Friday, 27 July 2012

Turing's legacy: AI in the UK today

Since Alan Turing first posed the question “Can machines think?” in his seminal 1951 paper, “Computing machinery and intelligence” the field of Artificial Intelligence (AI) has been a growing and fruitful one. While some might argue we are still as far away as ever from reproducing anything approaching human intelligence, the capabilities of artificial systems have expanded to encompass some impressive achievements. 

The defeat of world chess champion, Garry Kasporov, by IBM’s Deep Blue in 1997, was a significant moment, but perhaps not as surprising as Watson’s victory on the US quiz show, Jeopardy!, last year. The processing of spoken language proved to be a particularly hard nut for AI to crack, but Watson’s victory signals that a widening range of abilities traditionally thought of as uniquely human are beginning to yield to the arsenal of techniques AI researchers have at their disposal. 

In the wake of Turing’s centenary last month, it seems fitting to survey how far the field the London-born mathematician pioneered has come in the UK since he was tragically lost to us over half a century ago. This article therefore attempts to identify some of the most significant recent trends in AI research in the UK, as well as describing some projects that have either generated significant media interest, had significant social impact, or offer great promise for the future.

Computational creativity

One of the most visible recent successes has come from attempts to develop software which exhibits behaviour which would be judged creative in a human.  A growing community of researchers have been active in this area for some time, culminating in the first computational creativity conference in Lisbon in 2010.

An important player has been Simon Colton at Imperial College London, with his PaintingFool software artist. In a 2009 editorial Colton and others described the fragmentation of AI research from the ambition of early projects aimed at “artefact generation” into subfields, such as machine learning, planning, etc., in a “problem-solving paradigm”. They went on to claim that computational creativity researchers are “actively engaged in putting the pieces back together again”.  The aim is to combine various AI methods with techniques in areas like computer graphics, to automatically generate “artefacts of higher cultural value”.

He has argued that for software to be judged creative, it needs to demonstrate three key aspects of human creativity - skill, appreciation, and imagination. Skill is not difficult to demonstrate for software which can abstract regions of colour in images, shift colour palettes, and simulate natural media such as paint brushes, but the others would seem at first glance to be beyond the capabilities of current systems.
Nevertheless, an attempt to demonstrate appreciative behaviour, by using machine vision techniques to detect human emotion and paint appropriate portraits, won the 2007 British Computer Society's Machine Intelligence prize. More recently, the group’s use of generative AI techniques, such as evolutionary search, along with 3D modelling tools, have led to works of art deemed sufficiently imaginative to have been exhibited in two exhibitions in Paris last year. The pieces were neither abstract nor based on photographs, challenging the idea that mere algorithms cannot produce original, figurative art. The pictures were also featured in the Horizon documentary, The Hunt for AI, broadcast this April.

But it is the combination of many different computing and AI techniques, and their “layering” into a teaching interface through which the system learns, which demonstrates the bold claim of “putting the pieces back together again”.

Enactive cognition

A potential criticism of computational creativity is that it lacks what some might claim is the crucial purpose of artistic endeavour. Art is often thought of as a form of communication and any current artificial system lacks the intentionality necessary to motivate communication between what philosophy calls autonomous agents. There is a radical view in cognitive science, which claims that consciousness is a unique property of evolved, biological life. This is the life-mind continuity thesis of the Enactive Cognition movement. Life (and the potential for death) gives rise to self-generated, self-perpetuating action, and thus intention, which is precisely what an engineered system lacks.  

First articulated by Varela, Thompson and Rosch in 1991, enactive cognition, in its broadest sense, is an attempt to reframe the questions we ask when we investigate cognition and consciousness. Rather than focussing on individual components of mind such as neural activity or functional anatomy, a much broader focus is encouraged - on whole organisms and their interactions with the environment and each other. Notions entrenched in mainstream cognitive science, such as computation and representation, are rejected as inappropriate models of biological cognition which will ultimately thwart attempts to understand consciousness. Professor Mark Bishop, chair of cognitive computing at Goldsmiths, put it this way: “By reifying the interaction of my brain, in my body, in our world, enactivist theory offers an alternate handle on how we think, how we see, how we feel, that may help us escape the 'Cartesian divide' that has plagued Cognitive Science since Turing.”

Enactivists also see perception and action as inseparable aspects of the more fundamental activity of “sense-making”, involving the purposeful regulation of dynamic coupling with the environment via sensorimotor loops. The “sensorimotor contingencies” theory of psychologist Kevin O’Regan, where perceiving is something we do rather than sensations we have, fits neatly within this framework. Such embodied approaches to perception are gaining ground in psychology and attracting media attention, as evidenced by numerous features in NewScientist.

Chrystopher Nehaniv’s group at the University of Hertfordshire have also made progress using an enactive approach to robotics. A focus on social interaction has led to robots which play peek-a-boo and can learn simple language rules from interacting with humans.

Currently however, enactivism is primarily a critique of the classical paradigm and lacks a coherent research agenda of its own. As a first step towards remedying this, the Foundations of Enactive Cognitive Science meeting held in Windsor, UK, this February, brought together researchers from philosophy, psychology, AI and robotics, in an attempt to start building a framework for future research.

Bio-machine hybrids

The enactivist claim that biological cognition and computation are fundamentally different things raises interesting questions about another recent trend – that of bio-machine hybrids, or “animats”. A group of systems engineers and neuroscientists at the University of Reading, led by Kevin Warwick, have developed a robot controlled by cultures of organic neurons that is capable of navigating obstacles. Cortical cells removed from a rat foetus were cultured in nutrients on a multi-electrode array (MEA), until they regrew a dense mesh of interconnections. The MEA was then fed signals from an ultrasound sensor on a wheeled robot via a Bluetooth link. The team were able to identify patterns of action potentials in the culture’s responses to input, which could be used to steer the robot around obstacles.
Footage of the “rat-brained robot” in action, produced by New Scientist, currently has 1,705,230 hits on YouTube. Its performance is far from perfect of course, and to what extent such an artificially grown, stripped-down “brain” makes a good model for processes in a real brain is an open question. But the team hopes to gain insights into the development and function of neuronal networks which could contribute to our understanding of the mechanisms governing cognitive phenomena such as memory and learning. There is also the hope that by interfering with such a system once trained, new insights may be gained into the causes and effects of disorders like Alzheimer’s and Parkinson’s, potentially leading to new treatments.

The question alluded to above is whether such hybrids are constrained by the same limits to computation identified for “standard” computational devices such as Turing Machines (a mathematical abstraction which led to the development of computers). A related question arises from consideration of the emerging technology of neuronal prostheses. Given these technologies could be seen as extreme ends of a spectrum of systems with the potential to converge, it seems timely to ask whether, and at what point, their capacity in terms of notions such as autonomy, intentionality, and even consciousness, would also converge. Questions such as these where the focus of the 5th AISB Symposium on Computing and Philosophy: Computing, Philosophy and the Question of Bio-Machine Hybrids, held at the University of Birmingham, UK, as part of the AISB/IACAP World Congress 2012 in honour of Alan Turing, earlier this month.

AI and the NHS

A project which has had significant social impact in the UK is a system developed between the University of Reading, Goldsmiths and @UK plc, which makes it possible to analyse an organisation’s e-purchasing decisions using AI to identify equivalent items and alternative suppliers and so highlight potential savings. Using this SpendInsight system the UK National Audit Office (NAO) identified potential savings for the UKNHS of £500m per annum on only 25% of spend. Ronald Duncan, Technical Director at @UK plc, said: “The use of AI means the system can automatically analyse billions of pounds of spend data in less than 48 hours. This is a significant increase in speed compared to manual processes that would take years to analyse smaller data sets.” He added: “This is the key to realising the savings.”

An extension to the system also allows the ‘carbon footprint’ of spending patterns to be analysed and the results of a survey of civil servants suggests that linking this to a cost analysis would go a long way towards overcoming the institutional inertia that is currently costing the UK billions each year. In other words, strong environmental policies could, in this instance, save the UK billions in a time of financial crisis, through the application of AI technology.

Quantum linguistics

Finally, Bob Coecke, a physicist at the University of Oxford, has devised a novel way of simultaneously analysing the syntax and semanticsof language. The technique is being referred to as “quantum linguistics” due to its origins in the work Coecke and his colleague, Samson Abramsky, pioneered, which applies ideas from category theory to problems in quantum mechanics. 

Traditional mathematical models of language, known as formal semantic models, reduce sentences to logical structures where grammatical rules are applied to word meanings in order to evaluate truth or falsity and so draw inferences. Meaning tends to be reduced to Boolean logic by such models, which therefore don’t capture nuances such as the different distances between the meanings of “pony” and “horse” compared to “horse” and “cat”. Distributional models quantify the meaning of words by statistically analysing the contexts they occur in and using this to represent the word as a vector in a single, high-dimensional space. In this way, the horse-pony-cat distance might be quantified, in any number of dimensions, by context words like “ride”, “hooves” or “whiskers”. While these models have the advantage of being flexible and empirical, they are not compositional and so cannot be applied to sentences. 

Coecke, together with Mehrnoosh Sadrzadeh and Stephen Clark, devised a way of using the graphical maps he had used to model quantum information flow, to represent the flow of information between words in a sentence. Talking about this transition from quantum mechanics to linguistics, Professor Coecke said: "By visualizing the flows of information in quantum protocols we were able to use the same method to expose flows of word meaning in sentences. It is as if meaning of words is `teleported' within sentences."

This map, representing the grammatical structure of a sentence, acts on the vector meanings of the words, to produce a vector space representation of the sentence as a whole. In other words, this represents an approach to language analysis which is both distributional, so that meaning is derived empirically, and compositional, so that it is sensitive to syntax and applicable to sentences. As such, it holds great promise for a future leap forward in the automatic processing of natural language. It has been verified on a number of hand-crafted tests as “proof-of-concept“, and work has begun using real-world language samples.