I believe the best way to extend our scientific understanding of consciousness is to stop using the noun and investigate all the many mental processes that can and do occur in humans and other animals and future robots in very great detail and explain how they are possible. Then everything of substance about consciousness will have been covered, and the vacuous, incoherent unanswered questions generated in philosophical discussions will remain unanswered as they should be, because they are unanswerable.
My talk is an illustration of a small part of this project, starting from a comment made by Wittgenstein when discussing the experience of ambiguous figures. He wrote: The substratum of this experience is the mastery of a technique. I don't really know what he meant by that, but those words slightly modified thus: The substratum of an experience is mastery of a large collection of techniques available and ready to be deployed if required, possibly in new combinations. could be used to express a theory I am trying to develop in the context of trying to understand how to give a robot human-like (to be more precise, child-like) capabilities in the context of perceiving and manipulating 3-D objects.
The idea is that an infant-toddler-child-youth (and future domestic robot) develops by constantly actively and creatively exploring many aspects of the environment and thereby learning a very large number (possibly many thousands, certainly many hundreds) of different facts about the environment including facts about different kinds of stuff things are made of, different kinds of surface fragments that can occur, different kinds of ways things can be combined or decomposed, different kinds of relationships that can occur between simple and complex objects, different ways collections of relations can change, different kinds of actions that can be produced, and of course different consequences of all the above.
These facts are not expressed as propositions using what we would call a human language, but they must be somehow represented internally in a usable form, and in particular, for creative experiments to be performed and novel problems to be solved by combining prior knowledge the information must be recombinable in novel ways for some uses.
So, a child or future intelligent domestic robot is constantly learning orthogonal, recombinable, competences. (Actually, not totally orthogonal since independent variation of phenomena is limited in many ways, that have to be learnt.) It seems that precocial species either cannot do this or do it to a much more limited extent: they start off with the vast majority of what they need to know about the world and how to act in it pre- programmed by evolution (contradicting familiar arguments about the requirements for 'symbol grounding'). Altricial species that develop very complex and diverse cognitive competences probably evolved these powerful information acquiring, restructuring, mechanisms because (a) genetic mechanisms lacked the space to encode them and (b) evolutionary history did not provide all the opportunities that would have been needed to derive them.
Because they evolved for dealing with a world that is not only complex, but is also constantly changing, these abilities to cope with novel processes (i.e. perceive, represent, and use information about them) at very short notice had to be implemented in architectures that made them readily available to be invoked on demand in different combinations. I suggest that that fact determines requirements for the design and implementation of visual systems that have not yet been fully articulated. Moreover, the implementation will use mechanisms that have not yet been thought of by neuroscientists, psychologists or AI researchers.
One of the requirements for an organism that may need to monitor, evaluate, modulate and perhaps extend its own mental states and processes (e.g. improving its reasoning, problem-solving, learning, capabilities) is that it should be able to learn not only about the environment but also about its internal states. As with exploration of the environment, this could use a self-organising mechanism that adapts to what it encounters by chunking things and inventing labels for reusable chunks.
This could include labels for aspects of the contents of various sensory manifolds. Because of the manner of their development, such concepts will have a feature referred to as 'causal indexicality', i.e. their intension is intimately connected with their conditions of use. But because they are used for categorising states and processes in virtual machines that are not accessible by anyone else, these concepts will be inherently incommunicable: accounting for one aspect of what people who discuss qualia are trying to say.
When we have designed or discovered appropriate mechanisms for acquiring and using all these different competences, and the kind of architecture required to accommodate them I conjecture that this will explain a wide range of familiar phenomena including the variety of ways in which an individual can experience the world and some of the ways in which things can be experienced as ambiguous, flipping between different interpretations that make use of different competences (or 'techniques').
Some half-baked explorations of these ideas can be found in the html file referenced here http://www.cs.bham.ac.uk/research/projects/cosy/pa pers/#dp0601 COSY-DP-0601 Orthogonal Competences Acquired by Altricial Species (Blankets, string, and plywood)
One problem with the theory is that nothing I have learnt about brain mechanisms (on which I am no expert) seems to be capable of explaining how these competences are acquired, stored and recombined on demand.
For example, the kinds of models of neural nets that I am aware of just do not seem to be capable of meeting those requirements, though perhaps networks of networks could? Chemical information processing systems have more of the right features, but would probably be too slow, and could not easily be coupled with the processes that acquire and use the information.
It is possible that there are such mechanisms, but they have not been found because nobody was looking for them. They may be implemented in subtle ways as high level virtual machines on lower level physiological machines that seem to be doing something more mundane or something mysterious.
Note that the recombinability of orthogonal competences seems to require some sort of internal syntax. This could have been a crucial precursor to the development of external social language. It could not be based on human language because the learning and creativity I am talking about occur in prelinguistic children and some other animals.
These ideas have some echoes of global-workspace theory, though I think there are several workspaces of different sorts, supporting different kinds of concurrent processes in the architecture I envisage.
This work is partly inspired by collaboration with
Jackie Chappell who studies animal cognition,
especially New Caledonian Crows and
Parrots/Parakeets.