At
AAAI 2007 Fall Symposia
Call for participation:
(Leaflet
)
(Brochure)
November 9-11 2007
Tabula Rasa or Something Else? It may be of interest to see what can be done by giving a robot no innate knowledge about its environment or its sensors or effectors and only a totally general learning mechanism, such as reinforcement learning, or some information-reduction algorithm, to see what it can learn in various environments. However, it is clear that that is not how biological evolution designs animals, as McCarthy states: Evolution solved a different problem than that of starting a baby with no a priori assumptions. ....... Instead of building babies as Cartesian philosophers taking nothing but their sensations for granted, evolution produced babies with innate prejudices that correspond to facts about the world and babies' positions in it. Learning starts from these prejudices. What is the world like, and what are these instinctive prejudices? John McCarthy, 'The Well Designed Child' (1999) http://www-formal.stanford.edu/jmc/child1.html Members of most species are born or hatched with all the competences they will need (though they may be able to adjust minor parameters as a result of feedback). Some grazing mammals can walk to the mother's nipple and run with the herd very soon after birth, and chicks find their way out of the egg unaided, and can peck for food and follow a hen. This raises the question why other species, such as primates, hunting mammals and nest-building birds, seem to start so helpless and incompetent. This is especially puzzling in the case of species which, as adults, seem to perform more cognitively sophisticated and varied tasks, such as: hunting down, catching, tearing open, and eating another animal; building stable nests made of fairly rigid twigs (as opposed to lumps of mud) high in trees (a task you would find difficult if you could only bring one twig at a time); leaping through treetops; using hands to pick fruit in many different 3-D configurations -- and in the case of humans far more. Perhaps in those cases the appearance of starting totally incompetent and ignorant is very deceptive. Perhaps the prior knowledge of the environment provided by evolution in those cases is more subtle than in the case of foals and chicks. There is not just one problem with one solution We conjecture that evolution discovered more design problems and more design solutions than most learning researchers have so-far considered. In [1] and [2] we proposed shifting the precocial/altricial distinction from species to competences, arguing that within a species some competences may be 'precocial' (i.e. preconfigured in the genome, for instance, sucking in humans and some) while others are 'altricial' (i.e. meta-configured, namely produced epigenetically by preconfigured meta-competences interacting with the environment). Innate knowledge can be knowledge about how to acquire more knowledge Those meta-competences, far from being totally general learning algorithms, are specifically tailored to finding things out about environments containing 3-D configurations of objects and processes involving them, where objects have 3-D spatial structures (i.e. shapes, changeable shapes in the case of non-rigid objects) and can be made of different kinds of material stuff with different properties, where not all properties are detectable using available sensors. What is learnt though the application of meta-competences includes what sort of ontology is useful in the environment, as well as which laws using that ontology work well for making predictions in the environment. In addition, there are meta-competences which build on the early acquired competences to produce new meta-competences that extend the individual's learning ability. A university student studying theoretical physics could not have learnt the same material soon after birth. The ability to learn to learn can iterate, as indicated graphically in this picture:
So, not just the individual's knowledge about the environment is continually extended, but also its ability to learn new, more sophisticated things. Which layers of competence develop will depend not only on the learner's innate meta-competences but also on the particular features of the environment in which learning takes place. So, for example, a three-year old child in our culture will learn many things about computers and electronic devices that were not learnt by most of its ancestors. They probably started with the same sort of learning potential but developed it in different ways: bootstrapping can be highly context sensitive. Forms of representation for 'inner languages' In [3], we suggested that in some species the kinds of perceptual, planning, problem-solving, and plan execution competences that develop require the use of internal forms of representation that support structural variability and some form of compositional semantics; features normally assumed to occur only in human languages used for communication. But if some animals that do not use human languages, and pre-linguistic human children use these generalised languages (g-languages) for perceiving, thinking, planning, formulating questions to be answered, etc. then those representations must have evolved prior to the evolution of human language. If semantically rich information structures are available prior to the learning of human languages, that transforms the nature of the language learning task: for the learner already has rich semantic contents available to communicate, including possibly questions, and goals, depending on what the internal language is used for. This contrasts with more conventional theories of language learning, according to which the child has to learn how to mean and what to mean at the same time as learning how to communicate meanings. [Ref Halliday: Learning how to mean ??] There is no implication that all g-languages are restricted to linear strings of symbols or to Fregean languages using a syntactic form composed entirely of applications of functions to arguments. On the contrary, in [4] it was suggested that analogical representations are sometimes useful for representing and reasoning about spatial configurations. Analogical representations, including diagrams and maps are capable of supporting structural variability and (context sensitive) compositional semantics since parts of diagrams can be interchanged, new components added, etc. with clear changes in what is represented. In [5] it is further claimed that the ability to manipulate representations of spatial structures can be the basis of a kind of causal competence that enables a reasoner to understand why a certain event or process must have certain effects. This is the kind of understanding of causation discussed by Kant, in opposition to the view of Hume that the notion of 'cause' refers only to observed correlations, the predominant analysis of causation among contemporary philosophers for example. This Humean conception has recently been generalised to include conditional probabilities, as represented in Bayesian nets. We suggest that as progress in science typically starts with Humean notions of causation in each new domain, and then as deep theories regarding underlying mechanisms are developed the understanding of causation in that domain becomes more Kantian, allowing reasoning about structural interactions to be used, for example, to predict the effects of new events in new situations. In contrast, Humean causation supports only predictions concerning instances of previously observed types of events. This Kantian understanding of causation is closely related, in humans, to the ability to learn and do mathematics and to reason mathematically, especially the ability to acquire and use competence in proving theorems in Euclidean geometry. We don't know to what extent other animals are capable of Kantian reasoning, but the creativity shown by some of them suggests that they do have a Kantian understanding of causation in at least some contexts. Moreover, it is clear that for robots to have the same abilities as humans (or even nest-building birds, perhaps) they too will need to be able to acquire kinds of ontologies, forms of representation and theories, that allow them to use Kantian causal understanding in solving novel problems. For more on the problems of investigating causal understanding in non-human animals, see the presentation by Jackie Chappell at WONAC http://www.cs.bham.ac.uk/research/projects/cogaff/talks/wonac#chappell Ontology extension Learning about the existence of new kinds of stuff, new properties, new relationships, new events, and new processes that require the use of concepts that are not definable in terms of the ontologies that are genetically provided, requires learning mechanisms that support substantive as opposed to mere definitional ontology extension. (Which would be impossible if 'symbol-grounding' theory were true!) For more on ontology extension during development, see http://www.cs.bham.ac.uk/research/projects/cosy/papers/#pr0604 [To be continued, possibly]
References [1] A. Sloman and J. Chappell, 2005, The Altricial-Precocial Spectrum for Robots, Proceedings IJCAI'05, pp. 1187--1192, http://www.cs.bham.ac.uk/research/cogaff/05.html#200502 [2] J. Chappell, A. Sloman, 2007, Natural and artificial meta-configured altricial information-processing systems, International Journal of Unconventional Computing, 3, 3, pp. 211--239, http://www.cs.bham.ac.uk/research/projects/cosy/papers/#tr0609 [3] Aaron Sloman, Jackie Chappell, 2007, Computational Cognitive Epigenetics (Commentary on Jablonka and Lamb: Evolution in Four Dimensions), Behavioral and Brain Sciences, http://www.cs.bham.ac.uk/research/projects/cosy/papers/#tr0703 [4] A. Sloman, 1971 Interactions between philosophy and AI: The role of intuition and non-logical reasoning in intelligence, Proc 2nd IJCAI, pp. 209--226, http://www.cs.bham.ac.uk/research/cogaff/04.html#200407, [5] J. Chappell, A. Sloman, 2007 Presentations on causal competences, Kantian and Humean, in animals and robots http://www.cs.bham.ac.uk/research/projects/cogaff/talks/wonac/ International Workshop on Natural and Artificial Cognition Pembroke College, Oxford 2007
Maintained by
Aaron Sloman
School of Computer Science
The University of Birmingham