Introducing Artificial Intelligence (Henry Brighton and Howard Selina)

It is important to note that Searle is not denying the possibility of Strong Al. Indeed, Searle believes that we are nothing more than complex machines, and therefore we can build machines that think and understand. Searle’s objection is with the notion that machine understanding is simply a matter of coming up with the right program. Searle strikes at the heart of functionalism.

For Searle, to argue anything else means you must believe in a form of dualism, the position that the mental realm has no causal connection with the physical realm. According to Searle, this is precisely the position many AI researchers take. They believe that their models have a mental life purely on the basis of the right program being executed. Mental phenomena can be understood solely in terms of programs (mind), independent of machinery (brain).

For Searle, the nature of the machinery required for consciousness is a mystery. He makes no claim to have an answer explaining why computers cannot support properties such as understanding and consciousness, but brains can. In contrast, Roger Penrose, a mathematical physicist at the University of Oxford, proposes a candidate “stuff”.

Like Searle, Penrose argues that conventional computing machinery cannot support consciousness. A conscious mind requires very specific physical characteristics.

To support his argument, Penrose appeals to a fundamental theorem in mathematical logic – Gödel’s Theorem – which states that certain mathematical truths cannot be proved by using a computational procedure. Because human mathematicians evidently can arrive at these truths, Penrose claims that humans must be performing non-computable operations.

If human thought comprises non-computational processes, then how is the brain supporting these processes? To answer this question, Penrose appeals to physics, and claims that the theory of quantum gravity is likely to be the kind of physics required to explain a conscious mind.

The idea that quantum gravity may prove important to our understanding of consciousness predates Penrose, but he has stuck his neck out and specifically proposed that quantum gravity effects in the brain are likely to rely on microtubules – conveyor-belt-like structures inside

neurons.

The classical approach to AI encompasses a set of principles and practices used to explore the validity of cognitivism, and specifically, to investigate the hypothesis proposed by Newell and Simon. Cognition is best understood as the formal manipulation of symbolic structures.

The classical approach to AI has resulted in engineering projects such as the following, which we shall examine in more detail later. •Chess-playing computers capable of beating the best human players. •Attempts to equip computers with common-sense knowledge.

In 1963, after spending $20 million on research into machine translation, the US funding agency concluded: “There is no immediate or predictable prospect of useful machine translation” – National Academy of Sciences National Research Council, 1963. Faced with a hard problem, AI research will often begin by simplifying it. Two kinds of simplification are frequently made.

Fortunately, cognitive brain functions are not part of a complex mush that cannot be decomposed. Many have argued that our brain is structured rather like an interconnected set of sub-computers. Some of these sub-computers seem to work independently, which is good news for Al. The psychologist Jerry Fodor, in the 1980s, proposed that the mind is largely composed of a set of task-specific modules.

Consider the Muller-Lyer illusion. Line 1 and Line 2 are of the same length, although Line 2 appears to be longer than Line 1. Even though we have knowledge that tells us the two lines are of equal length, our perception of the two arrows is not privy to this information. Our perception

“module” must be working independently of this knowledge.

As a result of learning from its experience, the program carried on improving quickly and soon beat a checkers champion. The champion remarked after defeat in 1965 …

This victory of machine over man is widely cited, and for good reason. It demonstrates an important lesson: the capabilities of an artificial agent are not necessarily restricted by the capabilities of the designer. Samuel’s program plays better checkers than he does.

The space of all possible tic-tac-toe games is negligible in comparison to the number of possible games of chess. One of the world’s greatest chess masters, Garry Kasparov, perfectly expressed the difficulty.

With chess, looking ahead even a moderate number of moves becomes intractable – the number of combinations becomes too large to contemplate. The game tree for chess cannot fit into the universe, let alone a computer’s memory.

In chess, winning board positions are situated deep in the game tree. Chess-playing computers cannot reach these positions by search. It would take too long. Instead, they look ahead only a certain distance. Using a measure which reflects how advantageous a given board position is, these intermediate positions are ranked, and the best one is chosen.

These tactical rules of thumb are called heuristics and crop up in AI systems everywhere. Heuristics don’t guarantee success or correctness, but offer a good approximation. Heuristics are used when more exhaustive and precise methods are intractable.

The idea that knowledge can be formalized is not new. For centuries the act of thinking has been seen as calculation based on logical reasoning. Newell and Simon’s physical symbol systems hypothesis has its roots in the work of the philosopher Thomas Hobbes (1588–1679).

Hobbes’s “parcels” were the basic units of thought, just as symbols are basic to Newell and Simon’s Physical Symbol Systems Hypothesis. Hobbes’s ideas were furthered by the mathematician and philosopher Gottfried Wilhelm Leibniz (1646–1716) who tried to identify an appropriate system of parcels – a logical language. Leibniz imagined writing down all the facts known to man in this language, which he called Characteristica Universalis.

Logical reasoning requires the manipulation of sentences described in a logical language. These sentences can be interpreted as representing concepts such as states of affairs in the world – or knowledge. Using computers to automate this process, Al has taken the “logic as thought” idea and built on it.

The CYC project is entering its final phase, with Lenat predicting a 50% chance of success. Apart from the practical benefits of a successful CYC project, the theoretical objective is to test Newell and Simon’s hypothesis. Is commonsense something that we can formalize and automate using symbolic representations?

A recurring justification for the inadequacy of logic-based systems is the “just one more rule” defence. Rather than questioning the enterprise as a whole, the tendency is to persevere with the powerful idea of formalized knowledge dating back to Hobbes.

Shakey’s design mirrors the traditional view that an agent should be decomposed into four functional components. This model revolves around the sense-model-plan-act cycle. First, the agent senses the world. Then a model of the world is constructed on the basis of the sensory inputs. Using this model, a plan can then be constructed to guide how the agent will carry out actions in the world.

The planning design was monolithic. After the plan was set into action, Shakey largely ignored feedback from the real world. For example, if someone secretly removed the block Shakey was interested in, he would become very confused.

Connectionism gained popularity in the 1980s and is often depicted as a radical departure from the classical, symbolic approach to AI. Rather than viewing the processes of the mind as a computer program, connectionism draws a parallel between the processes of the mind and the processes of the brain.

The building blocks of neural networks are simplified versions of biological neurons called activation units. These units have a set of input connections and a set of output connections. These connections model the job performed by axons.

The brain is massively parallel, whereas most computers compute serially. This is why the brain is so fast, despite its relatively slow machinery. The property of parallelism present in neural networks makes connectionist models appealing. The manner in which they carry out the processing task is far closer to how the brain computes.

Machine Learning is a branch of AI that spans both the classical symbolic approach and connectionism. Here, models of learning capture the ability of an agent to improve itself in light of information in the environment. Often, the ability of connectionist systems to learn is cited as one of its defining characteristics, and a feature most attractive to AI researchers.

The kinds of information processing performed by neural networks can differ fundamentally in nature from those found in symbolic systems. Representations are often distributed in the same way that processing can be distributed. A distributed representation is spread out across the whole network, rather than being localized to a specific area or being built up from atomic units.

Of course, neural networks themselves are built up from atomic units – artificial neurons – but these units are rarely used by the designer to represent anything in themselves.

So, in a distributed representation, a single neuron is unlikely to be responsible for representing the number of legs of our candidate animal. Instead, the number of legs would be represented by a complex pattern of activity over a wide number of neurons. Some of these neurons will play a part in representing some other property in the system.

As a general rule, you can’t locate specific items of information by pointing your finger at part of a distributed representation in the same way that you can with a local representation.

Predictably, the resilient Searle stands firm and replies with the Chinese Gym. Instead of a room containing a lone Searle, he imagines a gym full of non-Chinese speakers, one for each neuron in the neural network.

However, the Chinese Gym does serve as an illustration that the whole can be more than the sum of the parts. In a sub-symbolic system, the atomic units, neurons and their structured relationship to other neurons, do not individually do much at all. Only when the collection is seen as a whole can we start to speak of concepts such as distributed representations and cognition.

Searle’s argument concerns the inability of the symbols being manipulated to mean anything. By themselves, symbols are meaningless shapes realized by, in the case of a conventional computer, a pattern of electrical activity. Any meaning we confer to the symbols is parasitic on the meaning in our heads.

Harnard views connectionism as a good candidate for achieving this grounding, especially when coupled with a symbolic system.

First, imagine a native English-speaker learning Chinese, armed only with a Chinese-Chinese dictionary. Harnard likens this to a cryptologist cracking a code.

Could you ever learn Chinese as a first language, with only the aid of a Chinese-Chinese dictionary? Harnard likens this to a symbol-symbol merry-go-round.

How can symbols ever be grounded by anything but other meaningless symbols? Part of the problem of ascribing meaning to a symbol requires that the circle of meaninglessness be broken. Harnard imagines a classical symbolic system sitting on top of a sub-symbolic connectionist system. Importantly, the connectionist system has inputs that are grounded in the outside world through sensors. In this way, symbolic representations are no longer defined in terms of other symbols, but are instead related to iconic representations which are directly linked to the sensory surfaces of the system.

It is the connectionist system that supplies the sensory images. By combining symbolic and connectionist systems, Harnard believes we can begin to break out of the closed world of meaningless symbols that Searle discusses.

“We used to argue whether a machine could think. The answer is, ‘No’. What thinks is a total circuit, including perhaps a computer, a man, and an environment. Similarly, we may ask whether a brain can think, and again the answer will be, ‘No’. What thinks is a brain inside a man who is part of a system which includes an environment.” – Gregory Bateson This observation has led to the adoption of a new set of principles. This new orientation is not yet fully-fledged – it lacks a commonly used name, but is often termed New AI.

In one sense, the problem of creating intelligent agents has already been solved. Over the course of the earth’s 4.5 billion years of history, evolution has solved the problem over and over again. Mammals arrived 370 million years ago. Our last common ancestor with the apes started milling around 5 million years ago.

The MIT roboticist Rodney Brooks takes the evolutionary basics as evidence that “hard” tasks like reasoning, planning and language might turn out to be easier to understand once the basics are in place.

Can our knowledge of evolution inform AI? Brooks believes that it can, and argues that we should first aim to build basic mechanical creatures before we try to build mechanical humans.

Maturana and Varela illustrate this point by first presenting a juicy fly to the top-left area of the frog’s field of view.

Next, they sever part of the frog’s eye, such that the whole eye can be rotated by 180 degrees.

Importantly, the frog will persist with this behaviour. It will never adapt its behaviour in light of the unsuccessful attempts to capture the fly. The moral of the story is that a frog’s eye does not act as a camera supplying information to the frog’s planning module, which then constructs a plan to catch the fly.

as Maturana and Varela went on to show, the fly-catching behaviour is solved by the retina itself, independent of the processes going on in the rest of the frog’s brain. This experiment illustrates how certain behaviours, such as foraging for food, are realized through a tight coupling between perception and action, independent of, and without the need for, any high-level cognitive processes.

Many of the concepts central to Al have their roots in the work of philosophers such as Descartes, Hobbes, Leibniz, as we have seen, and the Tractatus Logico-Philosophicus of Ludwig Wittgenstein (1889–1951):

Wittgenstein, in his later philosophy, and Martin Heidegger (1889–1976) strongly reject the formalist assumption of meaning.

This alternative philosophical standpoint suggests that our interpretation of the world cannot be made explicit, and any attempt to do so will render our insights grossly inaccurate.

This argument formed the backbone of one of the foremost critiques of AI. The philosopher Hubert Dreyfus, in the 1970s, declared that AI was misguided in its assumption that disembodied intelligence was possible. On the perceived failure of classical AI, Dreyfus noted …

The arguments from evolution, biology and philosophy stand in opposition to much of conventional AI research. But to put these arguments into practice, they need to be translated into engineering principles. Three principles characterize the new approach to AI.

The First Principle of Embodiment

The degree to which embodiment is significant remains a controversial issue. Rodney Brooks goes as far as to say, “Intelligence requires a body.” For example, the design of a robot body will determine the sensory phenomena it experiences.

Situatedness refers to an agent being located in a complex environment, rather than a highly abstracted micro-world. The complexities of real environments are taken to be fundamentally different from those of the abstracted “micro-worlds”. Indeed, being situated permits the exploiting of structure in the world, lessening the burden of internal representations. Elsie, Walter’s robotic tortoise, exploited the location of the recharging station inside the hutch. No model of the hutch is ever invoked by Elsie. Its function was the result of the interaction between Elsie’s sensors and the real-world. Rodney Brooks sums up this kind of relationship by arguing that “the world is its own best model”.

Given the goal of building an intelligent agent, the methodology frequently adopted by Al is to build from the top downwards. That is, higher-level functions such as knowledge and reasoning are targeted first, with lower-level Junctions swept under the carpet until later. New Al proposes bottom-up design. Start with the basics first… For example, Rodney Brooks builds basic machines analogous to insects. His idea is that only by understanding the basics first can we begin to understand the complexities of human cognition.

Behaviours are built on, to yield more complex behaviours. Unlike much of conventional robotics, which takes the sense-model-plan-act cycle as a starting point, Brooks’ robots contain pieces of machinery that are autonomous and operate in parallel. There is no central control. These behaviours implement a tight coupling between perception and action, and avoid the use of cognitive processes to mediate between perception and action.

Genghis has no central control. Nowhere in his construction is there a description of how to walk. “The software for Genghis was not organized as a single program but rather as fifty-one little tiny parallel programs.”

Although the principles of new AI translate most directly into the field of robotics, they are by no means limited to issues in robotics. A closer treatment of the interaction between agents and their environments can be applied to every branch of AI. Luc Steels, director of the AI Lab at the University of Brussels, takes another line to the “bottom-up” approach by investigating the evolution of both meaning and communication systems in collections of agents.

The aim of the experiment is to investigate how a shared language can emerge as a result of the interactions between agents. Crucially, nowhere in the experiment is language defined; it develops as a result of interactions between the agents. Starting with a blank slate, the agents autonomously develop their own “ontologies” – a sense of being in the world – that allow them to identify and discriminate between objects in the real world.

The key insight of the Talking Heads experiment is that agents develop their own individual and internal way of categorising the world they see. While, at the same time, through external communication, they negotiate a shared lexicon. Different agents may be talking about the same object, but they might conceptualize it differently, yet at the same time share words. Steels’ experiment illustrates how a communication system, grounded in the everyday world, can emerge through interactions between agents, yet not be defined in any one of them.

“In stark contrast to the largely unanticipated explosion of computers into the mainstream, the entire endeavour of robotics has failed rather completely to live up to the predictions of the 1950s." Hans Moravec Most people know what robots are and could perhaps even name a famous one. But other than industrial robots which are widespread in, for example, the car construction industry, robots are rarely seen outside research laboratories. Useful robots have failed to materialize.

By engineering our own offspring, many have proposed that Artificial Intelligence can lead to the Lamarckian evolution of our species. In contrast to Darwin’s theory of evolution by natural selection, Lamarck proposed that evolution allows characteristics acquired over our lifetime to be transmitted to future generations.

•Rodney Brooks, Robot: The Future of Flesh and Machines (London: Penguin, 2002).