Monthly Archives: October 2012

Intellectronics and living computers

Posted on by

The Polish writer Stanislaw Lem (1921-2006) is one of the most influential science-fiction visionaries of all time. Mostly known for his novel Solaris (1961), which was later made into a film by Andrei Tarkovsky, Lem has been prolific in his fiction, often blending social satire with engineering fantasy. Space travel, human contact with alien intelligence and societies of robots are among his most favorite subjects. Lem has written non-fiction as well, such as the monumentalSumma Technologiae (1964), a reference to Summa Theologiae by Thomas Aquinas.

(Left: Stanislaw Lem (1921-2006))

In Summa Technologiae Lem speculates and philosophizes on prospective social, cybernetic and biological advances. What is of particular interest to AI is a chapter entitled “Intellectronics” where Lem discusses the concept of what he calls the “intelligence amplifier”.

The intelligence amplifier would be similar to existing machines that amplify human physical strength, for example cares, excavators, airplanes or cranes. A human connected via a suitable control system to an intelligent amplifier would be able to increase her IQ by many factors. Lem envisions intelligence amplifiers that could turn a person of average intelligence (100-110 on current IQ tests) to a super genius with an IQ of 10,000!

One could argue that intelligence amplifiers already exist – they are called “computers”. Computers increase our capacity for calculations manifold, with monumental and historical consequences in the way our civilization, our economy and our society evolves. They truly “amplify” our intelligence, collective and individual.

But Lem means something deeper than that. In comparing the IQ of the personwith the IQ of the person-plus-machine he suggests that the machine itself must be intelligent. In fact, just like a crane is more powerful than its human operator an intelligence amplifier must be more intelligent too!

The problem that Lem aptly identifies in the design of such a superintellgent machine is the obvious: the machine would have to be more intelligent than its designer. From a classical engineering perspective this means that one cannot develop design specifications (how can you describe what intelligence higher than yours can do and how it thinks?), and therefore one cannot even begin to imagine how this machine would be like, let alone how  to control it. This is of course a paradox that strong AI faces when dreaming of “machine superintelligence”, i.e. machines smarter than humans that would, supposedly, usher us into the era of “AI Singularity”. One only wishes they will be friendly, or else…

Interestingly, in Intellectronics Lem suggests an exit from the superintelligence engineering conundrum. Instead of electronics he proposes the search for new substances, building materials which in certain aspects are similar to living organisms.

Forty-seven years after the writing of Summa Technologiae researchers in Caltech have invented a method for designing systems of DNA molecules whose interactions simulate the behaviour of simple artificial neural networks.

The researchers based their biochemical neural network on a simplified model of a neuron. The model neuron receives input signals, multiplies each by a positive or negative weight, and only if the weighted sum of inputs surpass a certain threshold does the neuron fire, producing an output. The model was built by synthesizing DNA strands in a test tube, i.e. a real “computing soup”!

Lulu Qian, the lead author of the paper (see reference) posed the main theoretical question of the research thus: “Instead of having a physically-connected network of neural cells, can a soup of interacting molecules exhibit brain-like behaviour?”

Although the invention is far from having practical applications any time soon, it could have a future in nanodrug design. More interestingly however, the Caltech invention explores notions of computability in living systems. As Qian explained in Science Daily, “Before the brain evolved, single-celled organisms were also capable of processing information, making decisions, and acting in response to their environment. The source of such complex behaviors must have been a network of molecules floating around in the cell. Perhaps the highly evolved brain and the limited form of intelligence seen in single cells share a similar computational model that’s just programmed in different substrates.”

It is a theory worth testing, and certain to have made Lem smile were he alive today. Unlocking the secrets of biological computation may indeed be the way to building the intelligence amplifier that the Polish novelist dreamt of, so many years ago.

Journal Reference: Lulu Qian, Erik Winfree, Jehoshua Bruck. Neural network computation with DNA strand displacement cascades. Nature, 2011; 475 (7356): 368 DOI: 10.1038/nature10262

Animism and AI

Posted on by

Animism transcends all human culture. It is considered the proto-religion of our species, the first explanation we humans had about the workings of the world. Animism comes from our cognitive inability to distinguish between our psyche and the external world of animate and inanimate objects (read my post of Piaget’s relevant theory here). It is therefore of fundamental importance if we are to understand our relationship with artifacts that move, speak, or think.

With time, animism transformed to more sophisticated explanations that we today recognize as religion. A key characteristic of this transformation, at least in Europe, was the gradual anthropomorphing of animistic souls and powers. Thunder became Zeus; Earthquakes Engeladus; Fire Hephaestus, the Sun formed into handsome Apollo, and so on. Although there were still zoomorphic monsters out there and chimeras of all description, pantheism, as evolved in pre-classical Greece, attributed conscious and human-like intentions to everything in nature.

(Left: A robot made in Greece)

The Olympian human-like gods ended up ruling over everyone and everything. Although in Greek proto-religion creation occurs through the union of the primal elements,  the Male-Sky and the Female-Earth, the Olympians (image reflections of humanity “below”) were now capable to giving life too. They could create animated artifacts, like Hephaestus who poured the golden life-blood of the gods (“ιχώρ” in Greek, pronounced “ichor”) into Talos, the metallic man placed in the service of King Minos of Crete. Animism, as developed through pantheism, arrived at an intermediate position of human-like gods inculcating life, and consciousness, into things.

The invention of democracy in classical Athens (6th century BC) changed the relationship between gods and humans forever. Subjects became citizens, a novel political breed with the right to decide their destiny through debating and voting. Now, there was no power superior to the will of the people, which meant that  in a democracy human beings were truly free men, including free from the will or whim of gods. The Athenian democracy was far from atheist, but the degree of elevating human beings to the level of gods cannot be underestimated. The Parthenon’s friezes in the 5th century BC depicted Athenian youths riding their horses during the city’s greatest celebration. Till then temple friezes in Greece were reserved only for gods .

(Left: Automatic diversions by Heron in Hellenistic Alexandria)

The Hellenistic period that succeeded classical Greece went a step further: it sought to imitate gods by constructing animated machines. In ancient times “animation”, i.e. movement, was considered the principal characteristic of life. Life was by definition conscious because it had intention; and intention was expressed through movement. Movement was the “Turing Test” of the ancient world.

Engineers in Alexandria build machines that moved, mechanical clocks, steam engines, copper birds that sung. Those machines were still used and displayed in the Eastern Mediterranean centuries later, when Greek-speaking Christian emperors ruled in Constantinople. Their ideology was much different. Medieval Christianity, particularly in the East, frowned upon engineering, or the practical application of any knowledge.  Knowledge ought to be kept “pure” and for the sake of worshiping God; applying knowledge to an end was regarded as a sinful. The mechanical automata of Byzantium were meant to impress diplomatic missions from the empire’s periphery, not to compete or imitate the life-giving power of the now One – and Only – God.

(Left: Following on the tradition of building machines that move.)

When Christianity was questioned in 15th centuryEurope and later, the thread that was lost since Hellenistic times was rediscovered. Retro-animism in the sense of reanimating dead matter resurfaced, as alchemical, as automata, and in the ensuing industrial revolution as the construction of machines that replaced human workers. But the ideology of the Hellenistic engineers was forgotten No one remembered the psychological and religious roots of the call. Economic utility was now the principal, and sufficient, reason for engineering machines that moved, spun, pumped, steamed and worked.

This modern, utilitarian ideology still holds true today. AI researchers, when we aim to build an intelligent machine, we do not aspire to imitate godly powers. Presumably many of us, if not most, do not believe in a god to begin with. We want – at the very least – to do some interesting research that will help people and business, and – when real ambition kicks in – to solve one of the greatest mysteries of all time; consciousness.

Nevertheless, although these objectives may appear detached and “scientific”, I would argue that understanding a deeper requirement for strong AI that seems to stem from primal animism may inform the scientific quest for artificial consciousness better and, perhaps, resolve a paradox in its objectives.

For it is often argued that building a machine that thinks is not such a useful application after all. Machines are better than humans because they do things without “thinking”, or “feeling”, or being “conscious” of themselves. There is not an apparent added utility in replacing a naturally conscious agent with an artificial one; except when extreme environmental issues are at play, for instance if the conscious artificial agent must operate in a hazardous environment or independently for prolonged periods of time (e.g. in deep-space travel). However, the benefits from such applications are marginal when judged under the light of scarce research resources that could be better deployed to solve the bigger problems of the world. Perhaps, the deeper reason we want to create artificial consciousness is because we need to confirm our cognitive instinct that inanimate objects can think and have intentions too.

Ghosts in the machines

Posted on by

The Swiss psychologist Jean Piaget noted that in children’s’ minds there is an implicit understanding of the world in which all events are the product of consciousness or intention. Things happen for a reason and never by chance. Piaget’s discovery has tremendous repercussions in the way we understand ourselves and our relationship with the world. It is apparent that we humans are born with a cognitive inability to distinguish the external world from one’s own psyche. We are born with our consciousness embedded in a continuum that extends beyond our physical bodies. Evolutionary biology is at play here: the seamless connection of one’s mind to the environment promotes survival by automating fear responses. Quick reflexes are far superior to reflective cogitation when something moves in the dark. Asking what is out there would have been a fatal question in the living conditions of our hominid ancestors. Better to respond with an a priori hypothesis deeply embedded in the brain, that whatever is out there has intention and that this intention concerns me.

In this evolutionary wiring of our brains must lie the roots of animism, our common proto-religion. Early humans saw the workings of a conscious, all-pervading lifeforce in everything that surrounded them. There was a soul, or a spirit, or a ghost, in everything: in the animals they hunted, in the caves in which they lived, in the forest, in the water, everywhere. We still believe that, deep inside. Amazingly, contemporary developmental psychology has confirmed that we manage to distinguish between inanimate and animate objects only through learning. Our belief that inanimate objects and natural forces are soulless is acquired through living in a society that has evolved away from the dangers of savanna-living. It is a social skill. As such, it is constantly challenged by our deeper, cognitive systems.

(Left: She is beautiful. But is she “alive”?)

We dream of childhood in terms of innocence, and within this framework we become nostalgic of the period in our lives when toys would come alive in our imagination. It turns out that these toys were indeed alive. We did not imagine them so. Our brains evaluated them as living creatures. As adults we often approach robots in a child-like fashion. For anyone who has been next to a sophisticated human-like robot with sophisticated behavior, the feeling of seeing something with an intention is formidable. It seems to stem from deep inside our psyche. It is. Despite the technical obstacles and hiccups (See my post on “uncanny valley”) we will ultimately accept androids in our society because they will appeal to our deepest, most primal cognitive beliefs that machines with intentions must be “alive”, that they must have a soul just like ours.

Dreaming of electric sheep

Posted on by

Renown neurobiologists Christof Koch and Giulio Tononi recently suggested an alternative Turing Test to examine whether an intelligent machine is conscious or not: instead of having a human-machine conversation they propose a psychological test where the machine decides through a dialogue if a series of photographs are “right” or “wrong”. Their strategy assumes that demonstrating such knowledge on the part of the machine implies that it has subjective understanding of the world.

(Left: What is “wrong” with this picture?)

Looking at the above photograph it should be obvious to any human observer from age of 6 upwards that it is not “real”. People do not fly into the air. Could an intelligent machine make the same inference? If yes then, according to Koch and Tononi, this machine must be regarded as conscious.

Koch’s and Tononi’s proposal is based on Tononi’s integrated information theory for consciousness. According to this theory, for consciousness to occur information must be highly integrated. There is a measure for the integration of information in systems called “Φ”. This quantity signifies how much information a system contains above and beyond the information contained in its parts. In other words, it expresses the degree that the individual parts are interconnected. The higher the degree (i.e. the higher the value of Φ) the more “surplus” information the system contains.

For example, in the cerebral cortex individual neurons contain their own specific information (say, their ion potential levels); but they also have many specific interconnections. The Φ in the cerebral cortex is high because the amount of information that is contained in the whole system far exceeds the information in its parts. The degree of Φ may be correlated to the degree of consciousness. Because Φ can be measured in any system – including brains and intelligent machines –  Tononi hopes that measuring Φ may pave a quantitative path towards ascertaining consciousness in machines.

This is a very promising approach to the problem of machine consciousness. Complex systems, such as brains, can be regarded as information systems that exhibit types of behavior which cannot be deduced from the behavior of their interconnected parts. Brains behave as if they have a “subjective” understanding of the world. It is this understanding that we generally refer to as “consciousness”. The correlation between the complexity of an information system (e.g. a brain) as measured by Φ, and conscious behavior of that system is very strong; and that is why I have  been such a strong proponent of applying systems theory to the problem of consciousness (read my “Noetics” paper). It follows that  a complex enough machine that exhibits similar types of behavior ought to be judged by the same criteria, therefore the psychological test (“what’s wrong with that picture?” ) suggested by Koch and Tononi.

The two neurobiologists concede that current technology cannot possibly arrive to the levels of integration required for consciousness. This has been a major problem for “general knowledge AI”: algorithms do not suffice to solve the problem of general inference not only because of the theoretical limitations posed by Godel’s theorem of incompleteness, but also because of limitations in machine architectures (see also my post on “brain-like” computers). Regardless how powerful modern computers are, or may become, information on their electronic components remains unintegrated. For Φ to reach levels of human-like information integration we need machines where knowledge is embedded on highly integrated systems.

If such machines arrive, then I would like to propose an extension to the test suggested by Koch and Tononi.

(Left: What is “right” about this picture?)

There are many states of consciousness. Although under “normal” circumstances we find it “wrong” to see a picture of a man flying, we may not have issue with such an event if we see it in a dream. Dream states, naturally or artificially induced, are a unique characteristic of consciousness. Arguably, they are the foundation of human creativity in the arts as well as the sciences. An angel is a flying human and, for many cultures around the world, there is nothing “wrong” with a human-like creature having wings. Culture can only be produced by conscious agents. By the same token culture can only be “understood” (or “appreciated”) by conscious agents too. Let me then suggest that a conscious machine will be able to convince us that is conscious when it manages to amaze us, and move us, by composing a piece of art; a poem, a novel, a drawing, or a piece of music that will speak to our hearts.

Reference: C. Koch and G. Tononi, “A Test for Consciousness”, Scientific American, June 2011, pp. 44-47.

Brain-like computers

Posted on by

Most digital computers are built on a simple, albeit revolutionary, principle suggested by Alan Turing in 1936 whereby data and instruction sets (the “programs”) are stored together (in your “Hard Disk”) whilst information processing takes place separately (in your “RAM memory”). This is what a “universal Turing machine” does;  most modern digital computers today are universal Turing machines.

Alas, this is not how brains work. Brains function by performing processing and storage together. This realization has obsessed researchers in AI since the beginning, and various approaches have been tried to overcome this disparity, including parallel processing and neural networks.

(Picture left: The fundamental components of electronics)

In 2008 a team at HP labs invented the memristor, the “fourth” fundamental component in electronics (the other three are the resistor, the capacitor and the inductor). Memristors are like mutant resistors with a memory. Their resistance increases when the current flows one way and decreases when it flows in the opposite way. And they “remember” what current flowed through them. As a fundamental component in electronics, the application of the memristor is potentially boundless.

The exciting news for AI research is that memristors store and process information at the same time. With memristors one can build computers other than universal Turing machines that can do very interesting things. Researchers have already demonstrated that by using memristors they could model the adaptive behavior of unicellar organisms. Taking this finding a step further one could use memristors to mimic certain single cells called neurons, which in turn opens up the possibility of neuromorphic architectures performing massive pattern recognition based on memristive properties.  Such architectures could learn and adapt their behavior accordingly, ushering a new era of “brain-like” computers.

(Picture left: A memristor under a miscroscope)

The research interest in this area is enormous and recently a research team at Exeter University led by David Wright announced a method of building “brain-like” computers by using phase-change materials to perform general purpose operations, such as the four basic ones of addition, subtraction, multiplication and division. Moreover, they have demonstrated that they can mimic neurons and synapses. Synaptic functionality was demonstrated by the “memlflector”, an optical analogue of the memristor.

Memristors are not only a major technological breakthrough but a theoretical one too, particularly in building intelligent machines. Mimicking brains with memristors suggests that we could develop machines that will program themselves, become thus independent of their programmers. We may imagine such machines exhibiting intuitive-like behavior by inventing novel solutions to unexpected problems. This may bring us closer to seriously begin to discuss technological methods for deciding whether such machines would be conscious or not.

References:

Turing, A.M. (1936), “On Computable Numbers, with an Application to the Entscheidungsproblem”, Proceedings of the London Mathematical Society, 2 42: 230–65, 1937

C. David Wright, Yanwei Liu, KrisztianI.Kohary, Mustafa M. Aziz, Robert J. Hicken. Arithmetic and Biologically-Inspired Computing Using Phase-Change Materials.Advanced Materials, 2011; DOI: 10.1002/adma.201101060

Making machines intelligent

Posted on by

The goal of Artificial Intelligence (AI) is usually described as the making of intelligent machines. This may seem like a well-defined goal however AI has been plagued with misunderstandings and misgivings since its modern reinvention in the 1950s. The word “intelligence” is laden with cultural, philosophical and political frustrations, and the only way to see clearly behind the smoke is to deconstruct intelligence in the context of AI research and ideology.

To begin there is the engineering, hands-on, objective of AI in trying to answer what an intelligent machine actually is and how it differs from any other computer program. The difference is this: an intelligent machine perceives its environment, makes inferences based on accumulated knowledge and current environmental stimuli, and thereof takes appropriate actions to maximize the chances of succeeding in its goals. Put another way, the engineering goal of AI is to develop techniques and technologies for autonomous decision-making in uncertain circumstances.  This is a well-defined, prosaic even, technical objective. It is therefore not surprising that, as a discipline of engineering and computer science, AI has enjoyed much success over the past years. There are numerous successful AI applications today, from analyzing markets to enhancing the experience of video games.

The cultural fascination with AI starts where engineering stops. Building “intelligent machines” unavoidably begs the more general question “what is intelligence?” The nature of this question is not mechanistic but psychological. It deals with how we can tell a machine is intelligent, as well as how we relate to that machine once we have decided that it is intelligent.

A technical distinction must be borne in mind when exploring the psychological meaning of intelligent machines. In AI there are systems that aim for “specific intelligence” (such as expert systems), where the application space is both narrow and deep; for example medical diagnosis or financial analysis. These systems can be quite successful because the programmer may use heuristics (e.g. rules of thumb) to readily encapsulate human expertise.

Another class of AI systems aspires for “general intelligence” where the application space is wide and shallow; an example would be speech recognition or automatic translation systems. Here the situation is different because of the complex interconnectedness of general knowledge. Knowing a priori the relevance between facts is impossible. Coding to infer from general observations is not only technically challenging but theoretically precarious as well. A “universal machine” that infers everything from everything beckons at Russell’s paradox: can this theoretical machine infer itself? Is the set of all sets a member of itself?

We may of course fall back to a behaviorist position and agree that machines can be said to exhibit intelligent behavior providing their performance is adaptive and therefore unpredictable. Things become interesting when people react to such machines. Affective computing coupled with AI can provoke emotive responses from human beings that include sympathy and empathy. Humans can relate psychologically to intelligent machines.

And this is how we find ourselves up against the deepest, philosophical, question of all: how “true” is this machine intelligence? Is it just an engineering illusion, not unlike the “mechanical Turk”, only more sophisticated? Shouldn’t true intelligence involve some degree of consciousness?

Alan Turing suggested that, for all intents and purposes, a machine may be regarded as intelligent if it can fool you in believing that it is. The infamous “Turing Test” posits that if you cannot tell the difference between the responses of a human and the responses of a machine, then the machine is “truly intelligent”.

The Test was debunked by the philosopher John Searle who suggested that a system may appear to give “intelligent” answers to human questions without necessarily having any intrinsic knowledge – or consciousness – of the answers it gave. It may just follow a set of rules that made perturbations of symbols that were meaningful to human receivers but not to the machine itself. According to Searle intelligence without consciousness is not “true” intelligence. If we accept his notion then we can ask whether it is possible to create artificial consciousness. After all this was the dream of AI’s modern godfathers in the 1940s and 50s. It still is the central dogma of “strong AI”. However, artificial consciousness is not (yet) an engineering problem but a philosophical question of the most fundamental importance and gravitas.

There are many ethical corollaries to the questions posed by AI in its broader techno-cultural dimensions. For example, should we aim to develop intelligent machines further, and if so where should we stop? We can imagine a future where we relinquish control of many human affaires to intelligent computer systems and networks. Energy, trade, defense, international relations, the economy, could arguably become better-run without the evolutionary defects of biological agents (humans). Should we aim for a techno-Utopia of a world “all watched over by machines of loving care”? Other moral issues relate to political rights that may or may not be given to machines that exhibit “true” intelligence. Such issues become increasingly complex, and therefore more interesting, when the demarcation line between human biology and machines blur, as in the case of cyborgs.

AI is a controlling technology. It is the “brains” (conscious or not) of our global technological infrastructure. As computer networks link every facet of our civilization, intelligent control means that we are relinquishing the keys of planet Earth to our digital brethren. Exploring the questions posited above may help us understand what we are dealing with and, hopefully, prepare us for what will come.

The “uncanny valley”

Posted on by

There is a feeling we humans get when confronted with human-like artifacts. The more human-like the artifact the more we tend to like it. Think dolls or mechanical robots. But when these artifacts start to look much more like human our liking wanes and we start to get a creepy feeling.

This phenomenon is called the “uncanny valley” and has befuddled researchers in affective computing and robotic design. Making robots look human is something of a dogma when we imagine the robots of the future. Given the literature that precedes and inspires current research, we expect robots to evolve into androids, the Marias of Metropolis or the Rachaels of Blade Runner. But will they ever? And if they do, are we going to accept them?

The “uncanny valley” phenomenon could potentially spell the end of android evolution. To look into the causes of this, an international team led by Ayse Pinar Saygin of the University of California, San Diego (see journal citation below) made an experiment scanning the brains of 20 subjects aged 20 to 36 while they were looking three different scenarios: (a) a human, (b) a mechanical-looking robot, (c) a human-like robot.

Interpreting the results from the fMRI scans the researchers suggested that the cause for the “uncanny valley” is a mismatch between at least two neural pathways, that of recognizing a human-like face and that of recognizing the robotic movement. Humans do not move like robots. So when someone looking like human moves like a robot it makes us feel that “something is wrong”.

The results of the experiment reminded me of Capgrass Syndrome, the personality disorder syndrome where patients feel that their world is populated by mechanical impostors of their family and friends. Interestingly, research byVilayanur S. Ramachandran in 1997 hypothesized that Capgrass Syndrome is caused by a mismatch too. This time it was between the pathways that make the patient recognize the face of a loved one and those that evoke an emotion about that person. Failing to feel anything about the one you see creates the creepy feeling that there is “something wrong” about that person; that he is not “real”.

The connection between Capgrass Syndrome and the UncannyValley phenomenon runs deep into the culture of AI. The paranoid feeling of doubles is a common theme in Philip Dick’s work, which informs our contemporary techno-cultural milieu. Indeed Rick Dechard’s (played by Harrison Ford) dilemma in Blade Runner is to decide if Rachel is “real” or “artificial”. Can he really “love” Rachel? The Turing Test could be seen also as the statement par excellence that blurs the borders between “real” and “artificial” on the basis of emotional perception from the human observer. If Rachel speaks and moves like a human, then she is. Or isn’t she?

The only “cure” for the uncanny valley phenomenon would be to retune our perceptual systems into accepting mechanical-like motion from human-like artifacts. But this may not be so straightforward. Human brains are haphazardly evolved objects that guide our actions, our imaginations, our horrors and our triumphs, conditioned by the most powerful emotion that nature has ever invented: fear.

Journal Reference: A. P. Saygin, T. Chaminade, H. Ishiguro, J. Driver, C. Frith. The thing that should not be: predictive coding and the uncanny valley in perceiving human and humanoid robot actions. Social Cognitive and Affective Neuroscience, 2011; DOI: 10.1093/scan/nsr025