Cyberlife Research

Cyberlife Research is an english company set up by Steve Grand. Steve was one of the original authors of the hugely popular artificial life game "Creatures". He is now working to improve the artificial intelligence (AI) technology which he created for this game, and is planning to implement the resulting neural network architectures in hardware. The plan is to build a neural network processor for use in small autonomous robots.

The Alife/AI Philosophy:
The approach that Cyberlife takes to AI is that intelligence can only emerge from truly life-like systems. Most so-called "artificially intelligent" systems are nothing more than products of very clever computer programming. The belief at Cyberlife is that if you want a system that really behaves like an intelligent animal, then you have to actually build the complete animal.

This was the driving philosophy behind the game of Creatures. Instead of building real creatures, however, life was modeled in a virtual computer environment. The environment was designed to be as biologically realistic as possible. The creatures themselves were built from modeled brain cells, blood-streams, and biochemical reactions, whilst the environment modeled diseases, hunger, and the needs to breed and evolve.

About Creatures:
Creatures was first released as commercial home-entertainment software in 1996. It is still considered to be one of the most sophisticated Alife games available.

In the game, a virtual world is inhabited by various species called Norns (image right), Grendels, and Ettins, etc. The idea is that these creatures can be reared as virtual pets. Genetic principles are modeled to give every creature its own realistic idiosyncrasies and life cycle. Each creature possesses around 700 genes. Almost half the genes control the character and appearance of the creature. The rest are responsible for the growth of organs, a modeled biochemistry, and brain structures. When creatures mate the genes can be duplicated, mutated, or deleted.

The Norn Brain:
A Norn, the original species of creature in the first version of the game, has a brain which contains about 1000 neurons. These neurons are organised into 10 lobes with a total of around 5,000 synapses. Only the basic large-scale structure of the brain is genetically specified. The synapses develop themselves as the creature grows and learns. Despite this relatively small number of neurons, the brain architecture is very sophisticated and elegant. It is able to produce surprisingly complex and intelligent behaviour.

Drives, such as hunger, are registered in the Drive lobe. When a Norn sees food, the "Food" neuron in the Stimulus lobe fires. When both the hunger and food neurons fire simultaneously, the Concept lobe recognises that the Norn is hungry and food is available. This in turn causes an "Eat" neuron in the Decision lobe to fire which motivates the Norn to eat whatever food he can see. The decision lobe has 16 cells, each one represents a single action such as "eat food", "walk forward", or "pick up object".

Memories are formed through the Norn's continuous attempts to reduce its drives, i.e. through satisfying hunger or preventing boredom. If a certain behaviour is successful in reducing drives, then this behaviour is reinforced.

The Brain Processor:
Research is now underway to develop the Creatures' brain architecture still further. Theories are being tested about how mental models, such as body image and imagination, can arise from neural networks. This work is performed using a PC based neural network simulator.

The developed brain architecture will then be implemented in specialised hardware. This hardware will be composed of standard digital components such as memory, logic, and digital signal processors. It will be specifically designed for the rapid processing of complex neural networks.

Test Robots:
The first robot on which the new artificial brain architecture will be tested is a glider. It is hoped that the glider will learn by itself how to fly, eg. how to stay upright and how to find and use lift. This glider is currently (as of May 2000) being built, whilst the control theories are being tested using Microsoft Flight Simulator.

Further details:
For a more detailed review of Steve Grand's work, look here.

Links:
Cyberlife Research: www.cyberlife-research.com

Introduction

In the 1990's the company Cyberlife Technologies was founded by an Englishman called Steve Grand. Through this company Steve developed a computer game called "Creatures". This game is a virtual world in which the player can breed virtual pets to play with. The Creatures game has proven to be hugely popular and today there are believed to be over one million creatures living in these virtual worlds. By the late 1990's the game was considered by the artificial intelligence community to be the most realistic example artificial life in existence. (??? Is this still the case? Have there been any clones of the game?)

The phenomenal success of the Creatures game stems from Steve's refreshingly common-sense approach to that age old question "what is life?". His approach can be summarized by the sentence "life is more than clockwork, even though it is nothing but clockwork". This is the mechanistic view of the world. He also believes that although intelligence itself can not be simulated within a computer, it can be coaxed into emerging spontaneously from a network of simulated neurons and simulated biochemistry. It was upon this philosophy that the Creatures software was developed. Each creature is programmed to be a bundle of interacting life-like processes. The creature's brain is an artificial neural net which is influenced by the creature's biochemistry, it's immune system, and genetic makeup.

In the year 2000 Cyberlife Technology split into two companies. The games producing part of the company became Creature Labs and is now based in Cambridge, England. Steve Grand's part became Cyberlife Research and is based in Somerset, also in England. Cyberlife Research is a purely research company through which Steve is working on intelligent robotics. This chapter describes in further detail the Creatures game, Steve's philosophy of life and how this was used in the game, and lastly, the current research being undertaken at Cyberlife.

The Creatures Game

The Creatures game was first released for the personal computer in November 1996. Creatures 2 was released in August 1998. Both of these were set on the fantasy planet of Albia. The general aim of the game was to explore the planet and search for eggs. These eggs could then be hatched to produce little creatures called "Norns". The player cares for the Norns and watches them grow, just as if they were real pets. The Norns go about their world eating, playing, breeding and dying. The world is a rich environment containing several habitats such as damp caves and musical rooms. Scattered all over the place are lots of different types of food and toys. The world also has cycles of night and day, as well as changing seasons and weather conditions. The Norns are terrorized by other creatures called Grendels. A third life-form are the Ettins which are relatively harmless.

The current version of the game, as of 2001, is Creatures 3. This game is set on an enormous spaceship which contains various environments called "terrariums". There are still the same three species of creature, but many new breeds can be created. The newer creatures have a more advanced genetic makeup and greater intelligence. In March 2001 a smaller version of the game was released for free. This is called the Docking Station and consists of a smaller spaceship which can dock with the Creatures 3 spaceship, or even to other docking stations over the internet. The Docking Station software is similar to the full blown Creatures 3 game but is lacking in size of the virtual environment and the number of Norn breeds.

One aspect of the game which makes creatures so popular as virtual pets is that each individual creature is unique. The characteristics of a creature, such as their appearance and personality, are determined by their genes. These genes are mixed when the creatures breed. Also, the memories and preferences of a creature are constantly modified by experience. In addition to the core game software, a number of tools are available which can be used to genetically engineer the creatures and to meddle with their biochemistry. Possibly the most interesting of these tools is the so-called "brain-in-a-vat". This tool allows you to see, in real time, the activity of each neuron in the brain as the Norn is thinking and going about his business.

Behind the Creatures games is a proprietary artificial life technology which underpins a number of other products from Creature Labs. The company has recently developed "Sea-Monkeys" which is a game played across third generation mobile phone networks. They are also working on smart robotic toys, although these are still under development. Meanwhile the Creatures game is being ported to platforms other than the PC including, for example, the Nintendo gamecube.

The Philosophy

In his book, "Creation, Life and how to make it" Steve Grand explains his philosophy of life and how it was applied in the development of the Creatures game. In a nutshell, he maintains that the life and the soul of all intelligent creatures is not the result of any magical essence. To believe that the secret of life is some kind of physical thing which has not yet been discovered is to belong to a school of thought called "vitalism". Steve's approach, on the other hand, is very strongly mechanistic. This is to say that living systems, although awe inspiringly complex, are nonetheless built upon nothing but clockwork. Life emerges from endless interacting cycles of cause and effect. Steve then goes on to say that artificial intelligence will not be achieved in machines which are modeled on the outward appearance of mental processes, i.e. intelligence cannot be programmed directly. Instead machines must be built which model life's processes at a level lower than intelligence. Intelligence will then naturally emerge within these machines, without having been created directly. What's more, Steve believes that these intelligent machines needn't be created in physical form, but can be built within the virtual realms of cyberspace. This philosophy is elaborated further below.

Things as Patterns

Instead of categorizing everything, the entire universe can be thought of as one large dynamic pattern. Take for example a cloud formation often seen hovering at the top of a windy mountain ridge. The particles of moisture in the cloud are not stationary. In fact, each particle only exists within the cloud for a very short period of time. It is the difference in air pressures around the mountain ridge which cause the moisure particles to condense. Although the molecules of air and water are moving very fast, the pattern remains stationary. This imagery is used to explain that idea that things are very often not the things they are made of. Instead things are often just patterns which appear on top of dynamic processes. Even the human body is simply a pattern. Your body probably contains a completely different set of atoms to that of ten years ago.

The theory is that everything which exists is not solid but a persistent pattern. Even light and atoms can be thought of, not as solid particles, but as ripples in electromagnetic and gravitational fields. What is the true fundamental nature of matter? Even many physicists are not sure whether quarks, the building blocks of atoms, are real things or whether they are simply convenient theoretical objects which successfully describe the behaviour of atoms. This theory that "everything is a pattern" can be equally applied to life, intelligence, and consciousness.

Layering of Patterns

Most objects, or phenomena, mind their own business. Some objects, or patterns of objects, however, combine with others to produce new complex phenomena. For example atoms can combine to produce molecules. Molecules interact to produce networks of chemical reactions. These networks can then become encapsulated within bubbles of lipid, i.e. cells. This layering of patterns upon patterns can continue to produce a whole hierarchy, at the top of which we find life and intelligence.

Emergence and Complexity

Life and intelligence are not separate physical things. Instead they are patterns which emerge on top of other interacting patterns. Consciousness is not a property of matter, it is a property of the configuration of matter. This also means that consciousness, or your soul, can not be maintained separate from matter. Consciousness is a pattern which needs a physical substrate from which it can emerge. Crucial for us, however, is the idea that consciousness doesn't need to emerge on top of wet biological matter. It could also emerge in some kind of man-made dry matter, perhaps even silicon.

Emergence is a basic principle which needs to be understood before artificial intelligence can be created. In brief, emergence is when following simple rules leads to a complex and surprising result. Often the results can not be predicted without actually carrying out the rules. This is related to the mathematical theories known as complexity theory. In complexity theory, even though objects interact with each other using totally deterministic principles, the interactions are so complex that the outcome of interactions cannot be predicted. It is not that outcomes are just practically unpredictable, rather they are theoretically unpredictable. This is because even the slightest change in the initial conditions can cause a change in the outcome. It is impossible to measure initial conditions to absolute accuracy. For example, it may be possible to measure the temperature to an accuracy of a thousandth of a degree. But a change in temperature of millionth of a degree may cause a change in outcome. Even if you measure the temperature to a billionth of a degree, a yet smaller change may also change the outcome. And so on ad infinitum.

Complexity theory has interesting implications for free will. Indeed, it implies that there is no such thing as free will. Every action that you take is the only action that you could have decided to take given the initial set of conditions. The intitial set of conditions, however, are impossible to measure to 100% accuracy so your decision, although deterministic, is absolutely unpredictable.

Feedback

Another crucial part of nature is feedback. The action of positive and negative feedback can lead to highly complex interacting loops of cause and effect. These loops will tend to stabilize towards attractors. It is the use of feedback which enables adaptive behaviour. Although each individual feedback loop is mindless, the interaction of many millions of such loops can give rise to emergent behaviour such as intelligence.

Recreating Life

Having come to some conclusions about what life is, Steve Grand maintains that life can be recreated within the memory of a computer. His argument is that a computer is more than just a device for carrying out long lists of instructions, it is also a tool for creating another reality, a place called cyberspace. Within this space a form of life can be created which is just as real and life-like as the natural world. This may seem counter-intuitive at first, but the principle is based on the idea of layered patterns as mentioned above. Steve says that although a simulated atom is not the same as a real atom, a molecule built from simulated atoms has every right to be called a molecule. A simulated atom is an example of first order simulation. A molecule built from simulated atoms is an example of second order simulation. The levels of simulation needn't end at the second order. Third order simulations of networks of chemical reations can be combined to create fourth order simulations of living cells. This argument can be extrapolated to propose that although a first order simulation of life and intelligence is not real, a higher order simulation which emerges from lower level simulations can be considered real. It is in this way that Steve Grand hopes to coax intelligence into emerging from networks of simulated neurons.

This is a controversial idea. There are many neuroscientists and computer scientists who believe that true intelligence can never be recreated within a computer simulation, regardless of how many orders of simulation are involved. A counter argument to this is the following thought experiment. Supposing you were to replace just one of the neurons in your brain with a neuron simulated on a computer. Would you still be intelligent? The answer is undoubtedly, "yes". Provided the simulated neuron was sufficiently realistic, the real neurons to which it was connected wouldn't notice any difference between the simulation and other real neurons. Now supposing each and every neuron was replaced in turn with a simulated neuron. At what point would your brain cease to be intelligent? The supposed answer is that you wouldn't cease to be intelligent. Your brain would no longer be biological, but you would still be as intelligent and conscious as before.

Another idea which was central to the development of the Creatures game is that there can be no such thing as half an organism, or half intelligence. Steve demonstrates this with the obvious example that if a frog is chopped in half it will not survive for very long. What is less obvious is that intelligence is composed of many parts and although none of them is singularly responsible, they are all essential for intelligence. Examples of the components of intelligence are memory, generalization, imagination, etc. None of these components on their own constitutes intelligence, but they are all required. Traditionally artifical intelligence reasearcher have believed that intelligence can be abstracted out. The realization has just about arrived, however, that it can't be abstratcted, instead it is deeply linked in with, amongst other things, emotion and the fight for survival. For this reason the game of Creatures attempts to simulate not only the intelligence within the brains of the Norns, but also to simulate all other aspects of life. The simulation includes a genetic reproductive system, a fertility cycle, susceptibility to disease and a simple immune system.

In creating this simulation it does not matter that the physical layout of life is not replicated. What is important is that the informational layout is recreated. Thus the building blocks of life have been implemented on a computer, and from networks of these blocks emerges the vital spark of life. Although intelligence is an inherently parallel process, the brain contains many billions of neurons all working simultaneously and in parallel, any parallel process can be simulated on a sufficiently quick serial computer. Al that is required is a fast enough time-slicing loop. Such a time-slicing loop is at the heart of the Creatures game. Using this, together with feedback, and a little bit of artistry, the simulated ecosystem bubbles along without ever settling down. What's more, many of the events that occur are completely unpredictable.

The Creatures Brain

The creatures brain is built from simulated neurons and biochemicals. These enable the creature to learn for itself in a complex and realistic environment. It is able to remember past associations, to generalize from these associations, and also to forget. It is able to control its attention so as to focus on salient things. There are also a number of drives built in which motivate behaviour and learning. These drives are integrated with the body biochemistry.

Structure

The brain contains 815 simulated neurons which are divided amongst 15 lobes and form a total of around 5000 synapses. The neurons are arranged in a three layer feedforward network. This means that information enters the network through the sensory input neurons, it is integrated by the middle layer of interneurons, and is finally translated into actions by the output neurons. This large scale structure of the brain is genetically specified. The pattern of synapses, however, develops by itself as a result of the creatures experiences. The table below gives an overview of the function of each brain lobe together with the number of neurons per lobe.

[neuron count table]

Sensing and Attention

There are forty categories of object in the in virtual world of Creatures. The categories are such things as fruit, toys, vehicles, buttons, tools, etc. Each of these objects sends out information which stimulates the Creature's sensory neurons. This stimulation is short or long range depending on the type and strength of the stimulus. For example the sound of a Norn baby crying might be audible throughout the whole of the complex, but a flashing button may only be visible when it is within arm's reach. Every object has four properties: noise level, smelliness, visibility, and movement. The Norn brain has input neurons for each property of each category. This gives the Norn a total of 4 x 40, or 160, input neurons. This also means that the creature is capable of knowing about only one object from each category at a time.

The simulated neurons of the Creature brains do not signal via pulse trains the way that biological neurons do. Instead, when a Creature neuron is stimulated its output line, or axon, asserts a continuous value between zero and one. This value changes depending on the strength of the signal. When there are a number of objects of the same category within range of the Creature, only the nosiest, smelliest, or closest object is noticed.

The outputs of the sensory neurons are connected to the attention lobe. This lobe has forty neurons, each neuron represents one category of object. This lobe is wired as a "winner-take-all" network, only the object which is providing the greatest stimulus becomes the focus of attention. It is the combination of signal strengths from the various properties that determine which object is becomes the focus of attention. This ensures that the Creature is aware of only the most important object. For example, a toy may be less important to a Norn than a dangerous monster. But if the monster is sleeping and the toy is making lots of noise then the Norn's attention will be focused on the toy. (??? Need to check this network topology with the diagram of the brain - need to download and install the Docking Station)

Decision Making and Actions

The actions of a creature are motivated by a number of drives such as hunger, boredom, and tiredness. These drives sense the internal state of the body. For example the simulated biochemistry contains a glucose/glycogen cycle. When the glucose and glycogen levels are running low the hunger drive is increased. There are eleven of these drives and each one is represented by a neuron in the drive lobe. The eleven neurons of the drive lobe and the forty neurons of the attention lobe are connected together in the combination lobe. The combination lobe thus contains 11 x 40, or 440, neurons. The neurons of the combination lobe are connected by a set of weighted synapses. These connections determine the actions which are appropriate in response to the various combinations of input.

The output of the combination lobe are the 13 neurons of the decision lobe. The decisions are of the form "do X to Y", i.e. simple sentences combining a verb with a noun. Examples would be "eat fruit" or "press button". The action is always directed at the currently firing attention neuron. Thus, when a Norn sees food the "food neuron" in the stimulus lobe fires. When both the hunger and food neurons fire simultaneously the combination lobe recognizes that the Norn is hungry and that food is available. This causes the "eat neuron" in the decision lobe to fire and the Norn is motivated to eat the food.

The brain is wired so that activities are chosen which make the drive levels go down. The biochemistry, on the other hand, acts to make drive levels go up. For example, activity increases the hunger drive, the presence of a member of the opposite sex increases the sex drive. In response the brain decides to eat to reduce hunger drive, or to chase the potential mate in an attempt to reduce sex drive. Because the drives are in a constant state of flux with each competing with the others a creature will not continue doing the same thing for very long. For example, if a creature decides to play with a toy, at first the signal from this toy will become stronger because it is closer and possibly making more noise or moving faster. But eventually boredom will kick in and the creature will abandon the toy in favour of something else of interest. This simulation technique also allows for opportunism. If a creature has decided to walk towards something, but on doing so comes close to something else, the creature will first play with that something else before carrying on as to the original object as intended.

It should be noted that the simulated arms and legs of the creatures are not under neuromotor control. There are no simulated neurons which control movements such as lifting a leg, moving it forward, and placing the foot on the ground again. This would be far too fine-grained for the purposes of the game. Instead the individual motor neurons control higher level actions such as "walk forward", "turn left", etc. This means that the Creature doesn't have to decide to approach an object and pick it up before he eats it. Instead the decision is made simply to eat the object, all the actions inbetween are scripted.

In addition to the creatures being able to make decisions for themselves, actions can also be suggested by the game player. The player types in simple commands, such as "eat food" and these are set to the verb and sensory input lobes. The actions won't necessarily be taken, it depends on whether the Norn itself considers other actions more important. These commands function as a very primitive language. The Norns sometimes use this language to comment on what they are doing.

Learning and Memory

None of the actions mentioned above are hard-programmed to occur in response to certain stimuli. Instead, all responses are learnt as a result of experience. For example, a Norn learns that eating food reduces the hunger drive, this behaviour is not programmed from the start. The way that learning works is as follows. Each neuron in the combination lobe is initially connected to a random set of two or three sensory neurons. The action neurons are also connected to a random set of combination neurons. These connections migrate and change in strength as the creature learns. Whenever a neuron is activated it will stay active for a few seconds before slowly decaying over a period of about 25 seconds. Whenever a drive is being reduced the synaptic strength of the currently active neurons is increased. This process means that decisions which resulted in successful actions are reinforced and more likely to be carried out again in the future. Neurocomputational scientists call this "Hebbian Learning", after Donald Hebb who first came up with the concept of learning through reinforcement.

The creatures are also able to forget. There are only a limited number of available synapses, so if an newer response becomes more important than an old response, the older synapses will be broken down and new ones formed. A constantly sized pool of unconnected dendrites is maintained through the cunning use of biochemistry. All unconnected dendrites release a chemical which decays over time. All synapses also decay over time, their rate of decay, however, is related to the concentration of the dendrite chemical. This means that synapses will decay faster when there are no free dendrites than when there are plenty.

Through this dynamic balance of learning and forgetting the creature is continuously learning. New and important responses are learnt whilst old and less important responses are forgotten. This process is thought to be very similar to that which occurs in the brains of natural living things.

The combination lobe is also capable of generalizing from previous experience. Each neuron in the combination lobe is connected to neighbours which express similar concepts. If one concept activated, but as yet has no associated actions, then the similar concepts are activated. These will hopefully lead to a sensible decision. If this decision turns out to be good then a new dendrite is formed between the new concept neuron and the decision neuron. (??? what are similar concepts?)

Responses

Genetic Structure

The brain lobes and their tracts (the bundles of connections between lobes) are genetically defined. The genetics define a number of large scale rules such as each lobe can have a maximum of two input lobes. Because the brain structure is genetic it is subject to variation. An offspring my be born with two brain lobes where normally only one would be present. Each creature contains about 700 genes which are inherited as a mixture from the mother and father. Just as in biology, the genes are susceptible to mutation, delete and duplication. Natural selection is then applied to the Creatures offspring. Children with healthy brain structures will survive whilst those with damaged brains will perish. Not quite all behaviour is learnt after birth. There are a number of instincts such as "eat food". These instincts are still learnt, but the learning is programmed to occur whilst still inside the egg.

Latest Research - Intelligent Robotics

The cute little animals of the Creatures game are life-like enough to fool many people into becoming attached to them, just as though they were real pets. Indeed, the biologist Rupert Sheldrake(???) has even ventured to say that the creatures are pseudo-conscious. There is no doubt, however, that the creatures are not really sentient, they are by no means aware of their own existence. The reason for this is clearly the simplicity of their neural networks. Their brain is basically a three layer, feedforward neural network. Although it makes cunning use of biochemical feedback for continual adjustment of synaptic weights, biological brains are considerably more complex and involve many more layers of processing and make heavier use of neural feedback.

The creatures' brains were built under commercial pressure as a relatively quick solution for a computer game. Even at the time of development Steve realized that they were not as realistic as they could be. Now, through the company Cyberlife Research, he is taking the time to develop newer, more sophisticated neural network models. Many of the neural dynamics from the Creatures game will be used in the new models, for example the pool of available synapses, the techniques for weight modification and generalization, and directed and diffuse biochemical signaling. The most important new feature, however, will be the inclusion of a kind of imagination. The new creatures will not be conscious, but will possess a certain level of sub-consciousness.

Steve believes that imagination arose from neural structures which were originally evolved to handle self-nonself determination, body image, and focusing of attention. This belief is motivating him to work on a key neural sub-circuit which is capable of two internal states: sensory and mental. In the sensory state the network receives input from the real world and creates an internal map of reality. In the mental state the network explores this internal map in order to formulate desires, beliefs about the world, intentions, and expectations. The idea is that conscious beings do not live in the real world. Instead they live in a virtual model of the world which exists within their heads.

It is building this model of external reality which is the center of current research. How can such a model exist in neural substrate such that internal representations inevitably and automatically lead to the formulation of plans and the sequencing of muscular movement. The hope is that this neural circuit can be replicated many times to build up a complete brain. Further details of the neural sub-circuit can not be revealed here for two reasons. Firstly because it is still protected intellectual property, and secondly, and perhaps more realistically, because it hasn't yet been designed. These ideas are very much still a work in progress.

Intelligent Robotics

As a platform for developing his ideas about the neural substrate of imagination Steve has decided to step out of the virtual world and develop a real world robot. The robot, which he calls Lucy, is loosely modeled after a young orangutang. This robot is being built with no specific useful purpose in mind, instead it is a purely a platform for research.

The new neural network will exert a finer level of control over motor actions. In contrast to the Creatures game where a generalized decision would be made to "walk forward", for example, the robot brain will be able to command lower level actions such as move arm up, twist hand, etc. The first stage of development involved creating the peripheral nervous system, i.e. the eyes, ears, sense of balance, and proprioception. This will enable to the robot to see, hear, and to move it's arms. Initially it will consist of a head, torso, and two arms, but no legs. The eyes have a 1024 pixel peripheral field of vision, and 1024 pixel foveal vision. The motor control and sensory signal pre-processing are carried out by onboard digital signal processors (DSP's).

The robot's central nervous system will continue to be a simulated neural network running on an external personal computer. Each of the creatures in the computer game had about a thousand neurons. It is anticipated that without the need for a graphical user interface, and with increases in hardware speed, that the number of neurons in the newer networks can be increased considerably. Initially the network will have tens of thousands of neurons.

The ultimate plan is to build this neural network in specialized hardware. This could be used as a kind of "neural accelerator", or something which Steve calls a "psycoprocessor". This hardware would be built from standard digital components including memory chips, logic circuits, and digital signal processors, but it would be designed specifically for rapid processing of complex neural nets. It will contain many silicon neurons all working in parallel. No-one has yet figured out a way of creating silicon neurons which are able to rewire themselves in the way that is required, although there is a great deal of research in this field (see the chapter on aVLSI Neuromorphic Engineering). When such circuits have been built, Steve then imagines a whole brain being built from massive repetition of these small basic neurocircuits. The circuit units may contain even just a few dozen neurons. This unit may turn out to be very reminiscent of one which has already evolved in nature and is used in the neocortex. Networks of these units could carry out a vast range of different tasks depending on their configuration and interconnections.

Conclusion

Despite the relatively small number of neurons in the brain of a creature, surprisingly complex and life-like behaviour still emerges. This suggests that many of Steve's ideas about artificial life and intelligence may be on the right track. There seems little doubt that some of these ideas and technology will eventually be used in a truly intelligent artificial creature. It seems unlikely that the first intelligent robot will be built in a garage in Somerset, but it is significant that someone is at least trying.