cortex: “Let’s move the eye gaze to the right since there seems to be something interesting there”, but at the same time, the replay system might insist on replaying a past episode while ignoring whatever may be happening in the outside world. The result may be a bit chaotic, and having, say, wandering thoughts would not be surprising. To the extent that we define the self as the central executive, there would be no self, in line with Buddhist philosophy.
While such a philosophy is fascinating, it has to be pointed out that there are also neuroscience results claiming that some brain regions in the prefrontal cortex are actually the central executive.19 Moreover, in the design of distributed computing architectures, it is well-known that having some kind of a central processor actually makes communication easier. The point is that there is a good compromise to be found between the two extremes of completely distributed computation and computation in a single processor. Such a compromise can in fact be found in computation which is mainly parallel and distributed, but, crucially, includes a central processor that coordinates the computation, which is still mainly performed by the other processors. In the example above, with a million processors, we saw that a fully distributed system might need a trillion wires to connect all the processors with each other. But suppose that all the communication happens through a central processor, which further selects and processes the information to be transmitted to each of the other processors. Then, all that is needed is wires from each processor to the central one and back (figuratively called a “hub-and-spoke” architecture), which means about two million wires, enabling a reduction by several orders of magnitude. Still, the computational power of the system need not be restricted by the central processor if it is skillfully designed to “delegate” the hard computation to all the processors and only take a coordinating role. Such architectures are currently of great interest in artificial intelligence.20
In fact, the whole dichotomy between a powerful central executive and no central executive is a bit artificial. There can be varying degrees of control that a central executive is able to exercise. While it is not possible to say much with certainty on this topic, the reality in the brain may well be that there is a relatively weak central executive that controls some things to some extent, perhaps many things to a limited extent, but it does not control everything. It may be in control a lot of the time, but not when emotional interrupts, wandering thoughts, or similar processes take control of the mind. Thus, while parallel and distributed processing is inherently without central control, it may be advantageous to introduce some limited form of central executive, and this may turn out to be the best description of what happens in the brain.21
Yet, what is undeniable is that I clearly feel that I can control my body and do things such as raising my arm. A central executive is often intuitively assumed to exist based on exactly such a feeling of self, or a feeling of control. But why should we assume that there is a central executive simply because it feels like there is control? The feeling of control is just another form of perception, and as we have seen, perception may not be accurate. Perception follows certain rules outlined in Chapter 10. It is usually based on incomplete information which has to be combined with prior assumptions to arrive at a conclusion, and this conclusion or inference is what we perceive. Mistakes do happen in this process.
The perception of control in the brain seems to be based on predictions—like so many other things in the brain. Every time you engage in any action, your brain tries to predict the outcome of the action. In particular, when the brain sends detailed motor commands to the muscles, it uses an internal model to predict how the limbs should move as a result. The brain then computes an error signal, comparing the predictions with the actual outcome. In humans, small errors in such predictions are actually quite common because of constant physiological changes in your muscles due to fatigue; or it could be that you are holding something heavy in your hand, which increases the force required to lift the arm. Computing the prediction errors is useful since they enable the brain to learn or adapt its motor commands to such changing circumstances.22
Now, if the prediction error is small (the actual outcome of the action is not very different from the prediction), you feel that you generated the action, and you are in control, according to current thinking in neuroscience.23 This is the computational mechanism underlying the perception of whether you are in control. In contrast, if the errors are very large, the feeling of control is disturbed, and various pathological symptoms may arise. You may even feel the arm is being controlled by somebody else (by “them”, or by “spirits”), as typical of some schizophrenic patients.24
Based on his extensive psychological experiments, Daniel Wegner25 proposed a related theory: the perception of control is simply based on one part of your brain observing a correlation between two things, which are the formation of an intention to act (intention being used here in the ordinary sense of the word, not in the AI sense as usually in this book) and the action actually taking place. If the action happens soon enough after the formation of the intention and the action happens as you intended (and you cannot explain the action in any other simple way), the brain concludes that “you” actually performed the action out of your own “free will”. A strong correlation between intentions and outcomes is not very different from small prediction errors, and thus this psychological theory is very much in line with the neuroscience results cited above. Interestingly, just like visual neuroscientists who construct optical illusions, Wegner then devised clever experiments where the perceptual system makes the wrong conclusion about control, thus showing that the feeling of control can be fooled like any other perception.
Any of these computations are actually quite simple and could be easily implemented in a robot. A robot can assess whether it is able to control its arm by comparing the results of its motor commands and the actual outcome. Suppose some kind of central processor sends a command to the joints in the arm that the arm should be lifted by 10 cm. A couple of seconds later, the input from the camera (or input from specialized sensors in the joint) tells the central processor that the arm was, indeed, lifted by 10 cm. The central processor then concludes that it is in control of the arm.
This logic demystifies the concept of control, which is no longer anything deep or philosophical. The perception of control by the robot above is due to computations of a rather practical nature. In fact, any agent should have a model of what parts of the world it can control (e.g. its limbs) and which parts are outside of its control (e.g. mountains). This is in contrast to my everyday perception that it is I, or myself, that is in control, which is the result of a very complex inference process, and possibly exaggerated, misleading, or even false and illusory. Our everyday perception of control by ourselves is, therefore, no proof for the existence of a central executive, or “self”, that controls actions. One could say that our perception only indicates that there is control in the simple sense of the limbs moving as expected, but it does not necessarily mean that there is any particular entity that is in control. In other words, our feeling of control simply means that certain systems are working in a predictable way, correctly and in harmony with each other; in particular this is about the decision-making system, the motor system, and the actual limbs (or “actuators” as they are called in robotics).
Free will is a celebrated and highly controversial concept in Western philosophy—the idea that you decide your actions “yourself”, that is, your actions are not merely a function of external circumstances, such as your past or other agents. Free will is very closely related to control and feeling of control: most of neuroscience uses the terms almost interchangeably. There are some nuances, though: talking about free will emphasizes your capacity to decide what you will try to do, while talking about control emphasizes your ability to actually do it, i.e. change the state of the world. A very clear difference is, moreover, that free will is almost always considered a conscious phenomenon, while control need not be, as we saw above. Even a completely unconscious robot would benefit from knowing which events are due to its own actions and which parts of the world it can control.
Philosophers have been debating about free will for hundreds of years. Democritus claimed already around 400 BCE that everything, including humans, consists of atoms, and follows strictly deterministic causal laws, thus excluding any free will. A bit earlier in India, the Buddha had debates against philosophers who held similar, strictly deterministic views.26
In a famous series of experiments, Benjamin Libet recorded an EEG response that is known to precede any action decision. The results showed that conscious experience of the decision started up to half a second after the beginning of the EEG response. From this, it is tempting to conclude that consciousness cannot cause the action decision, and hence there is no free will. The EEG presumably measured some unconscious processes which started the decision-making process long before any involvement by conscious processes. Libet’s own interpretation, though, was that consciousness could still participate in the action decision by having the possibility of “vetoing” any decision that the unconscious circuits were trying to implement. This would imply some weak version of free will, but his interpretation is controversial.27
Some would argue that denying free will may be dangerous: people have to believe in free will in order for our moral systems to work. If people don’t believe in free will, they might not feel they have any moral obligations and might behave just as they please. Our justice system in particular is based on the idea of free will: if it can be proven in court that a murderer acted without free will, say, because of a brain tumour, that will usually lead to a reduced sentence. This is, of course, not saying that there is free will, just that it may be useful to think that there is one.28
One of the most influential psychologists in the 20th century, B.F. Skinner, had a more computational viewpoint. He thought human behaviour is simply determined by rewards and punishments. That is also where the “moral” behaviour comes from; no special metaphysical beliefs are necessary. Reward people for good behaviour and punish for wrong behaviour; that is all that is needed to make them follow moral rules, in Skinner’s view. From the perspective of this book, I can partly agree with Skinner on the importance of learning from the environment, where learning, again, includes evolution.
Let’s go back to the “no-self” quote by the Buddha which we saw earlier (🡭). In Buddhist philosophy, it is the historical basis of a celebrated doctrine claiming that there is no such thing as the self. This is clearly a much more general idea than the mere claim that there is no entity which is in control, but in this chapter, we focus on the aspect of control and free will. We can now recapitulate the ideas in this chapter in view of justifying some claims regarding the existence of self.
First, if the brain, and thus the mind, is composed of many different processors all working simultaneously and to a large extent independently of each other, how could we speak of a self? If we admit there is no central executive, that is one form of “no-self”: there is no particular part of the mind that actually is in control and could be called self in that sense. (It is not clear if it is neuroscientifically correct to deny the central executive, but let’s admit it for the sake of argument here.) The conscious part of the mind does not control actions according to neuroscientists such as Libet and Wegner, thus contradicting our everyday perception that we decide actions on the conscious level. Decisions seem to be actually taken by various unconscious neural networks, and it may be difficult to point out any single entity making the decision.
Some parts of Indian philosophy actually formulate a more specific doctrine of “no-doer”, which means that there is nobody that “does” anything in terms of taking the actions—or that at least, it is not “you” that does anything. Instead of “you” making conscious decisions and being in control, your body and mind are constantly on some kind of autopilot, and your consciousness is merely observing it all.29 The points above give some credence to such a variant of no-self philosophy.
But we may go further. If it is not your conscious self that decides, is it even necessarily your neural networks? In our framework, we could say that control is ultimately exercised and decisions are ultimately taken by the input data that the agent learns from. Our computational models have assumed that our actions are determined by past input data, together with the design of our learning and inference machines—even though the mapping from input data to action can be extremely complex and impenetrable.30 From this viewpoint, nobody at all is in control, and there is no free will—while it is possible to say that there is some control in the specific sense of sufficient predictability. Furthermore, the existence of a central executive in the brain becomes irrelevant: if it exists, it is still only a vehicle for the evolution and the input data to steer our thinking and behaviour. Even a brain with a strong central executive could be seen as having no self: even though a central executive was seen as the hallmark of a self earlier in this chapter, if we now see all its actions as simply following from learning based on input data, it may not actually qualify as a self.31
Ajahn Brahm, a famous meditation teacher, once said that when he sits down to meditate, he always remembers the instructions of his own meditation teacher in his head; thus, it is not really Ajahn Brahm who meditates, it is his teacher—or, if I may, it is the input data he received from his teacher.32