This report summarises views expressed in the third of three BCS Thought Leadership Debates run in association with the UK Office of Science and Technology as part of the Cognitive Systems Programme of the UK Foresight Programme, which explores areas for advanced research.
The event was on 10 February 2005 at the Institute of Directors in London. Two expert speakers stimulated debate with short talks and then the 40 participants discussed the topic in small groups over dinner.
The participants were mainly senior computer scientists and neuroscientists from UK universities, with others from specialist companies and research organisations. After dinner each table reported back to the entire gathering.
Cognition is something that humans do well and computers don't really do at all. The brain on the other hand can combine perception, learning and reasoning and take into account a plethora of issues as it decides on a course of action.
If we want to build machines that show cognition it would be useful to know how the brain works.
Although progress is being made in understanding how information is represented and processed by the brain, and in areas of perception such as vision, neuroscience still tends to offer theories rather than facts when it comes to cognition.
So a fundamental obstacle to progress in artificial cognition is that we have no way of even beginning to transfer human cognition into machines.
Can Computer Scientists Help Neuroscientists?
In neuroscience there are advances at the fundamental level of molecular biology relating to many aspects of brain function, at the level of cells and at other levels up to that of brain function: the 'integrated system' level.
What is missing is an architecture, a hierarchy of abstractions to join the various levels together.
Computer scientists can contribute here: they work on system architectures and routinely use hierarchies of abstractions to describe and understand complex machines. From this neuroscientists can gain a deeper appreciation of how the brain works and computer scientists can gain ideas on how to build better machines.
There is an issue over what an architecture means in this context, and whether computer scientists and neuroscientists have the same understanding.
Brains are complex systems, and computers are useful tools for simulating and understanding complex systems, so computer science can support neuroscience here.
Neuroscientists have drawn on aspects of machine learning, content-addressable memory and other areas understood by computer scientists, mathematicians and physicists.
Neuroscience involves the collection and analysis of vast amounts of data; this is clearly an area where computer science can and is helping. For example, neural networks are used in the interpretation of physiological data to study the action of synapses (the links between neurons) and other elements of the brain.
However, computer people don't want to be seen just as helpers to 'real' scientists: it might be their destiny, but computer scientists have hypotheses of their own.
Can Neuroscience Help Computer Science?
An obvious area where neuroscience can help computer science is that of technology interfaces with humans.
One category here is that of interfaces between a machine and the brain or nervous system: for example the use of implanted devices to control artificial limbs or to stimulate muscles affected by damage such as a spinal lesion.
The other category is that of cognitive interfaces, including systems that understand speech, or intuitively search databases - working much like human memory. This type of cognitive machine would be able to meet humans half-way. Humans currently have to adapt to systems to use them; life would be easier if machines had to adapt to human foibles instead.
Understanding of the brain could help in the design of database systems called on to handle large number of queries demanding big combinations of data items.
Different disciplines are gradually coming together. For example, computer scientists and brain scientists have been working separately on face recognition but are now collaborating: this has practical application in CCTV for recognising criminals in public places.
Modelling the Brain
One issue for links between computer science and neuroscience is that much of the computer science work on cognitive systems has little or nothing to do with biology.
Models are devised by computer scientists but we have no reason to believe the brain performs the functions in the same way: the fact that a computer model gets the same result as a human is not proof.
Could we build a computer model of the brain, and would it be any use to us in any case?
If we had the resources of the CERN nuclear research centre could we build a machine to scan the brain, map the 10bn neurons, each with about 1,000 connections, and then record every neural spike?
The data would not tell us how the brain works, because it would still require a great deal of interpretation, but it would offer a source against which theories could be tested in the spirit of e-science.
A separate and complementary idea is that of building a silicon neural substrate on a scale large enough to model the human brain; it would need 10 square metres of silicon, based on optimistic assumptions. This would form an experimental test bed for testing theories of brain function.
There would be issues surrounding the handling of such big machines. Much existing work on artificial neural networks is on a small scale, and it would be hard for humans to get to grips with such a huge and multi-dimensional machine.
There is also an issue of working out how to use such machines: how to configure a huge set of neurons to carry out particular functions we were interested in.
An evolutionary approach is arguably better than putting lots of money into developing such big structures.
Too Big a Challenge?
Perhaps the brain is too much of a challenge for computer science. Computer scientists have already been disappointed when asking neuroscientists how human memory works, so they could apply it to computers, only to find that their colleagues don't know. Perhaps we should start with something simpler, such as ants, instead of human brains.
Such disappointed expectations, the need for common understanding of terms like architecture in this context, and the technical terms used by computer scientists and neuroscientists in their separate fields highlight the problems of getting the two disciplines together.
The BCS might have a role here, perhaps by enabling people to publicise their areas of interest on its website so they can get contacts from others.
There is also an academic research funding issue. When disparate groups work together they can each get marked for funding by different criteria, and a group from one discipline might not be able to meet the criteria for the collaborative project as a whole.
Individuals Take the Initiative
Many major advances in understanding in this whole area have been made by individuals who have seen a problem and crossed traditional discipline boundaries to make progress.
So it is at least as important to help individuals follow an interesting question across disciplines as to encourage people in different disciplines to talk to each other. There is a danger that exhorting people to talk to each other creates a 'them and us' feeling, which is itself a problem.
One risk in the quest for machine intelligence lies in public understanding. Science fiction has prepared us for what might happen, so fears and phobias are established, ready to emerge at the first sign of the fiction becoming fact. We need to be prepared for this if or when the breakthrough comes.
A Middle Way?
If computer science and neuroscience are at opposite extremes in any discussion of how they can help each other in understanding the brain for their own ends, there is middle ground where much practical work is being pursued on robots and other devices that help elderly and housebound people and play an almost companion role.
These show some human characteristics but do not involve an understanding of how the brain works, and do not use large-scale computing.
A Never Ending Quest?
We cannot know yet whether computer cognition will take a decade, a century or a millennium. We will either succeed or keep trying until the age of machine intelligence begins.
The risks are high but the potential rewards for success great, and a multidisciplinary approach stands the best chance of making the breakthroughs needed to move forward on one of the final frontiers of science.