AISB/IACAP World Congress 2012 - Alan Turing 2012
AISB/IACAP World Congress 2012 - Alan Turing 2012
Varieties of Meta-Morphogenesis in the Bootstrapping of Biological Minds.
(DRAFT: Liable to change)
In his 1950 paper he also suggested that production of a normal human mind could be achieved by a powerful learning machine with a large empty memory, interacting with an environment containing a teacher. an idea that is often re-invented. In contrast, I suggest (as did John McCarthy in 'The Well Designed Child', in 1996) that human genomes, and many other animal genomes, specify large, extendable, collections of both information and competences, including control competences and learning competences, some of which interact with the environment and with each other to produce more sophisticated learning and control competences, while acquiring more factual information. For example, learning to perceive, learning to act, learning to think, learning to communicate, and learning to generate new goals and evaluate alternatives, all develop from primitive prenatal and neonatal versions, through mutual interactions, partially under the influence of the environment. These are forms of physical and computational morphogenesis. Later abilities to learn and develop may be influenced by earlier environmental interactions: a type of genetically seeded, but not genetically determined, meta-morphogenesis.
Like the mechanisms of physical morphogenesis, the collections of useful information and boot-strapping competences, including information processing competences, must have been acquired piece-wise across many generations over millions of years, with the previously acquired versions contributing to, or at least enabling, evolution of new forms, by providing new "platforms" on which evolution could build under changing pressures from the environment: producing new forms of morphogenesis. This is another type of meta-morphogenesis, which itself can take different forms. It may require some recapitulation of computational aspects of phylogeny in ontogeny.
Mechanisms for acquiring and using information, like mechanisms producing physical growth and physical behaviours, develop both within individuals and across generations. A genome that enables offspring to learn from observation of adults enables learning to be accelerated. A genome enabling adults to become aware of and to contribute explicitly to learning processes (using meta-semantic competences to provide various forms of 'scaffolding') allows cultural evolution to be very much faster. Mechanisms for one to many communication further speed up change, but also allow mind-binding religious indoctrination, and control of the many by the few.
Understanding all of this fully may be too difficult for human minds (especially those without experience of building information-processing systems), but use of AI experiments can help us develop, test and debug far more complex theories of evolution, development and learning, than ever before. We thereby alter our environment to support yet more rapid meta-morphogenesis.
A key feature of complex human-created information processing systems has been increasing reliance on virtual machinery, including stacked layers of virtual machinery and coexisting interacting virtual machines, and increasingly also self-monitoring and self modulating virtual machines, some of which could never have been implemented purely AS physical machines, even though they are implemented IN physical machines, and partly also in the environments of the machines.
Some of the information contents of self-monitoring virtual machines using self-generated ontologies, have several analogies with contents of minds (including the impossibility of physical detection from outside the machine and other sorts of 'privacy'), and this may provide clues to meta-morphogenetic steps required for evolution of biological minds.
We can see biological evolution as producing proofs: proofs of the possibility in a physical world of various kinds of biological machinery, including various kinds of physical structures and processes and various kinds of information processing capabilities. As in mathematics, different proofs can lead to closely related theorems. As in a complex mathematical proof, later stages can depend in complex and varied ways on multiple early stages, and different proofs can share earlier stages. But the biological proofs require not just one information-processing machine with a tape or memory that it controls completely, but changing collections of machinery interacting with each other and with a changing multi-dimensional environment, all supported by a reasonably large planet also going through geological and other changes.
This is a first-draft rudimentary theory of "meta-morphogenesis" that may one day show how, over generations, interactions between changing environments, changing animal morphologies, and previously evolved information-processing capabilities, might combine to produce increasingly complex forms of "informed control", starting with control of various kinds of chemical and physical behaviour at the level of large molecules then microbes, and so on, including eventually informed control of information-processing machinery. Later stages could explain philosophically puzzling features of animal (including human) minds, such as the existence of qualia; and also enhance our still incomplete understanding of requirements for future machines rivalling biological intelligence. We still have much to learn about the space of possible minds, and the requirements different sorts of minds need to satisfy -- many of which are unobvious, including consequences of embodied cognition that often go unnoticed.
Developing these ideas is a worthy Turing-inspired multi-disciplinary project for the 21st Century and beyond, unifying science, engineering and philosophy.
Maintained by
Aaron Sloman
School of Computer Science
The University of Birmingham