This variety of opinions is, to a big part, due to the fact that we do not have an overarching theory of intelligence. We do not have a theory about which consent would exist and which would tell us what intelligence is, what it takes to build it, and what the limitations are.
These three questions:
a) what intelligence is,
b) what it takes to build it, and
c) where the limitations are,
one may expect to be answered by brain and cognitive sciences. However, despite the optimistic picture that popular articles sometimes paint, as an insider, I can tell that science does not properly understand how our brain works. The principles of brain operations that we managed to decipher so far are only partial. Our understanding of the whole picture is, to say the least, incomplete. We certainly do not have an adequate explanation of how our mental operations—that is, our abilities to perceive, decide, think, and act—emerge from our physiological hardware.
The today's situation in the brain-mind sciences reminds of that in chemistry during the times of blooming alchemy. There were a lot of cool phenomena to be observed and replicated in labs, but no one had an overarching insight of what really was going on. The lack of a theory made it difficult to know what chemical substances were, what it took to create new ones, one and what the limitations were. This produced a fertile ground for speculations, leading to unfulfillable promises, and ungrounded fears. The modern situation in the brain-mind sciences is no different in that respect; we witness both far-reaching promises and doomsday fears.
But this may change soon.
A theory has been proposed recently that can, for the first time, offer an overarching account of how biology managed to create intelligence. This theory, under the name “practopoiesis”, proposes a whole new set of future experiments in studying the brain and mind and, by doing so, paves a new pathway for understanding the physiological substrate of human consciousness. But equally so, the theory has implications for understanding what intelligence is, how to create it artificially, and where the limitations are. It turns out that the theory can give us a few ideas of what we need to do to reach high levels of AI. The theory can also help us identify limitations, and establish that not everything that is currently being promised can be actually done.
Bellow I explain these implications in more detail and lay down the general steps needed to create biological-like intelligent machines (or strong AI). But first I would like to establish that the critique of the current AI is not at all new. A number of scholars have been pointing out throughout years that the classical approach to AI, the one that we have utilized until today, may be fundamentally insufficient.
One of those people is philosopher John Searle who has formulated his Chinese Room thought experiment with which he illustrated a claim that looking-up into a database and computing outputs based on that database is not an equivalent of human thought. He claimed that humans do not just execute programmatic statements, or look up in databases for information. Instead humans understand what they are doing, while databases cannot possibly understand and hence computers (based on such databases) cannot either. He then distinguished syntax (what computers do) from semantics (what humans actually do). For example, a computer would apply certain rules on a picture to decide whether there is a car. A human would understand that the presented object is a car. Importantly, he did believe that in principle human mental operations are based solely on physical processes and thus, that machines could mimic them. But this would have required another approach than databases. He thus, distinguished weak-AI, or what machines did at the time, from strong AI, as a hypothetical machine that would be able to understand in way humans and animals do.
Searle was by no means alone in his critique and in calls for better approaches to AI. A popular list of the limitations faced by the dominant approach to AI was extracted by philosopher Hubert Dreyfus in his 1972 book "What computers can't do", and then reiterated twenty years later in "What computers still can't do".
But let us see what it means to understand. What kind of performance we need from strong AI?
Please take a look from the following picture. Can you see a car in it?
It should be no problem for a human. A human can understand that this picture can represent a car. For today’s AI this is a problem. Unless the AI was trained explicitly on these types of picture, it is impossible to “see” in car in this image. If an AI has been trained only on real-life photographs of cars it is hopeless in such a task.
Humans, in contrast, can understand how this image can represent a car even if they see the image for the first time. Children can do it without any problem too. Moreover, a child does not need 1000s of examples of real cars in order to see a car in the above image. It is sufficient if a child has had a chance to play with one single car and it automatically becomes capable of understanding the above image as a car too. This is because the child has the power of understanding what the car is about, e.g., where the wheels are and where the cargo is. This power of understanding (semantically) the car, rather then extracting statistical properties of pictures of cars, is the difference between the strong, biological intelligence and the weak, machine intelligence of today.
But this is not where the advantages of strong AI stop. Consider the next picture. If the preceding one was a car, than what is this?
This becomes immediately a truck. Even a child who has played with just one truck-toy can immediately generalize from a car above to a truck here.
And a similarly easy generalization happens here:
whereby the image becomes a train. For a child that had seen and understood the concept of just one train, this image becomes easily a train.
Humans understand. Today's AI does not. There is a big difference on how we go about our lives and how machines do their jobs. To be biological-like intelligent means to be capable to use previous knowledge in order to immediately understand and interpret novel situations.
But there is even more to our understanding. We can flexibly change our interpretations. For example, let us, for a moment, consider the image above not as a car any longer, but instead as a table with two chairs:
No problem!
If this was a table, how about this one?
Is this a setting in a restaurant? Again, no problem for a human.
And do you see a classroom here?
Everyone does.
A change of perception through understanding is easy if you are strongly intelligent and impossible if you are a machine. We acquire understanding of the relations between the functional and structural aspects of tables, chairs, cars, trucks, trains and so on. This is what makes semantics and what makes the above tasks easy. Today’s AI just learns stimulus-response associations from a number of images, and this is defined as syntax, and is insufficient to satisfy tasks that probe domains outside those images.
Irrespective of how the above example can look convincing, it was by no means that AI researchers easily accepted the critique. The relationship between AI enthusiasts and those who pointed out limitations did not seem to have been a productive one. AI researchers largely dismissed Searle and others. In fact, it was quite easy to ignore such critique. The problem with Searle’s argument was that it was not based on examples like the one above or on hard facts. Rather it was a philosophical argument—abstract, and based on thought experiments. Most of all, a constructive proposal of how to do things better was missing. As a result, many saw Searle's argument as flawed, a situation that remained until today. For example, in his 2005 book "The singularity is near" Ray Kurzweil calls Searle's arguments a tautology—and he is by no means alone in this view.
Nevertheless, the predictions that followed from the critique of Searle, Dreyfus and others seem to have turned out true. Repeatedly, there was a big discrepancy between the promised and the delivered, and this had at some moments in time created a somewhat negative perception of AI. Late seventies and end of eighties were especially bad, leading to the serious consequences of reduced funding and reduced general enthusiasm. These periods are known as AI winters.
Although AI made admirable progress, it has a history of over promising, leading to waxing and waning of enthusiasm. Today, we have a hype again—and AI “summer”. A winter may come again if we don’t find a way to create new type of technology, one that would allow solving the Searle’s problem of understanding.
In fact, the AI technology of today is not much different than that from eighties. The main improvement is more powerful computers, and bigger datasets for training. But computing power and big data may not be enough.
I want to describe significant theoretical insights that the theory of practopoiesis can bring for understanding the nature of intelligence, what the limitations are of strong intelligence, and what would it take to build one.
What is the nature of (strong) intelligence?
Any intelligent computer software would need to have two components: For one, it needs something that we call algorithms, and these are programs that are conceptualized and specified by human persons. Algorithms have the purpose of collecting knowledge, storing and manipulating it. But algorithms alone cannot make a machine intelligent.
A second component is thus needed. And this involves an extensive interaction with the outside world. A smart machine must acquire a bulk of its knowledge by its own work. In other words, it has to learn.
For example, an artificial neural network may need to be exposed to thousands of images of cars before it learns to classify them accurately.
The result of this learning is stored in a “box” (for example, in “synaptic” weights). The box later makes the decisions and classifications.
Importantly, in classical AI it is the box that requires large memory resources for storage of information. Inside, it has crammed a huge amount of knowledge. In comparison, very little information is stored in learning rules. In fact, learning rules can be often described with just a few equations, and the code specifying their execution occupies also a comparably small amount of memory.
Therefore, the classical approach to AI can be illustrated like this:
In contrast, strong AI organizes its knowledge in learning rules, and not in the box. The “box” contains only the knowledge needed to deal with a current situation—the one occurring right now. For example, one can possess box-knowledge only for interacting with one particular car that is being encountered at this very moment.
In strong AI, it is the learning rules that take up most of the memory resources, and not the box:
Strong-AI perceives and decides through learning. For example, perceiving a car may involve several quick learning steps: Is that a car? Maybe it is not a car. Oh, I it is a car. But what kind of car? Could it be my car? No, it doesn’t seem to look like my car. It is definitely not my car.
This cognitive process based on learning changes the box continually and is the very reason why biological intelligence has working memory, and today’s machines do not. Also, this fast learning process makes our cognition situated.
Knowledge stored in learning rules is more general than that in a box. This is what gives strong AI the capability of semantics and understanding. For examples, it is the application of such learning rules to the pictures above that make us recognize a car or a truck.
Therefore, the problem of creating strong AI, is the problem of making machines learn new learning rules. In biological systems this happens through an inborn set of learning rules stored in our genes. What we assume normally under “learning”—as in “I learned to drive a car”—is an application of these genetically-encoded learning rules to create our long-term memories in a form of more specialized learning rules.
Therefore, a more complete illustration of how to organize the resources of strong AI is:
According to practopoietic theory, an intelligent agent that has these three components is called a T3-intelligence. In contrast, today’s AI is a T2-intelligence as it has only two components.
Limitations of strong AI: no bootstrapping
One of the commonly discussed scenarios about growth of intelligence is the one in which an AI develops far above human capabilities on its own and very quickly. This is known as “intelligence explosion”: an already smart intelligence would be capable of changing itself to become even smarter, which would make it even more capable of changing itself making it in turn even smarter, and so on.
But according to practopoiesis, intelligence cannot bootstrap. An AI explosion would be for intelligence what a perpetuum mobile is for energy. One can't get something from nothing.
The fundamental property of intelligence is the need for learning from environment. And an intelligence cannot predict what knowledge it will receive from the environment before it has actually received that knowledge from the environment. Hence, it is not possible to boost intelligence by a new algorithm such that no time would need to be invested in learning.
This also means that understanding of the world is limited. This holds true for us and for machines. Hence, neither we nor machines can directly engineer a strong AI in a way we engineer for example, a car.
Higher levels of intelligence can only be evolved.
AI-Kindergarten: How to build strong AI
The problem of building strong AI (or T3-AI) boils down to creating a proper set of learning rules at the level of “machine genome”. For biological evolution it took millions of years to create proper genetic knowledge, and we don’t have that much time.
That means that we have to “steal” knowledge from biology. AI-Kindergarten is about transferring knowledge from our biological genome into a “machine genome”. AI-Kindergarten works through interactions between machines that need to acquire knowledge and humans who have that knowledge.
To accelerate the evolution of machines (acceleration factor: millions of times) machines need to get proper challenges in a correct order and a quick feedback. This process is somewhat similar to the transfer of our civilization to a new generation of kids. The 20,000 years or so that it took to bring our civilization to the present level, we manage to transfer in just a few years, applying adequate schooling, teaching, training, playing, etc.
AI-Kindergarten provides adequate training and feedback to evolve machine genomes, and bring them gradually up to our own level. Hence, AI-Kindergarten deals in total with four learning components (a T4-system):
The details of how AI-Kindergarten works are described in this video and are specified in a provisional patent. AI-Kindergarten functions similarly to selective breading of animals—when for example, we turn wolfs into dogs. The differences are that AI-Kindergarten has better efficiency and that it starts from the earliest stages of evolution. It would be something like selectively breading bacteria with a plan to eventually create dogs.
Here is an artist impression of T3-robots learning in AI-Kindergarten:
It is possible to apply AI-Kindergarten to creation of commercial products based on T3-technology. We begin with development of specialized strong AI systems for applications in time series analyses and forecasting. Later we will expand to sound and image processing, and other applications, with a long-term goal to approach gradually human-level intelligence.
This technology is safe because intelligence cannot grow without extensive training and is completely controlled by the contents of that training. Thus, it will be solely up to us to develop, or not develop, AI that obeys Asimov’s laws of robotics.
Further information:
The droids we’re looking for, by Anthony Miccoli
The original paper on Practopoiesis
A blog on practopoiesis
TEDx talk on AI-Kindergarten, organized by the European Space Agency
Acknowledgment
The author would like to thank Pascal Weinberger for help during the preparation of the manuscript.