Evolving Intelligence: The Only Proven Method

The following paper is an attempt to circumvent the traditional roadblocks in the development of artificial general intelligence by proposing that researchers use the only method for developing intelligence that has been proven to work: evolution. I argue that the best, if not the only, way to develop artificial general intelligence is over time with a selection process.

The brain is a very complex computer which evolved over billions of years to reach its current state. The complexities of the mind-brain system are so great that they cannot possibly be reconstructed so perfectly as to resemble human thought. Instead, the framework of the mind’s representational system should be an artificial neural network evolved to resemble the brain.

It also must be recognized that many of the complex representations in the matured mind are in some way related to and based upon the physical world, evident by the way infants spend their very critical first six months exploring the world in any way they can (Lobo, Kokkoni, de Campos, & Galloway, 2014). If from birth a child was placed in a sensory deprivation chamber, it would not be expected to become in any way intelligent, yet many AI researchers expect a machine under the same conditions to become intelligent. Thus it is necessary that the machines used to develop AI can experience the world, and have bodies, which will of course be robotic.

Through back-propagation across several generations of intelligent robots, the machines will become capable of greater and more complex tasks. This will continue until the robots have evolved sufficiently to resemble human intelligence. Of course some dissenters would argue that these resulting machines, which will have a representational language of thought (see Fodor, 1975), are not capable of any real understanding. I do not seek to convince such people, as this is a method for creating machines intellectually indistinguishable from humans, not an argument for the consciousness of such machines. However, I expect some of these people to be convinced anyway, as it is quite hard to differentiate in any meaningful way between an organism whose evolved neural networks produce a representational mind and a machine whose evolved neural networks produce a representational mind.

Functionalism: A Goal, not a Tool

It should be clear by now that this paper endorses a type of functionalist, or even computationalist, theory of mind. This endorsement comes simply because such a theory is, as far as I can tell, the best theory to explain thought so far. Unlike other theories, functionalism does not break down when you ask how the brain evolved or what happens when a person makes a choice. Although I would gladly accept any new theory that could better explain the mind without falling victim to such pitfalls, no such theory has been proposed, and therefore computationalism has been endorsed.

While, as I have stated, there will be some objections to the functionalist assumption, I believe that most of my audience will not only accept but also expect that functionalism will be assumed in a paper regarding artificial intelligence. Contrary to most artificial intelligence research, however, it is my belief that functionalism cannot be used as a tool in constructing artificial intelligence. It is my belief that although thought is a representational system, which is the result of calculations in the brain, such a representational system cannot be artificially constructed.

The reason for this should be quite obvious if one considers the implications of the mind being a functional system of representations, specifically the necessary complexity of such a system. How many representations must go into a working concept of what it means to feel sad? What thoughts, memories, and experiences come to mind when one thinks of sadness? What physiological responses come with those thoughts? An artificially intelligent system with human-like cognition must be able to achieve its own unique version of those thoughts, memories, and experiences to have a proper concept of sadness.

And sadness is only a tiny part of what it means to be human. Could anyone ever hope to properly implement every single fragile piece of such an inconceivably complex system without error? How could one ever hope to imprint in a computer every representation needed to be human while avoiding all representations that are non-human? The task is impossible. Consider a billion-piece jigsaw puzzle that comes with five billion pieces, most of which don’t fit into the final puzzle but look very similar to the pieces that do. AI researchers that attempt to use functionalism as a tool to build artificial general intelligence are essentially trying to solve this puzzle, perhaps constructing sections of it, only to realize much later in the process that some said sections are in the wrong place and most of them actually shouldn’t be put together like that at all.

Researcher Rodney Brooks (1991) saw this problem when he began his work in AI. He considers the functionalist method analogous to taking engineers who were trying to figure out flight in the 1890s and bringing them to present day, where they would get to ride on a 747. After being transported back to the 1890s, the engineers get to work reconstructing the plane, perfectly replicating the seats and windows in the cabin, but of course their model won’t take flight. We have a very vague idea of what the mind is exactly, so our attempts to replicate what we see from the ground up will not be successful.

The method of creating artificial intelligence through the construction of a representational system is not viable, but functionalism should not be dismissed, as it is still our best description of what a mind is. Though a different method of developing artificial intelligence must be utilized, functionalism should still be used as a guide because these creations will still use representations in order to think.

Robots over Representations

As an alternative to a representational system, Brooks(1990) has taken to developing intelligent robots. These robots do not use a representational system at all, but rather are built to have one goal, which is survive in the world. This means that Brooks’ robots are built to adapt to unique situations for self-preservation. The idea is that if robots are to be like us they must have the same driving forces as we do.

Robots are ideal (necessary, even) for the development of intelligence. From birth, humans have a constant input of sensory experience. We are used to distinguishing objects by sight, judging the position of something by sound, and feeling the constant pull of gravity towards the earth’s surface. To assume that these constant inputs have no effect on the intelligence of a person is absurd. It is highly unlikely that one’s interaction with the world has no influence on the structuring or development of one’s brain. In fact, our ability to experience and move about in the world very likely plays a key role in how our brains and minds develop. It would be very difficult for someone to gain a concept of a cat, for instance, if one had no way of ever experiencing what a cat is.

I agree wholeheartedly with this stance, and it underlies the whole of my argument henceforth. If a system has no way of experiencing and deriving information from the world, it is impossible for the system to achieve human-like intelligence, because human intelligence fundamentally concerns the world in which the human exists.

Brooks began with robotic insectoids, which used a simple program to control the legs and which began to adapt to their environments. These robots were quite successful, and before long Brooks had his insects developing their own gaits and navigating their way through increasingly complex obstacle courses. For all intents and purposes, the machines created by this work were examples of artificial intelligence - insect intelligence.

After succeeding in robotizing a simpler organism, Brooks decided to “skip robotic lizards and cats and monkeys and jump straight to humanoids”(2008). I do not expect this line of research to find general artificial intelligence for the same reason that it has not been found yet: this method still attempts to instill human intelligence in a machine directly. Though Brooks has a successful history with insects, I would argue that human thought is significantly more complex than that of an insect. A humanoid robot running the insect program would be able to perform only as well as the insectoid robots. Obviously, the program must be modified before a humanoid robot can have human intelligence. But what aspects of the program must change, and how?

The Two Stages of Acquiring Intelligence

It should be obvious from how complex the mind is that human-like intelligence cannot be created at once, but rather must be evolved. Perhaps beings which we would consider superintelligent would be able to figure out and create a human program, much like how we can probably perfect an insectoid program. However, I believe it would be very difficult, if not impossible, for a species to so fully and accurately comprehend its own intellectual properties so well to be able to create a computer program that is identical in intellect.

Rather, artificial intelligence must be evolved over time, much like natural intelligence. In fact, if we take an evolutionary approach to AI, the distinction fades. How can there be both artificial and natural intelligence if they both came about via evolution? Would there not simply be “intelligence”? I believe that society will eventually move towards this new perception of intelligence, focusing more on the capabilities of a system than the material from which it is made.

To bridge the gap between robotic insectoids and robotic humanoids, the foundational software must be evolved. In order to do that, we must have a basic concept of what a complex learning program would probably look like. Most importantly, it must be understood that there are two distinct periods of “learning” for any organism: evolution and classical learning. Classical learning is obvious, and is what is usually meant when one discusses learning. It is the process of an individual taking information from experience and education and incorporating it into their worldview.

The other type of learning is not done by the individual, but by the species. This type of learning is evolution, where the foundation of the mind improves generation to generation. Not taking into account traditional concept learning, it is still clear that something was learned between the time of the humans and the time of our great ape ancestors, for example. The evolution of the mind could be called “learning how to learn” because the major change that takes place is in the species’ ability to accomplish the traditional type of learning. What we’ve learned since the time of the great apes is how to better participate in concept learning. This means that our representational system, most crucially our ability to form representations, has improved.

The type of learning that AI researchers need to focus on is this second type, the learning that allows something to observe the world and form or modify representations in accordance with what is observed. This process, the abstraction from reality to thought, is the key process required to have intelligence.

Researchers can (and have) attempted to construct a system that contains every representation required to have intelligent thought. However, even if such a system were possible to construct, the machine would not have thought, at least not in the sense that we do. This is because the fundamental aspect of intelligence, what separates thinking from non-thinking, is not having a representational system, it is being able to construct and modify representations. For imagine if we took every representation that I have regarding anything and we manage to instill that exact system on a machine. If the machine has all of my representations, it should act exactly like me, speak exactly like me, and so on. However, if the machine was not capable of dynamically altering itself, its representations would never change. Therefore, the machine would continue acting like me… but I wouldn’t. I would be continuously changing my own representational system, adding and enhancing based on my new experiences. In time, I would be quite different from the robot with whom I once shared a system of thought. In this case, one would hardly consider the robot intelligent, as it wouldn’t be able to adapt to or perform in the world.

The key to being intelligent is the ability to learn, and that ability has been developed through evolution, which is a sort of genetic learning process in itself.

The Evolutionary Learning Process

What sets us apart from less intelligent beings is very likely our ability to create and alter representations better than them. Our representations are more complex and greater in number, allowing us to make connections and entertain thoughts that other creatures cannot. If the leading scientific literature is to be believed, humans reached that representational system strictly through evolution. Because it is fair to assume that intelligence didn’t pop out of nowhere, it is a necessary truth that our genetic ancestors were less intelligent than we are, and that our current intelligence was evolved from that lesser intelligence.

During this evolutionary process, something was “learned” by us as a species. The ability to form representations improved little by little until we reached what we are today. This ability must be imprinted in computers in the same fashion if we are to ever create true AI. The key to evolving the system is very likely constructing a neural network and backwards-propagating the system towards something that resembles human thinking (see Churchland, 1990). This seems to unify the language of thought hypothesis with neural networks, where the neural net establishes a sort of medium over which the representational system can function.

The first step in the process must be perfecting the way the computers interact with the environment, as that is achievable by even very simple organisms. Squirrels and sparrows have rather simple minds, but these creatures do not struggle to get around. At the same time, movement is in some sense learned by the individual. We don’t walk out of the womb with a developed gait; it takes time to learn how to use our legs. From birth our brains do have control of our leg muscles, but we don’t yet have an idea of the proper combination of muscle movements that keep us on our feet, let alone walking.

What’s important to note here is that we seem to be born with our physical motor skills intact, but our mental motor skills nonexistent. This makes sense in relation to the computational theory of mind, as we have not yet formed the representations necessary for walking. Likewise, our robots must be “born” with control over all of their parts, but no concept of how to use them to move about in the world. If this sounds familiar, it is likely because this is the approach taken by Rodney Brooks. His insectoid robots could use their legs, but they had to learn and adapt to be able to use them well enough to move about in the world.

Once these motor skills are essentially perfected, the robots can be evolved, but at this time it is still important to focus on the basic features on life. At this point it may be a good time to implement the desire to survive, for instance. This could include the need for it to eat or plug itself in, the desire to avoid obvious dangers, and so on. But how can these complex features be implemented? Again, the robots can acquire them in the same way we acquired the same features, which is the method of selection of the fittest individual of the group. Humans have individuality, and so must our robots. Thankfully, computers are pretty good with randomization. Consider the following procedure:

An insectoid robot is developed that starts with the ability to use its parts, but no concept of how to use them to its advantage. The robot is able to take in some amount of knowledge from the world and conceptualize some aspects of its environment. The robot is capable of simple learning, and with time it swings around its appendages until it is able to push itself around, and then walk. It continues to improve until it has developed a unique gait and can traverse simple obstacle courses. You are happy with this stage of life and would like to move on to develop more features. So, you evolve your insect.

Take the base code that you started with, what the insect was hard-wired with from the start. This is the neural network. By the end of your insect trials there will be a lot more information, what the robot internalized during its lifetime, but you just want the program that started its life. In a sense this neural network is similar to our DNA - it is the hard-wired code that we are born with and is passed down to subsequent generations. Remember, nothing an organism learns in its life goes into its DNA or is passed down. Evolution works because those with unfit DNA don’t live long enough to pass it down, not because the more fit organisms became more fit after birth.

Next, create a plethora of child robots, each using the same neural network as the first, but with slight alterations in the weightings and number of connections in the neural network. These alterations can of course be man made attempts at implementing whatever the next feature is, but it is important that there is also some randomization in this stage. Raise and test each child program in a robot, and select those that best display the desired trait. Those more fit robots can be modified until that stage is satisfactory, and then the most fit version of the program can be given this same treatment of randomization in order to implement the next features, and thus the cycle continues.

How Evolution Differs from Other Methods

It may have occurred to the reader to ask how this method differs from traditional methods, which also seek to implement representations one at a time into a machine. To some it may seem like this method is the same, but spread out over time. However, there are two very key components of this method which set it apart from most other methods.

First, we have the aspect of genetic randomization. It is true that adding random gene-like components could be messy, but from any angle we can look at it from, the human mind also appears to be quite messy, and that is because of the same randomized structure. Sometimes a component may be added that helps or makes no difference at one stage, but then can have weird effects later.

For example, humans have a very complex and precise nervous system, which evolved over countless generations. This nervous system includes anxiety, a feature that is key to the flight-or-fight response and has probably saved many lives in the human and prehuman eras because it helped organisms react to dangerous situations, such as staring down a hungry tiger. If humans felt no emotion in that situation, they could easily be eaten because they didn’t react to the threat. However, this same part of the nervous system is what makes you sweat and choke up when presenting to a room of people. Clearly humans did not evolve in that way specifically so that we could have panic attacks in conference rooms, but rather to save us from predators and other dangerous situations. However, the genes that determine the formation of this system did not come about in order to save us from tigers. They were accidental genetic mutations which made their way into some human ancestor at some point and never left the gene pool because those with the gene outlived those without it (see Darwin, 1859). Because they are randomized accidents and not specifically made to warn us about predators, this system also kicks in when we’re staring down an audience and in no actual danger.

That is an example of the messiness of the human system. Any intelligence system that currently exists has this type of messiness, and it would be foolishly optimistic to believe that any intelligent system we create can be devoid of similar noise. The aspects that go into human intelligence have not been tailored specifically for the world. It would be more accurate to say that intelligence was entirely random and the forms of intelligent life that exist do because the intelligent beings with which their ancestors competed were less fit for survival.

So building capabilities from the ground up and then popping them in a robot is a very unnatural way to acquire intelligence, and so far it has not been proven to work. Evolving a genetic system using some randomized variables, on the other hand, has been proven to be viable by every intelligent creature on earth.

The second place this method differs from most is the stance one must take on functionalism in order to entertain my hypothesis. Specifically, there is no such stance. Most functionalist methods of developing AI rely heavily on a belief in functionalism in order to make sense because they involve implementing representations into an intelligent system. My method, on the other hand, does not require implementing representations.

There has been much talk of functionalism and the computational theory of mind in this paper. In fact, my entire proposal is an attempt at creating an artificial representational system equivalent to humans in capabilities. However, nothing about my method requires verification of the functionalist theory. Suppose the functionalist theory is wrong entirely, completely missed the mark. That doesn’t matter because my method is based entirely on results. The fittest organism of a generation is judged by how well it performs, and subsequent generations are produced semi-randomly. This method guarantees results in producing human-like intelligence, but it makes no promises of producing objective consciousness because those who believe consciousness is exclusive to living things cannot be convinced.

If a generation’s software is kept constant and the only randomization the next generation gets is new stuff, new DNA, added onto the preserved string from the last generation, the new generation will have the old abilities, as well as new ones. Some of these new “abilities” will only hinder, but through backwards propagation, some ought to do the opposite, improving memory, perception, and the conceptual system as a whole. These offspring will be better than the parents at whatever tasks their different genetic makeup influence, and therefore they will be more fit creatures for the environment.

You can throw away all talk of concepts and representations, but the hypothesis remains the same: some robots will be more fit than others, and you can pick those robots to be the model for subsequent robots, some of which will surpass the parent. If this cycle repeats the result will be a robot far more fit for the world than the insectoids with which we started, regardless of the theory of mind which one subscribes to. Of course one may not believe that these systems actually have subjectivity, but this method will result in machines capable of at least acting like highly intelligent beings.

A Formal Symbol Manipulator Capable of Understanding

As for considering the final humanoid machines “actually intelligent,” or capable of understanding, I believe that there is a powerful case for it. Clearly the functionalists would say that a robot indistinguishable from humans in ability and intellect is capable of understanding, as would the behaviorists. Those who may not be convinced are those that that believe that a system of formal symbol manipulation cannot have understanding, in accordance with the Chinese Room experiment (Searle, 1980)

The machines that I have described are formal symbol manipulators. They are digital computers, and so they manipulate symbols according to a set of rules. The general argument I seek to oppose is that the system cannot understand anything except for those rules, as all it does if perform formal operations. The argument makes intuitive sense, but only if the rules are grossly simplified.

For example, imagine a machine that passes one instance of the Turing Test where you are the interrogator. You thought that the machine was human based on its seemingly intelligent replies. You then find out the way the program works is it has several thousand or million questions built in, with the correct answer ready to go. You just happened to only ask questions that it had the answer to. Of course this machine is not intelligent because all it does it follow the rule and give the pre-determined answer. Even more realistic chatbots could be subject to this argument. The program parses the sentence, looks up the words and their meanings in its memory, and gives an appropriate response structured specifically for keywords in the sentence, such as returning one type of structure in response to “what” and another in response to “why.” Looking up the words and determining the structure is still algorithmic and in some way predefined by intelligent beings - the programmers. The intentionality comes from the programmers, not the system which they created.

However, the machines I have described are not subject to this argument regarding formal symbol manipulation. Though the machines do manipulate symbols according to rules, these rules are not rules in the sense that they are in the aforementioned examples. In those examples, the rules were explicit and predetermined. In these machines, however, the “rules” are not specific statements which determine the actions of the agent. Instead, these rules are the vast, precise, and complex neural connections and weightings and the representational system which they produce. The rules, much like the rules which govern humans, are not known exactly because they are never explicitly stated. The rules are simply how the neurons work and are connected to one another, as well as the learned efficacy of some connections.

Humans certainly have rules that govern us, even if we cannot determine these rules. If we did not follow rules there would be no order or predictability. Some rules are simple, such as when you get pinched, you feel pain. Every neuron involved, from the spot of the pinch to where the pain was experienced in the brain, helped enforce this rule. If there was no neural connection there, this rule would not exist. Other rules are less easy to explain, such as the rule which makes you sad when a loved one dies, but we know that they exist. Otherwise, we couldn’t be certain that you would be sad when a loved one dies.

Another difference between the grossly simplified rules and the rules that humans follow is that human rules can be modified. For example, a child’s mind might have the rule that when he sees a black, four-legged animal, it is a dog. When the child is shown a black cat and told that it is a cat, the rule for distinguishing dogs must change to exclude black cats. So, there exist rules in the brain for changing other rules. That makes sense because the rules are so poorly defined in the first place that it becomes difficult to think of them as rules. The rule for identifying a dog is not as simple as “if it is black and if it has four legs, it is a dog.” In fact, the rules for identifying a dog are unknown. All we know is that if the image of a dog is presented to the human mind as an input vector, the output would be the knowledge that it is a dog, not a cat. What makes this rule exist are the various layers of neurons which affect the input vector in ways which produce the appropriate output vector.

The artificial neural networks in the brain of the evolved robots work in exactly the same way. The rules are unknown, but with additional knowledge the rules can be modified to be right more often. This is important because it makes the rules the robot’s own. The rules which govern what the robot does are entirely produced within its mind as a representational system. The rules which govern how that system works are the result of genetically randomized neural connections. None of the rules which the robot follows are artificial. They are not programmed in. The only human rules in this system are the rules governing the backpropagation of the neural network, and those rules are only in place so that we evolve human intelligence rather than some intelligence which is most fit for another task. Humans don’t decide how the robots act; they only decide which ones act well enough to have their genes passed on.

I don’t believe that there is any difference between this machine and a naturally evolved animal except for the material they are made of. The only software that will be developed will be the one which establishes the neural network. The cognitive activities of the machine will be entirely self-constructed, the result of an optimally weighted neural network. Were these machines to be created, I would not ask whether or not they are really intelligent; I would ask whether or not the evolved intelligence was really artificial.

Works Cited

Brooks, Rodney (1990) “Elephants Don’t Play Chess” Robotics and Autonomous Systems

Brooks, Rodney (1991). “Intelligence without representation” MIT Computer Science and Artificial Intelligence Laboratory

Brooks, Rodney (2008). “I, Rodney Brooks, Am a Robot” IEEE Spectrum

Churchland, Paul (1990). Cognitive activity in artificial neural networks. In Daniel N. Osherson & Edward E. Smith (eds.), An Invitation to Cognitive Science. MIT Press. pp. 3--372.

Darwin, Charles (1859). On the origin of species by means of natural selection, or, the preservation of favoured races in the struggle for life. London, J. Murray.

Fodor, Jerry (1975). The Language of Thought ISBN: 0674510305

Lobo, M. A., Kokkoni, E., de Campos, A. C., & Galloway, J. C. (2014). Not just playing around: Infants’ behaviors with objects reflect ability, constraints, and object properties. Infant Behavior & Development, 37(3), 334-351. doi:http://dx.doi.org/10.1016/j.infbeh.2014.05.003