I will give a quick, cognitive/computer scientist-skewed overview of how the brain works. I will complain briefly about anyone who produces homogenous / monolithic models of animal-like learning or control systems. I will talk about attributes of dominant architectures for complete (animal-like) agents, and how they may correlate to things we know the brain does. If time permits I'll do a slide on Carruther's recent BBS article about language as THE cross-modular organizational structure and why I don't believe it.
I believe that attempting to specify and design a robot equivalent to a 3-5 year old child is a very ambitious goal. But I also believe that it is a worthwhile goal, and an achievable goal, but not necessarily in ten years.
The human brain and mind are extremely complex mechanisms but I believe that with appropriate resources and approach we can unravel much of their mystery sufficient to be able to achieve our goal.
We should probably support more than one research methodology. I wish to propose a methodology that is appropriate to a long-term research project of this type. I have used it successfully in two long-term research projects of my own: the mathematics understander (15 years) and imitation learning (10 years+).
Essentially we need to collect very detailed data about a child's behaviour. This data can be analysed in depth to determine patterns and generalities. From this emerges proposals for architectures, mechanisms and data structures.
Various strands of knowledge need to be linked together:
* neural architecture
* computational mechanisms
* computational theories
It is probably not possible to design the robot entirely bottom up from an analysis of a child's behaviour. Indeed it would be rash to ignore the vast amount already known about the brain's architecture and its methods of processing. Thus as we attempt to build an architecture for the robot it will have to relate to the actual architecture of the human child's brain, or at least our current understanding of it.
Similarly, it is likely that many of the computational mechanisms that we invent for the robot will be related to already known mechanisms in AI and computer science. But we probably need completely novel mechanisms and need to ensure that our vision is not blinkered by our knowledge of the existing AI mechanisms. Indeed few existing AI mechanisms are actually consistent with the sort of architecture that we will be proposing.
In relation to learning I believe that the human child has many different learning mechanism, and only a few of these are already understood by cognitive science. I propose that we should do detailed studies of one or more children from about 6 months until the age of five. These studies would be very invasive and clearly require parental consent and appropriate discussion of the ethical difficulties. But a full 100% video recording of a child as it grows up would be invaluable data for this grand challenge project. Ideally we would want five or more such children for study.
If we believe in the strong AI hypothesis, as I do, then it seems reasonable to me that we should be able to analyse the videos of the child growing up and produce a detailed predictive model of his or her behaviour. But we would need really detailed and complete data for such research to work. For example, in relation to language learning, if we know everything that the child has heard in the past, then we should be able to make reasonable predictions about future utterances in specific situations. Naturally as the child gets older exact and equal predictions become impossible because the child becomes autonomous and there are too many variables, not to mention randomness in the way behaviour turns out.
The architecture of brain and mind is an excellent grand challenge, but I believe that we need to ensure that our methodologies are consistent with the goal of the research. We must avoid just building a robot based on our current understanding of AI, cognitive science, neuroscience, linguistics etc. The project's goal should focus research in these disciplines working together to achieve the objective.
Early development in the human neonate is at its most intense during the first year of life. I will give a brief presentation explaining why this is so vital for machine intelligence and describe some work that aims to approach this difficult area by combining results from existing psychological data with algorithmic architectures that are compatible with both the neural substrate and the behavioural domain.
I will outline the case for putting the Imagination centre-stage in any architecture that could have a chance of fulfilling the ambitions of Grand Challenge 5. The Imagination, according to my definition, rehearses trajectories through an abstraction of sensory-motor space, and potentially forms the basis for endowing a robot with a capacity for spatial reasoning, planning, inner speech, and creative problem solving.