|UK COMPUTING RESEARCH GRAND CHALLENGES|
The UK Computing Grand Challenges Initiative
The mission statement for the UK Computing Research Committee (UKCRC) includes:The UKCRC aims to promote the vitality, quality and impact of Computing Research in the UK.
In October 2002, under the auspices of UKCRC, Tony Hoare and Robin Milner initiated discussions of "grand challenge" research projects in computing. A web site for the initiative was set up here: http://www.ukcrc.org.uk/grand-challenge/index.cfm
Seven main proposals came out of those discussions, listed here: here (GC-1 to GC-7): http://www.nesc.ac.uk/esi/events/Grand_Challenges/proposals/
That file includes pointers to summaries of each proposal produced earl in 2003, along with archives of discussion lists for each proposal, and an 'overview' or 'master' discussion list, GC-0. It was important that these were primarily scientific research projects aimed at increasing knowledge and understanding rather than practical engineering projects aiming to serve a useful purpose, although it was expected that advances in scientific understanding would inevitably lead to important new engineering advances.
Conferences were held in 2004 to discuss progress and make plans for promoting and extending the Grand Challenges initiative. As a result of continuing discussions during 2004 a booklet edited by Tony Hoare and Robin Milner was published by the British Computer Society (BCS) in 2004, summarising the current grand challenge proposals.
At the conference in 2006, additional challenges were added listed here. That web page also specified a procedure for proposing further grand challenge projects., although this is not currently available
This web site: Grand Challenge 5
This web site is about one of the challenges that emerged from the original discussions, "GC-5: Architecture of Brain and Mind". It is concerned with a multidisciplinary attempt to understand and model natural intelligence at various levels of abstraction, demonstrating results of our improved understanding in a succession of increasingly sophisticated working robots. The booklet mentioned above included a four page summary of GC-5 (also available in PDF format.)
This extremely ambitious project brings together work in neuroscience, cognitive science, various areas of AI, linguistics, philosophy, biology, and other relevant disciplines, so as to produce a new integrated theory of how a single functioning system can combine many human capabilities, including various kinds and levels of perception, different kinds of reasoning, planning, problem solving, curiosity, many varieties of learning (including grasping new abstract concepts and developing new fluent skills), many kinds of action of varying complexity, different uses of language, varieties of affect, including motivation and emotions, social interaction, and various forms of creativity.
Current robots can perform many tasks but each one is capable of only a very limited range of behaviours. Usually they do not combine perceptual and manipulative skills with the ability to communicate and cooperate, and they do not know what they are doing, why they are doing it, what difference it would make if they did things in a different way, etc., and they cannot give help or advice to another robot or a person performing such tasks. They do not have the variety of competences, the integration, or the self-understanding of a 4 to 5 year old (or child (or even a much younger child that can move about manipulate things, and interact with other people, not necessarily using language) and current robots can barely learn anything a child can learn.
Robots produced within this grand challenge project should at least have an interesting subset of the capabilities of a child aged somewhere between 2 and 5, including the ability to go on learning, and the ability (some of the time) to understand what they are doing and why. One way for such a robot to demonstrate all of that functionality would be being capable of helping a disabled person who wishes to avoid being dependent on other humans, at least around the house, without the robot first having to be programmed explicitly with knowledge about that house and its contents, and that person's needs and preferences.
Gaps in our knowledge
The summary of GC-5 in the 2004 booklet listed some of the gaps in our knowledge. Many processes in brains and minds are not yet understood, including how we:
These all involve both abstract mental processes and concrete physical processes. The project aims to explain how both kinds of processes work, and to demonstrate this in robots with a large collection of human-like capabilities, unlike current robots which are very limited in what they can do.
- see many kinds of things around us;
- understand language, like the language you are now reading;
- learn new concepts;
- decide what to do;
- control our actions;
- remember things;
- enjoy or dislike things;
- become aware of our thoughts and emotions;
- learn about and take account of the mental states of others;
- appreciate music and jokes;
- sense the passage of time.
Implementation in biological mechanisms
The aim of this grand challenge, is not merely to understand how such diverse functions can be integrated in single system at a high level of abstraction which might be modelled on computers or future information processing machines, but also to explain how they can be implemented in actual biological mechanisms. So an aim of the project is to continue developing our understanding of brain mechanisms (e.g. chemical, neural, etc. mechanisms) including showing how those mechanisms are able to support the high level functionality required by a child or robot. For this purpose, natural minds can be viewed as virtual machines implemented in brains. Since human minds surpass artificial minds in many ways at present, we may discover that this is partly due to using a different kind of physical implementation from current computers. There could be other reasons: it may be that our current designs for AI systems are simply far too simple because we have not yet understood what kinds of functionality they need nor what kinds of architectures, forms of representation and algorithms can provide those kinds of functionality in an integrated system.
Concurrent top-down, bottom-up and middle-out
Instead of taking sides on debates over whether it is best to use a top-down or bottom-up research strategy (i.e. starting from high level cognitive functions and trying to model them using any available mechanisms, or starting by trying to model biological information processing mechanisms in human or animal brains and then later using those mechanisms to implement higher level functions), the recommendation of GC-5 is that both extremes and hybrid versions should be pursued in parallel, with researchers of all kinds engaging in rich communication. A major feature of the top-down approach is the attempt to clarify the requirements to be met by the mechanisms produced in the bottom-up approach. At the same time, work on the mechanisms may help to clarify constraints on possible ways of meeting the requirements.
Three parallel activities
The 2004 summary of GC-5 stated that several mutually-informing tasks will be pursued concurrently:Task 1
Bottom-up specification, design, and construction of a succession of computational models of brain function, at various levels of abstraction, designed to support as many as possible of the higher level functions identified in other tasks.
Task 2 Codification and analysis, partly from a software engineering viewpoint, of many typical, widely-shared, human capabilities, for instance those shared by young children, including perceptual, motor, communicative, emotional and learning capabilities, and using them:(a) to specify a succession of increasingly ambitious design goals for a fully functioning (partially) human-like system,Task 3
(b) to generate questions for researchers studying humans and other animals which may generate new empirical research leading to new design goals.
Top down development of a new theory of the kinds of architectures capable of combining all the many information-processing mechanisms operating at different levels of abstraction, and testing of the theory by designing and implementing a succession of increasingly sophisticated working models, each version adding more detail.
Why focus on the capabilities of a young child?
The 2004 summary summary of GC-5 stated:As a 15 to 20 year target we propose demonstration of a robot with some of the general intelligence of a young child, able to learn to navigate a typical home and perform a subset of domestic tasks, including some collaborative and communicative tasks. Unlike current robots it should know what it is doing and why, and be able to cooperate with or help others, including discussing alternative ways of doing things. Linguistic skills should include understanding and discussing simple narratives about things that can happen in its world, and their implications, including some events involving capabilities, motives and feelings of humans. The robot could be tested in various practical tasks, including helping a seriously disabled or blind person cope without human help. This long-term target will be broken down into a large collection of sub-goals, with different subsets used to define intermediate milestones for judging progress. Achieving all this will require major scientific advances in the aforementioned disciplines, especially if one of the sub-goals is production of biologically plausible mechanisms capable of supporting the robot's functionality.Although there are many different forms of research that would contribute to the aims of GC-5, the proposal to aim for an integrated system with the capabilities of a young child was motivated by the following observations:
Success could also provide the foundation for a variety of practical applications in many industries, in unmanned space exploration, in education, and in the ever-growing problem of caring for disabled or blind persons wishing to lead an active life without being totally dependent on human helpers. Perhaps some people reading this will welcome such a helper one day.
Although many practical applications are possible, the primary goal is to increase our understanding of the nature and variety of natural and artificial information-processing systems. This is likely to influence many areas of research including psychology, psychiatry, ethology, linguistics, social science and philosophy. It could transform ideas about education.
- Young children provide an existence proof of the possibility of combining many different sorts of capabilities in a highly competent individual (including highly competent pre-linguistic individuals).
- A typical child that is only a few years old can be moved to any human culture and will be able to develop as a typical member of that culture, demonstrating that there is something very powerful and generic, presumably to a large extent biologically rather than culturally determined, in the child's capabilities: understanding how such a child develops will provide insights of great generality and wide applicability.
- Although many people have proposed starting with something that has capabilities more like a newborn infant, and this is not ruled out by the project, it was argued that newborn infants are much harder to study since most of what they do is very inscrutable. Understanding a later stage of development may provide a 'window' through which we can look backward in attempting to unravel the mechanisms producing that later stage.
Dependence on multiple disciplines
The 2004 summary also pointed out that although this was proposed as a grand challenge in computing research, because it was concerned with understanding very complex kinds of information processing systems, including both physical machines and virtual machines that process information, it depended on close interaction between researchers in several disciplines, including.
- neuroscientists studying brain mechanisms and architectures;
- psychologists, linguists, social scientists, ethologists and philosophers studying what minds can and cannot do;
- researchers in computer science and artificial intelligence developing techniques for specifying and implementing many kinds of abstract mechanisms and processes in present and future physical machines;
- researchers in mechanical engineering, materials science and electronic engineering extending materials and mechanisms available for robot bodies and brains.
No earmarked funding was allocated for the grand challenge projects. Instead it was left to proposers making use of standard grant proposal procedures to submit procedures, referring, where appropriate, to the grand challenge documents to legitimise what might otherwise be regarded as extremely risky or over-ambitious projects.