TITLE: New developments in the Meta-Configured Genome theory.
4pm Thurs 5th March 2020
Computer Science Common Room
Aaron Sloman and Peter Tino (Computer Science)
[Absent collaborator: Jackie Chappell (Biosciences)]
OVERVIEW (EXTENDED ABSTRACT) This joint presentation follows on from the presentation by AS on 2nd Sept 2019, on the idea of a Meta-Configured Genome (MCG), developed previously with JC, in his practice talk for the Models of Consciousness conference September 9 - 12, 2019 in the Oxford Mathematical Institute: https://models-of-consciousness.org/ Recordings of the conference talks are available here: https://www.youtube.com/channel/UCWgIDgfzRDp-PmQvMsYiNlg/videos Key MCG ideas, developed in collaboration with Jackie Chappell, are elaborated in some notes and a short video here: http://www.cs.bham.ac.uk/research/projects/cogaff/movies/meta-config/ After AS explains some of these ideas PT will make some remarks about how discoveries concerning gene expressions may be relevant to the implementation of our proposed mechanisms. The central idea of MCG is that an organism's interactions with its environment during earlier phases of gene expression can cause information to be collected that provides 'parameters' for use during later stages of gene expression that are only partly specified genetically. Examples where such parameterisation is required include genetically specified control mechanisms for producing various kinds of motion (e.g. walking, swimming, flying) where the details of control have to be changed as an individual becomes larger and heavier, and relationships between body parts change through growth. (Striking examples include the different motion patterns available to horses at different stages of development.) In many cases the parameters specifying control of action have to be acquired from the environment, e.g. sizes, weights and shapes of objects used as tools, parts of nests, or missiles! In the case of cared for infants, such as young orangutans, the mother will gradually extend the challenges of tasks that the infant has to master, e.g. by moving to different parts of foliage, remaining less close to the infant, and changing her speed of motion. In animals with high spatial intelligence this learning produces ablities to reason about novel situations -- using brain mechanisms that are not yet known, but may be related to the mechanisms used by humans in making mathematical (e.g. geometrical and topological) discoveries about possible and impossible structures and processes. Since statistics-based, probability-inferring, neural nets cannot represent impossibility or necessity, and since a great deal of genetically triggered or controlled development must depend on chemical processes involving the genome, we conjecture that some of this learning about structures, including discovery of impossibilities and necessary connections is implemented chemically, not neurally, e.g. using unknown sub-neural chemical mechanisms, that may be extensions of the chemically controlled developmental processes involved in construction of nervous systems and other parts of a foetus or hatchling. We know that chemistry-controlled gene expression is used to assemble extremely complex biological structures of many kinds during individual development (perhaps most spectacularly in building up a whole mature organism from a single fertilised egg, including in some cases transforming one whole organism (caterpillar, or larva) into another (so-called metamorphosis) there is a vast amount of sub-neural chemical activity, we conjecture that unexplained aspects of animal intelligence, including abilities to reason about The resulting information structures formed from patterns plus parameters can then provide parameters for later stages of gene expression. Such gene expression is inherently compositional, with multiple layers of compositionality. http://www.cs.bham.ac.uk/research/projects/cogaff/misc/compositionality.html This means that at least some of the later evolved, later expressed, genetic specifications are parametrised, and get some of their parameters used during individual development from previously stored information acquired during earlier stages of gene expression, by individuals acting in the environment when triggered by a variety of cues, often without expectation of reward, since the rewards cannot be known in advance. (This can be described as Architecture-Based Motivation (ABM), as contrasted both with Reward-Based Motivation (RBM), as explained in http://www.cs.bham.ac.uk/research/projects/cogaff/misc/architecture-based-motivation.html) and "Free-Energy Minimization" theories, e.g. https://www.frontiersin.org/article/10.3389/fpsyg.2012.00130 Stored information (e.g. about which parameters work in which situations) acquired as a result of such triggering, can be used later, when more recently evolved aspects of the genome are activated. E.g. word forms and word meanings acquired at one stage can be used as parameters for grammatical forms acquired later. The competences acquired can then be used as components for more complex competences that develop later. New "genetic layers", based on more recently evolved portions of the genome, generating new motives and behaviours, can be activated at various stages of development, e.g. into puberty and beyond. We'll try to show that this can produce far more powerful forms of learning than a multi-layer neural net created from an early stage, e.g. soon after birth. In particular, as AS and JC have previously suggested, the MCG theory seems to be able to explain the ability of a shared multi-layer human genome to produce multiple stages of language development in thousands of languages, that differ at all levels from the basic sounds used (or signs, in the case of sign-languages), up to modes of composition of phrases, sentences and beyond (e.g. However..., Therefore..., Although...). [This work is partly inspired by the linguistic theories of Chomsky and other linguists, without accepting that these features are unique to evolution and development of linguistic competences. It is also likely to apply to internal forms of representation that are much older than human languages some of which are shared with intellignt non-human species.] The MCG mechanisms may also be able to explain the development of spatial understanding over many generations, each extending the learning opportunities of successor generations, leading eventually to deep ancient mathematical and engineering competences, such as produced discoveries of Archimedes, Euclid and others -- still in use worldwide. JC's work on crows, parrots and orangutans suggests that there are also relevant (but simpler) non-human examples. Since genetic mechanisms are chemistry-based, the MCG theory implies that important aspects of learning and development in humans and other animals make use of changing sub-neural chemical structures and processes rather than the currently fashionable statistics-based neural net mechanisms (which are incapable of discovering or even representing topological and geometric theorems). Identifying those chemical mechanisms is mainly a task for the future. Peter Tino recently pointed out that changes in our understanding of the complexities of gene expression -- since the early theories following the discovery of the structure of DNA -- seem to be particularly relevant to the role of parametrised structures postulated in the MCG hypothesis. He will describe aspects of gene transcription discovered by biologists that vastly enrich the ways in which proteins are produced based on genome codes, going far beyond the original "Central dogma of molecular biology": genome-triplets of base pairs --> aminoacid --> protein. Here is an excellent collection of mini videos totalling just over 7 minutes, illustrating some of the processes involving gene expression: https://www.youtube.com/watch?v=7Hk9jct2ozY Many variants have been discovered, including immune responses, as summarised here: After an encounter with a new pathogen, the adaptive immune system often "remembers" the pathogen, allowing for a faster response if the pathogen ever attacks again. https://www.khanacademy.org/science/high-school-biology/hs-human-body-systems/ hs-the-immune-system/a/hs-the-immune-system-review This is *chemical*, not *neural*, remembering. We suggest that, in a similar way, a shared *meta-configured* genome implemented chemically can produce dramatically different effects in members of the same species in different physical terrain requiring different forms of behaviour, or in different cultures, in the same species, though using far more complex mechanisms than antibodies. This may explain why neurons contain such varied and complicated chemical mechanisms, not required for the simple arithmetical oprations postulated by neural net learning mechanisms. As remarked above, one of the most striking examples of what the MCG theory seeks to explain is the ability of the shared multi-layer human genome to produce multiple stages of development of competences in thousands of languages, that differ at all levels from the basic sounds used (or signs, in the case of sign-languages), up to modes of composition of words, phrases, sentences and beyond. We suggest that that power is used in many other contexts. The theory has deep implications for philosophy of mathematics and philosophical questions about consciousness, insofar as mathematical discoveries provide new contents of consciousness -- e.g. awareness of impossibilities and necessary connections, which could be related to spatial intelligence in other species, as well as a huge variety of human examples. This is part of the core of Kant's philosophy of mathematics, and does not seem to be explained by anything in current neuroscience: e.g. statistics-based neural networks cannot discover or even represent impossibility/necessity. This aspect of ancient mathematics is usually ignored in modern discussions of logic-based foundations of mathematics, e.g. using axiomatic systems -- of types unknown to ancient mathematicians -- or human toddlers, squirrels, orangutans and others with sophisticated spatial reasoning and manipulation capabilities. (There is no evidence to suggest that they are all unwittingly using logical forms of representation and have logic-based theorem provers in their brains.) Although there is a great deal of evidence for chemistry-based information processing in organisms, including chemical control of development of body parts and control mechanisms, including nervous systems and brains, I don't think anyone knows how those mechanisms could produce spatial intelligence including abilities to build and use tools, make nests, peel and eat bananas, crack open and eat nuts, manipulate unfamiliar objects and watch over their offspring as they develop spatial intelligence in increasingly challenging contexts. This research is still in its early phases. We hope this seminar will inspire other researchers with relevant knowledge and expertise to contribute relevant background information, criticisms, and suggestions for improvement. In particular, its implications for mathematical consciousness may require major changes in philosophical, psychological, and neural theories of consciousness. Alan Turing's thoughts Can digital computers ever replicate all the capabilities of sub-neural chemistry-based computations using a mixture of discrete and continuous processing? Perhaps Turing thought not, in 1952, when he published "The chemical basis of morphogenesis". He had previously (in 1936) distinguished mathematical intuition and mathematical ingenuity, suggesting that computers are capable only of the latter. (There are debates about whether he changed his mind later.) Time permitting, examples of various kinds of spatial reasoning, including familiar geometrical discoveries reported long ago in Euclid's Elements will be presented that provide challenges both for logic-based geometry theorem provers and statistics-based Deep learning systems. [A disorganised collection of examples can be found here, including links to more examples: http://www.cs.bham.ac.uk/research/projects/cogaff/misc/impossible.html] ==== NOTE: The evolutionary products we describe are examples of the *metaphysical* creativity of biological evolution, which continually produces new *types* of entity, illustrating the concept of metaphysical causation developed by Alastair Wilson, Birmingham philosopher. https://onlinelibrary.wiley.com/doi/abs/10.1111/nous.12190 ==== We invite anyone interested in learning more about this research, or in contributing to it in any way to contact any or all of: Aaron Sloman Peter Tino Jackie Chappell We are particularly keen to learn about information known to neuroscientists that could be relevant to explaining the phenomena we describe, including ancient forms of mathematical intelligence concerning spatial impossibilities or necessites in humans (some mentioned above) and related forms or spatial intelligence in other species. ==== Please report errors or omissions to Aaron Sloman (a.sloman AT cs.bham.ac.uk) http://www.cs.bham.ac.uk/~axs 5 Mar 2020