快猫短视频

WILL there come a day when a machine declares itself to be conscious? An increasing number of laboratories around the world are trying to design such a machine. Their efforts are not only revealing how to build artificial beings, they are also illuminating how consciousness arises in living beings too.

At least, that鈥檚 how those of us doing this research see it. Others are not convinced. Generally speaking, people believe that consciousness has to do with life, evolution and humanity, whereas a machine is a lifeless thing designed by a limited mind and has no inherent feeling or humanity. So it is hardly surprising that the idea of a conscious machine strikes some people as an oxymoron.

It鈥檚 certainly fashionable among biologists looking for the roots of consciousness to be suspicious of computer-based explanations of consciousness. The psychologist and writer Susan Blackmore insists that the brain does not directly represent our experience. She implies that constructing a machine that is conscious like us would be impossible. For her our 鈥渋nner sensation鈥� is some kind of illusion. She claims that any attempt to understand inner sensation scientifically is like 鈥渓ooking in the fridge to see whether the light is always on鈥� (快猫短视频, 22 June 2002, p 26). Susan Greenfield of the University of Oxford is another vocal objector to the idea of machine consciousness. In this magazine she has argued that such computer models 鈥渇ocus on tasks such as learning and memory which a PC can do without subjective inner states鈥� (2 February 2002, p 30). Greenfield uses a quaint engineering metaphor for consciousness, likening it to 鈥渁 dimmer switch鈥�: it grows and wanes as the number of brain cells working together rises and falls.

My view is that Greenfield鈥檚 theory does nothing to help us understand consciousness. And while her argument that researchers are focusing on tasks that a PC can handle may be true of the efforts of some, the computing research with which I am involved attempts to put flesh on what it is for living organisms to have memory and learning, which has nothing to do with the capabilities of PCs.

Blackmore鈥檚 contention that consciousness is an illusion is similarly useless when thinking about the mechanisms of consciousness. It is based partly on the assumption that no one can explain why some electrochemical activity in the brain is conscious and some is not. But I think that assumption is wrong. If anything, the problem is that there are several explanations to choose from.

Trying to explain the mechanisms that make us conscious is not going to be simple. But I am convinced that one way to face this complexity is to try to design conscious machines.

Laboratories around the world approach machine consciousness at a variety of levels. At one end of the spectrum are researchers creating detailed neurological models of the brain. At the other end are the unashamed users of pre-programmed rules that control the behaviour of an artificial intelligence, essentially a computer program that gives a particular output for a specified input.

The latter might seem a rigid approach that misses the whole point of creating consciousness, but Aaron Sloman of the University of Birmingham in the UK believes it neatly sidesteps the confusions and contradictions that surround what consciousness is. He argues that, when it comes to consciousness, nobody really understands what they are talking about, whereas the rules he writes are unambiguous. If these rules lead to apparently conscious behaviour in machines, they must form a basis for an explanation of consciousness.

According to Sloman, his creations are conscious within the virtual world; the computer itself is not conscious. With his colleague Ron Chrisley, he has built various virtual creations based on rules. For example, one rule might be: if an object appears to come towards me, enter the word 鈥渇ear鈥� in one of my memory banks. This rule is not emergent 鈥� it doesn鈥檛 develop automatically 鈥� but would be pre-defined by a programmer.

For some philosophers, subjective sensations like this are the central features of consciousness. Known as qualia, they include things such as redness, the sweet smell of a rose, and so on. 鈥淪ome of our designs are likely to produce systems that will discover in themselves the very phenomena that first led philosophers to talk about sensory qualia and other aspects of consciousness,鈥� says Sloman.

Closer to the centre of the machine consciousness spectrum is Bernard Baars, a psychologist at the Neurosciences Institute in San Diego, California, who has developed a model that accepts that the brain is different from a programmed machine. His idea, called 鈥済lobal workspace theory鈥�, represents consciousness as a phenomenon that emerges when a number of sensory inputs, such as images or sounds, activate competing mechanisms in the brain, such as memory or basic emotions like fear or pleasure. These momentarily activated mechanisms then compete with each other to determine the most relevant action.

Stan Franklin, a computer scientist at the University of Memphis in Tennessee, has turned Baars鈥檚 idea into 鈥渃onscious software鈥� called IDA (short for Intelligent Distributed Agents). Each agent represents one of the competing mechanisms in Baars鈥檚 model. Franklin has created a system using IDA to help the US navy automate the work of some personnel, such as deciding how and where to billet a service person when they come off a tour of duty. This work usually includes a great deal of human knowledge, judgement, understanding and emotion. The feedback IDA gets from its users is akin to the emotional feedback humans get for performing a task well or badly, Franklin says. This helps IDA improve the way it performs its tasks, by modifying the relevance value of the rules used in the task being appraised, so that it doesn鈥檛 repeat its mistakes.

Modelling living brains

Detailed neurological models of the brain lie at the other end of the machine consciousness spectrum. Rodney Cotterill, a neuroscientist at the Technical University of Denmark near Copenhagen, analyses brain scans from living brains, both human and animal, to identify the neurochemical interactions in the brain that he believes are essential for consciousness. On a computer, he builds simulations of these interactions to model how consciousness might emerge.

His idea requires us to understand the way in which an organism probes its environment and acts on it 鈥� the 鈥渁ction plans鈥� which he says are the basis of conscious thought. In the evolution of conscious organisms, the gaining of greater complexity and improved control of the environment must imply the emergence of thought. Cotterill points to specific structures in the brain, such as the basal ganglia and the cerebellum, that appear to have evolved to help the organism interact with its environment. He has produced a full map of the brain that highlights the role that many constituent parts play in generating action plans.

Going further in the direction of neurophysiology, Pentti Haikonen, a principal scientist at Nokia in Helsinki, has recognised that to model the activity in brain modules, such as the basal ganglia and the cerebellum, the corresponding artificial modules need to contain many neurons and be highly interactive. This suggests that it is no accident that real brain modules are made of vast numbers of neurons.

Haikonen鈥檚 work supports my own idea, which is based on an overwhelming body of neurophysiological evidence suggesting there are cells in the brain that compensate for motion, such as eye movement, in order to represent objects as they are in the real world. This allows us to get a sensation of the real world despite the constantly changing stream of sensory inputs, such as smell, vision and so on, that feeds our brains. To me, this evidence implies that our brains contain some sort of persistent representation of the outside world, encoded in the electrochemical impulses in their neurons.

And so my own design for a conscious machine starts by assuming that there is a neural 鈥渄epiction鈥� in the brain that exactly matches every scrap of our inner sensations. In order to form consciousness, these depictions have to have at least five major qualities (see 鈥淭he five axioms of consciousness鈥�). First, there is a sense of place. Depiction makes me feel that I am in the middle of an 鈥渙ut there鈥� world. Second, I am aware of the past. I know that depictions of the past can occur simultaneously with depictions of the present. Third, I can focus. I am conscious only of that to which I attend. Fourth, I can predict and plan. Depictions can occur which lay out in my mind alternative scenarios of the future 鈥� how the world might respond to my actions. Finally, I can feel emotions. Emotions guide me in my choice of which plans are good for me and which are not.

I believe that these five major axioms can be accomplished by what scientists have called artificial neural networks; these are simple approximations to the way neurons in the brain actually work. We have built machines that incorporate the first four axioms. The fifth axiom is still the subject of intensive work.

Computers hold the key

But how do we know that our machines really do have something like inner sensations? The key to all this is that depictions in a brain model can be displayed on a computer screen because we know exactly where the depictive neurons are (they are not confined to one region in the brain) and we can decode their messages. At the moment this cannot be done with a real brain as even the most accurate brain scanner only shows very roughly which parts of the brain are active. But demonstrations of sensory depiction, depictive memory, attention and planning all currently run on our machines.

We have applied the technique to investigate known deficits in the visual consciousness of Parkinson鈥檚 patients. They find it difficult to perform tasks that require them to plan some moves based on objects in the scene before them: making tea, for example. Dopamine deficiency is thought to cause Parkinson鈥檚. Our hypothesis was that the lack of dopamine somehow affected eye movement. In a coordination and mental planning study of 20 people with Parkinson鈥檚 we found that four of them became confused and did not know whether to, proverbially, take the lid off the pot, or pour the hot water in the pot first. In these cases, consciousness can be distorted because the lack of dopamine in the brain could affect the basic system of locking objects in a scene to a mental map in the patient鈥檚 mind. The moment the patient looks away from the pot, he forgets where it is, thus massively complicating the process of making tea. Our laboratory models show that this is a feasible hypothesis.

To verify our hypothesis that much of consciousness depends on muscular interaction with the world, we have also built a mobile robot equipped with most of the first four axioms. It has learned to 鈥渄evelop an interest鈥� in the objects in its environment so as to plan its movement from one to another.

Will building machines like this help us understand what it is to be conscious? I believe so. Are five axioms adequate? From a deep inner questioning of what is important to me in my own claim that I am conscious, the five axioms seem to me to be a necessary minimum. But the field is open for others to add to the list.

Of course, my robots will be infinitely less conscious of their worlds than I am of mine. But if their five axiomatic mechanisms are up and running, I wonder by what argument one could deny them their embryonic piece of consciousness? I may regret having said this, but I predict that machine consciousness will become a commonplace way of talking pragmatically about human consciousness. I would also predict that, in the same unspecified future, many machines will themselves claim to be conscious.