Transitions in biological information-processing
Varieties of evolved, developed, learnt, invented...
forms of biological computation ____________________________________________________________________________
On 5th Jun 2012, Stuart Wray, after
reading a draft paper
on Meta-morphogenesis and the Creativity of Evolution:
produced this sketch of the ideas in the project ...
(CLICK TO EXPAND JPG)
CLICK HERE FOR PDF VERSION.
This is part of the Turing-inspired Meta-Morphogenesis project:
See also Tutorial on Meta-Morphogenesis at AGI Conference, Dec 2012
(DRAFT: Liable to change: Please do not save copies -- save a pointer.)
School of Computer Science, University of Birmingham.
(Officially retired philosopher/cognitive scientist
in a Computer Science department)
NOTE ADDED 18 Jan 2017
Some parts of this are superseded by or expanded in the discussion of the role of construction-kits in biological evolution and development:
Installed: 15 Oct 2012
22 Jun 2014; 26-27 Jul 2014; 8 Aug 2014; 15 Aug 2014 (Wray, Birner); 18 Jan 2017
25 Oct 2012; 2 Nov 2012; 19 Nov 2012; 25 Nov 2012; 30 Dec 2012; 3 Jan 2013; 14 Jan 2013;
6 Feb 2013; 5 Mar 2013; 2 Aug 2013; 19 Aug 2013; 8 Jan 2014; 27 Apr 2014
This file is
A PDF version derived from the html (possibly a bit older):
A partial index of discussion notes is in
(Please do not save copies of this document -- as they will get out of date quickly.)
The attempt to identify and analyse those transitions in information-processing is
the Meta-Morphogenesis project, so named because the mechanisms that produce the
transitions sometimes produce new mechanisms for producing such transitions: for
instance, some of the types of evolution, learning and development that exist on
earth now are themselves products of evolution, learning and development, and did
exist in the earliest life forms.
This document presents and attempts to explain the importance of a growing collection
of examples of transitions in information-processing capabilities in evolution, in
development, in learning, in society/culture, and perhaps also in ecosystems. The
transitions created by information engineers since the 1940s could also be regarded
as products of biological evolution (like the cathedrals built by termites), but for
now they are used merely to illustrate types of information-processing phenomena.
Recent information-processing technology provides several pointers to problems and
solutions that previously turned up in biological evolution (e.g. the advantages of
control by virtual machines rather than physical machines, when virtual machines are
easier to design, monitor, debug, modify, extend and combine with other mechanisms,
as explained here.)
Others have asked some of the questions raised here, but I am trying to collect a
wide variety of examples of transitions that may show patterns not visible to
researchers in a single discipline focusing on narrower sets of examples.
For example, natural selection may seem to be a uniform process, but what it
does depends both on the mechanisms generating options between which selections
can be made, and the selection mechanisms. The points summarised above imply
that both the types of options and the selection mechanisms can change
For discussion of ways in which learning and development in an organism can
produce results that combine with genetic mechanisms to influence later
development see Chappell and Sloman (2007). Closely related ideas are in
Although the scope of this project seems to be larger than most others, this is not
the first work to be concerned with evolution of information processing mechanisms.
A similar concern can be found in many other publications, e.g. here's a tiny sample:
Modularity in Development and Evolution
Eds. Gerhard Schlosser, Gunter P. Wagner
University of Chicago Press, Chicago, 2004
Living is information processing; from molecules to global systems,
K.D. Farnsworth and J. Nelson and C. Gershenson, 2012,
At home in the universe: The search for laws of complexity,
Penguin Books, 1995,
Chapter 15 of Margaret A. Boden,
Mind As Machine: A history of Cognitive Science (Vols 1--2),
Oxford University Press, 2006,
Note added 2 Aug 2013: I have been reading Merlin Donald's 2002 book
A Mind So Rare: The Evolution of Human Consciousness
The book is spoilt by excessive rants against reductionism, and a
seriously ill-informed account of symbolic computation, but is a superb
introduction to many of the evolutionary transitions that involve
information-processing, e.g. Chapter 4. Donald seems to understand the
importance of the fact that what exists now, e.g. in human minds, builds
on many layers of previously evolved function and mechanism which may be
shared with many other species. He raises many important questions about
how and why various features of human minds evolved, even though he lacks
(or lacked) the engineering expertise to provide deep answers.
A common thread in the work on evolution of information processing is the importance
not only of the sensorimotor morphology of organisms, and the mechanisms in brains
and nervous system, but also the nature of an organism's environment, the problems it
poses, the opportunities it provides, and the kinds of information-processing systems
required for dealing with it.
In the last few decades there has been much emphasis on the importance of embodied
cognition, or enactivism. I think it will turn out that much of the work done under
that banner, especially the polemical pronouncements, merely illustrate the dangers
of following narrow fads instead of trying to get a deep understanding of the variety
of design requirements for organisms and robots, and the variety of possible
solutions and their trade-offs.
In particular a narrow approach to the study of embodied cognition tends to emphasise
the importance of "online intelligence" as if "offline intelligence" either did not
exist or had no major biological function, whereas I argue that it is crucial to
understanding the variety of types of affordance and their perception and use (going
far beyond the ideas of James Gibson on affordances). This is also essential to
understanding human mathematical and scientific theory-building competences, for
example. The distinction between online and offline information-processing is
discussed further below.
For a more detailed critique see:
Some Requirements for Human-like Robots:
Why the recent over-emphasis on embodiment has held up progress
In Creating Brain-like Intelligence,
Eds. B. Sendhoff, E. Koerner, O. Sporns, H. Ritter, and K. Doya, pp. 248--277,
Sources of variety in types of Meta-Morphogenesis:
For any biological (e.g. genetic) changes B1, B2, B3,.. etc. and for
any environmental states or changes E1, E2, E3,... there can be influences
of the following forms ...
It is clear that evolution, learning, development, and cultural changes produce new
biological information used in reproduction and in many forms of behaviour.
However, the mechanisms for producing new forms of information-processing have
themselves been changed -- including new forms of reproduction, learning,
development, cultural change, and "unnatural selection" mechanisms such as
mate-selection, animal and plant breeding, and more recently cloning and use of
genetic manipulation to control reproduction.
The meta-morphogenesis project seeks to identify (a) all such changes in
information contents, and information-processing mechanisms and their
consequences, especially the many unobvious changes that are needed to answer
old philosophical questions and shed light on the relations between nature and
nurture and relations between minds and brains, and (b) the processes and
mechanisms that drove those changes.
If identifying all (including future) changes is impossible, we can attempt to
identify as diverse a range as possible, with as many intermediate points as
possible, along as many divergent evolutionary lineages as possible.
In the earliest phases of evolution, the mechanisms, and the changes in information
Ideally where we should start....
(Picture by NASA on Wikimedia: protoplanetary-disk.jpg)
Conjecture: In similar ways, new products of biological evolution, and products of
its products, enhance evolution's ability to produce more complex products.
This document presents some examples of transitions in information-processing competences,
starting with very simple cases and moving to increasingly complex examples, but without
presuming that there's a fixed order in evolution, or development. The diversity of
possible trajectories, is clearly indicated by human learning and development and by
differences in evolutionary lineages. Whether there are any absolute restrictions on
possible trajectories is a question to be investigated later.
NOTE: A failure to recognise diversity in developmental and learning trajectories can ruin educational systems for many learners.Many of the transitions in biological information-processing are closely connected with
Some of the competences are illustrated and discussed briefly here:
Other transitions in information-processing were required to allow attention to be
switched between objects, events and processes in the environment and objects, events and
processes in perceivers, for example the ability to notice, when looking at unchanging
external structures from a moving viewpoint, the changeable intermediate results of
perceptual processing, such as aspect ratios, optical flow patterns, texture gradients,
and assumed but unperceived parts, e.g. 'far sides' of objects. Such changes in contents
of awareness have produced philosophical puzzles about the relationships between
experience and reality, since ancient times. (Think of Plato's Cave, for example.)
NOTE: Hayek also had ideas about evolutionary transitions expressed in his book
F.A. Hayek, 1952, The Sensory Order
For example, he wrote (page 82):
Jump to CONTENTS list
The concept of "information" used by the Meta-Morphogenesis project is not the
technical concept introduced by Claude Shannon in 1948. The older, more familiar
non-technical concept of information about something, with content that may be
correct or incorrect, and which can be used in formulating questions, forming
intentions, controlling actions, forming explanations, making predictions, and
helping others, was already familiar to Jane Austen in 1813, as demonstrated in
In that sense information is what is sometimes called "meaning", or "semantic content".
This is completely different from the relatively new usage of the word introduced by
Shannon, in 1948, which subsequently confused philosophers, composers, scientists, and
many others. In this document I never use the word in Shannon's sense.
Moreover, most of the attempts to define the older meaning are either erroneous, or
circular, or else misleading in various ways. Like many powerful theoretical terms (e.g.
"matter", "energy", "gene", "electrical charge", "valence",...) the word "information"
cannot be explicitly defined. Rather it is implicitly defined by the theories in which it
occurs, which through their structure partially identify a class of models, which can be
more precisely identified by adding links with observation and experiment to the theory --
a process I call "theory tethering", not to be confused with the seriously misleading notion
of "symbol grounding". The concept of information is discussed more fully in
A. Sloman, What's information, for an organism or intelligent machine? How can a machine or organism mean?, In, Information and Computation, Eds. G. Dodig-Crnkovic and M. Burgin, World Scientific, pp.393--438, 2011 http://www.cs.bham.ac.uk/research/projects/cogaff/09.html#905The notion of "representation" is often defined in a very narrow way, e.g. by specifying
A. Sloman, 'The mind as a control system', in Philosophy and the Cognitive Sciences, Eds. C. Hookway and D. Peterson, pp. 69--110, CUP, 1993, http://www.cs.bham.ac.uk/research/projects/cogaff/81-95.html#18 A. Sloman, What enables a machine to understand?, Proceedings 9th IJCAI, Los Angeles, pp. 995--1001, 1985, http://www.cs.bham.ac.uk/research/projects/cogaff/81-95.html#4__________________________________________________________________________________
Jump to CONTENTS list
Added: 14 Jan 2013
Place holder: roles of information in control
We can distinguish different sorts of functional roles for mechanisms involved in use of
information for control, in biological and non-biological systems. Examples:
Many thinkers discussing information processing or computation consider only formal
manipulations of structures within some sort of machine, e.g. a Turing machine or
computer. This raises questions about how any semantic content can be involved in what the
machine does. We can answer those questions by explaining how information can be related
to control, whether in organisms or human-made machines. Ian Wright attempts to present
and extend these ideas in this slideshare presentation (slides and audio):
Such uses of everyday language in asking scientific questions can be seriously
misleading because the concepts are not based on a deep explanatory theory, and
as a result group together things that are superficially similar but deeply
different (like sharks and whales, both originally thought of as fish) or treat
as different things that have deep commonalities, e.g. use of manufactured
tools, like hammers, cutters, spears, and use of kinds of pre-existing matter to
perform manipulations on other objects, including the use of body parts.
For example, use of one hand to hold an object that is being peeled by another
hand, has deep structural and functional (mathematical) similarities with the
use of a space between two rocks to hold something, or use of a manufactured
vice to hold the object being manipulated.
Of course, there are also differences, but without understanding how the
differences and similarities interact (e.g. providing extra functionality in
some cases) it's possible to miss important cognitive processes.
More subtly, asking questions about whether or when human infants, or other
animals, have or acquire concepts like "enduring object", "causation", "false
belief", "number", "error", or "emotion", will typically cause researchers to
group together processes, competences and mechanisms that are deeply different,
or fail to notice similarities between examples that are superficially
different, like the similarities between marine mammals and land mammals that
were initially not noticed.
Another source of deep traps is the word (or concept) "language". When
which animals can learn or use a language? or when do children start to use language?they nearly all make use of a shallow common-sense notion of "language" as
Another example is the ordinary concept "teaching", which some biologists have
attempted to apply to animal behaviours often ending up squabbling about what is
or is not real teaching. (Compare Nigel Franks on Teaching in tandem-running ants)
Since I have no option but to use ordinary language for most of this project,
still in its early stages, I try to specify what I am asking, conjecturing, or
proposing by giving examples. But in many cases I am likely to be guilty of the
mistakes I have just criticised, and I welcome critical analysis of examples,
from the point of view of a designer of working systems, showing that my
examples need re-organisation or re-labelling. The ultimate test will be ability
to contribute to a broad, deep and precise, explanatory theory that can be
applied to both the explanation of natural phenomena and to the construction of
Conjecture: Learn about possibilities before learning about utility
The more complex the organism and the more possible internal and external actions
available, the more important it is to explore possibilities and their consequences
before those explorations have any evident utility. (See the discussion of
architecture-based motivation, below.)
Jump to CONTENTS list
Below is a very sketchy summary list of examples of transitions in biological information
processing in evolution, development, learning, etc., some of which also involve
transitions in physical structure or new sensors, many of which are related to
changes in the environment, e.g. new problems, challenges, dangers or opportunities.
In some cases the transition is initially merely related to acquisition or
manipulation of information, without any practical application, though, as the
history of mathematics shows repeatedly, such 'useless' changes can later provide the
basis for massive practical advances.
[NB: the numbering of points below is likely to change, as new items are inserted
and old ones rearranged, or merged, or split.]
Changes in abilities and ontologies include (in no significant order): Ontologies required for learnable abilities including: abilities to label, abilities to describe, abilities to predict, abilities to manipulate directly or to control indirectly, abilities to perform actions of varying complexity and difficulty, abilities to classify, abilities to explain, abilities to evaluate, abilities to appreciate, abilities to plan, abilities to carry out plans, including modifying or extending them during execution, abilities to design things, abilities to make things to satisfy a need, abilities to prove things, make inferences, calculate or reason, abilities to discover things, abilities to communicate (using language or other media), abilities to understand communications, abilities to teach or assist others, abilities to collaborate as leader or subordinate or equal, abilities to empathise, abilities to introspect in various ways, abilities to resolve conflicts, within oneself or between individuals, and many more related abilities.
All those abilities are generic and may have sub-cases that have to be learnt separately, and in some cases the learning can include increasing speed, fluency, reliability and accuracy of performance.
Intelligent abilities require use of knowledge about types of things that can exist or happen, i.e. knowledge of an ontology. A simple homeostatic controller, e.g. a thermostat, may use a very simple ontology perhaps including contents of a form of sensory input (e.g. temperatures) and contents of a form of output e.g. 'raise' or 'lower' signals to a heater or cooler.
Organisms (and future robots) with multiple modes of sensing and acting on a complex independently existing environment need ontologies that straddle modes of perception and action, for instance the ability to express where something is or how it is moving, irrespective of how the object's location is sensed or changed.
If we had a better understanding of how various ontologies used for various
purposes in organisms evolved, and how they develop in individuals, we might be
better able to design machines with intelligence that matches those of animals,
including humans. Instead we can now only produce machines with very shallow and
restricted abilities, that often turn out to be very brittle when dealing with
For more on changes in ontologies and associated forms of representation, see:
E. BremerIt is likely that several homeostatic mechanisms developed in the earliest life forms and perhaps their precursors.
"Synthesis and uptake of compatible solutes as a microbial defence against osmotic
and temperature stress",
It seems very likely that during evolution of some species several changes of continuous control mechanisms occur, either directly through alterations in the genome, or indirectly through improved learning mechanisms. However this document is more concerned with intelligent deliberation where control is not continuous. [To be expanded]
(There are very many biological mechanisms that do this, some relatively simple, e.g. phototropism, geotropism, hydrotropism??, others much more complex, e.g. carnivores seeking prey that can move, or animals seeking mates.
Note that in general change detection requires more complex mechanisms than detection: e.g. it may require storage of previous information to be compared with new information. So puzzlement about change blindness is mis-directed: change detection, not non-detection, is what primarily requires explanation.)
-- combined to detect more complex phenomena -- used to detect unrelated phenomena in parallel, leading to possible conflicts in reactions (e.g. choosing what to consume, what to avoid).
Some of these kinds of discretization will be done by individuals separately, while others may depend on developing a consensus among members of a community, and others may result from species differentiation -- producing different ways of dividing a continuum (plants and their pollinators?).
Far more relations are important in interacting with a complex environment than unary predicates, yet very many writers seem to assume that all or most concept formation is formation of unary concepts (predicates), e.g. straight, square, box, shoe, house, dog, etc. Contrast different ways of generating new concepts:
Unfortunately current AI students seem not to learn about some of the deep pioneering work done in the late 1960s and early 1970s in which use of structural descriptions, including comparisons of structural descriptions was central, for example the work by T.G. Evans on a geometric analogy program which inspired work by P. Winston on learning structural descriptions from examples, and G.J. Sussman on A Computational Model of Skill Acquisition, referring to thinking skills, not physical skills (both of which are important in humans and many other species).
One of the deep questions related to this is how the differences are represented between
and similar examples relating to past and future, or different locations (what could or could not happen here and there).
It is sometimes proposed that such information contents require use of a modal logic, that adds operators such as "possible", "impossible", "necessary" and "contingent" to a formalism for expressing facts. But that presumes that all information is represented propositionally (in a form expressible in sentences), but, for many reasons, including observations about mathematical competences below I suspect the ability to think and reason about counterfactuals uses architectural extensions (illustrated by the work of John Barnden and Mark Lee on counterfactuals and ATT-Meta).
Frank Guerin informed me that his three and a half year old son asked at a restaurant "Why didn't we get them last time?" when the restaurant provided wet serviettes for wiping children's hands. This requires
There are anecdotes about other animals being able to remember past events involving individuals who have helped or harmed them. [REFs needed.]
Compare: toddler theorems about numbers and numerosity (often confused by researchers).
Some ideas about
this are presented in this discussion of the speed, power, and flexibility of human
A Multi-picture Challenge for Theories of Vision
How do you know that if a vertex of a planar triangle moves along a median away from the opposite face the area of the triangle must increase, no matter what the size, shape, orientation, colour or location of the triangle? Compare the two cases (a) and (b) in the figure below? I suggest this uses deep functions of animal vision that have mostly been ignored.
Abilities to perceive and reason about possibilities and constraints on possibilities in a mathematical context are deeply connected with the ability to perceive and reason about affordances, which must have evolved earlier. This sort of requirement is one among many aspects of cognition that are blindly ignored (as opposed to being temporarily postponed) by most researchers on "embodied" or "enactive" cognition. More examples are e.g. here, here and here. Compare "toddler theorems" about how what's visible through a doorway to another room changes as you move your location relative to the doorway in various directions.
For example, I know of no animal whose visual system uses a lens to project light onto a regular rectangular grid of optical sensors, as almost all artificial visual systems do. The physical and functional design features of biological eyes may provide deep clues to the information processing functions and mechanisms of natural vision systems that have largely been ignored.
An overview of some of the functions of vision in humans is under construction here and http://www.cs.bham.ac.uk/research/projects/cogaff/misc/vision
A larger project is to identify major transitions in biological visual information-processing since the earliest forms. I have many online papers and presentations related to functions of vision, and will later attempt to organise them. One way to do that is in terms of the 3x3 CogAff Schema grid, (outlined here), which combines three columns of functionality:
and three layers of functionality, listed here from the bottom up:
Many evolutionary and developmental transitions are concerned with either adding new kinds of functionality within these layers or columns, or connecting functionality in different parts of the grid, across columns or layers, to develop more complex systems, e.g. producing social actions (such as smiling, beckoning, teaching, that involve not just the low level motor control system but also meta-semantic competences generating intentions and actions, and visually interpreting actions and responses of other agents.
Because of the nature of this grid there are many possible sequences in which particular competences can be added by evolutionary and developmental processes.
Although the CogAff grid/schema provides a useful framework for thinking about design alternatives it must be considered as a very crude approximation, especially the obviously inadequate implication that there are only 9 major subdivisions among types of information processing.
In particular, he attempts to explain how human vision (and presumably also vision in some other animals), can use the information sent to the primary visual cortex which is constantly changing because of saccades and other eye movements as well as head movements and movements of the whole body. His "retinoid" theory proposes a constantly changing mapping between the retinal information in V1 and the enduring information structure that encodes what is seen.
I would summarise this by saying that instead of regarding F1 as the first level of visual processing in the usual manner, we should regard it as an extension of the hardware evolved for collecting visual information (collecting photons).
It is part of a "sampling" mechanism for rapidly sampling different portions of the optic array (Gibson), using saccades, with the samples immediately "forwarded" to several other subsystems for further processing and for absorption into various enduring information structures holding different sorts of information about contents of the environment, not contents of retinal stimulation -- which would be of far less importance to many subsystems in an intelligent animal or machine. Since the main sampling is done by the high resolution fovea, and the fovea will not find any gaps in the optic array, the other subsystems do not obtain information about gaps that one might suppose the "blind spot" on the retina would produce.
A corollary is that theories that associate contents of self-consciousness (contents of visual self-awareness) with contents of V1 (in humans) may be completely misguided, since useful meta-cognitive mechanisms for inspecting current contents of visual processing will normally need to know now what the raw, unprocessed data are, but what various results of intermediate data are. (Compare the sort of visual self-awareness needed by a good artist drawing or painting realistic pictures.)
Whether all the details are correct (including Trehub's proposal that information is stored in regular grid structures, which I doubt) a theory of that general sort has several merits in addition to explaining why no "blind spot" is perceived, even during monocular vision. There are still many unanswered questions about forms of representation used and types of processing associated with various aspects of visual intelligence.
We may be in a better position to answer them after the meta-morphogenesis project has uncovered many more intermediate stages in the evolution of current animal vision systems. [REF DISCUSSION PAPER IN PREPARATION]
Much is known about physical, chemical, and morphological aspects of many kinds of eyes, and also about their functions, which typically depend on the needs of the organism, its optical sensor morphology, the features of the environment (including available food, predators, mate-features), the actions of which the organism is capable, and the available types of information processing mechanism.
Several animals have two or more eyes, and in some cases these seem to operate as independent sensors. But humans, and various other animals, including primates, hunting mammals, and many birds seem to be able to use two eyes pointing roughly in the same direction to drive two collaborating streams of information processing to compute distances of perceived objects by triangulation.
Unfortunately, Julesz and others discovered that humans are able to see 3-D structures in random dot stereograms, and this led many researchers to assume that the methods required for doing that, by first finding low level correspondences in the images, are used for all stereo vision. However, most natural scenes do not produce random dot patterns on the retina, and it is easy to confirm that a great deal of 3-D structure can be seen monocularly (e.g. try wearing an eye-shield for a few hours). So it is at least possible that animal systems use the results of monocular perception to identify corresponding locations in the left and right percepts and use those correspondences to perform triangulation. I offer this merely as an illustration of how easily an experimental discovery leading to a large tranche of computational modelling can distract research attention away from an important biological function.
Conjecture: binocular depth perception first occurred as a transition from monocular perception that made use of monocular structure to find corresponding items in left and right visual fields to use for triangulation. Later, additional mechanisms evolved to deal with cases like textured surfaces or sand dunes, where the monocular percepts do not provide sharp points of comparison. Normally the two mechanisms function in parallel. If this conjecture is correct it could lead to much improved stereo vision systems in robots, primarily using the results of monocular vision. (Perhaps that has already been done.)
Some references on evolution of binocular vision (Added 2 Nov 2014)
Pettigrew JD (1986) Evolution of binocular vision. Visual Neuroscience (Sanderson KJ, Levick WR, eds.), pp 208-222. Cambridge, UK: Cambridge University Press. http://www.uq.edu.au/nuq/jack/BinocVisEvol.pdf
Andrew N. Iwaniuk Peter L. Hurd (2005) The Evolution of Cerebrotypes in Birds Brain Behav Evol 2005;65:215-230 DOI: 10.1159/000084313 http://www.psych.ualberta.ca/~phurd/papers/BBE_05.pdf
Both of these miss the requirement to identify 3-D features of shapes that may or may not be visible from all views, and which may or may not be relevant to possible uses and behaviours of the object, and may take account of various combinations of topological properties of the objects, metrical properties of the object, qualitative semi-metrical properties, e.g. constancy, increase or decrease of curvature of part of a surface, or "phase transitions" in orientation or curvature, e.g. regions where curvature changes from concave to convex or vice versa, regions where curvature is constant, and many more.
When can a typical human infant use vision to take in the information required to answer such questions? Which other species can do it? What forms of representation and capabilities had to evolve to make such tasks possible?
Many mathematically well educated researchers assume that animals (and intelligent machines) must express all those spatial properties and relationships using an ontology that assumes an all encompassing space, with global metrics for length, area, volume, curvature, orientation, angle, etc.
However, it is far from obvious that many animals (or any animals) can do that, and moreover there are many unsolved problems about how to derive such information from visual and other sensor data. Perhaps that is a poor analysis of the problems evolution and its products solved.
For various reasons, to be explained later, I suspect that the ability to think about, make use of, or acquire information expressed in terms of such global metrics, e.g. using a global cartesian coordinate frame, or a global polar coordinate frame, is a very late, relatively sophisticated achievement (only developed in 1637 and thereafter by Descartes, Fermat, and their successors -- without which Newton's mechanics would have been impossible).
An alternative ontology might instead make use of collections of spatial and topological relationships between objects, and object parts, where the relationships could be binary, ternary, etc., with partial orderings of size, area, volume, angle, distance, direction, straightness, curvature, regularity, where some of the relationships are detected and represented in far more detail than others, e.g. relationships between objects or surfaces (including surfaces of manipulators) in the immediate environment, or relationships between objects on which actions are being, or are intended to be performed. A network of partial orderings of size, distance, could be enhanced by semi-metrical relationships, e.g. A is longer than B, and the difference is more than three times and less than four times the length of B. If B is a pace for a walking animal that could be relevant to choosing routes for walking. Different kinds of information about partial orderings might be relevant to grasping and manipulating objects in the immediate environment.
There's lots more to be said about the alternatives, their biological uses, their evolution, their development in individuals, the forms of representation used, the forms of reasoning, their roles in perception of different sorts (visual, haptic, auditory, or multi-modal perception, or a-modal reasoning), and about how organisms differ. E.g most of this would be impossible for microbes.
I suspect this ability to perceive and reason about semi-metrical partial orderings is part of what accounts for the early discoveries leading to Euclidean geometry, including the examples summarised here, and that in humans many transitions in representation of spatial structures, relationships, processes and interactions occur in the first few years of life that have not been noticed or studied by developmental psychologists. (Though Piaget seems to have thought about some of them.) They also have not been noticed by roboticists, especially 'enactivist' roboticists who focus mainly on online intelligence ignoring offline intelligence, briefly mentioned below.
Too often researchers think that what they do effortlessly needs no explanation -- so they look for explanations of failures (e.g. change blindness, lack of "conservation", etc.) instead of first looking for explanations of successes, without which it is impossible to construct explanations of what goes wrong.
Some examples of matter-manipulation competences:
I suspect that biological evolution changed the information processing architectures of some organisms so as to allow more intelligent 'look ahead' to guide choices, or to allow different exploration strategies to be selected explicitly on the basis of information available instead of being 'hard-wired' in search strategies.
Such transitions have happened many times in the history of programming language
development. An example was the transition from the Planner AI language to the Conniver
language at MIT in the early 1970s.
[Ref Sussman and McDermott, 1972
There are other transitions where failures discovered during a search process can be found
to be detectable at an earlier stage, an example being the process of "compiling critics"
modelled in Sussman's Hacker program [[REF]].
Compare recording 'ill-formed' substrings during parsing to constrain future search, and the development of 'caching' mechanisms, mentioned below.
[[Other examples of transitions from cognition to meta-cognition needed.]]
Finding out when such transitions occurred in evolution will be much harder. Some cases of occurrence in individuals may turn out to be examples of what Karmiloff-Smith refers to as "Representational Re-description" in Beyond Modularity (1992)
Conjecture: There are MANY more transitions made explicitly by programming language designers that are analogous to transitions made implicitly in changes of biological information processing.
http://www.cs.bham.ac.uk/research/projects/cogaff/talks/wonac 1. Evolution of two ways of understanding causation: Humean and Kantian. (PDF), 2. Understanding causation: the practicalities 3. Causal competences of many kinds
These mechanisms, and related mechanisms involving architecture-based motivation discussed above, are probably deeply involved in the later development of processes involving aesthetic enjoyment -- creating, observing, taking part in processes and structures that do not necessarily directly serve any obviously biological need, e.g. some of the play in young mammals. [Much more needs to be said about this.]
For example, servo control can make use of physical/mechanical compliance (as in use of padded skin, or physically compliant manipulators) or virtual compliance e.g. allowing perceived/sensed changes in relationships and in forces to affect changes in applied forces, where the perceptual component could be haptic/tactile, kinaesthetic or visual. See http://www.cs.bham.ac.uk/research/projects/cogaff/misc/information-based-control.html
Such advances in online intelligence do not necessarily provided advances in offline intelligence, e.g. the ability to think about the past, or future, or what might have happened under different conditions.
Varieties of deliberation are discussed here.
See also this discussion of some of Karen Adolph's work on young children: http://tinyurl.com/CogMisc/online-and-offline-creativity.html
Note (modified 25 Nov 2012): Serious muddles about ventral and dorsal "streams" of visual processing arise from failure to understand the different information processing requirements for (a) servo-control based on transient, constantly changing (mostly scalar?) information, and (b) acquiring and storing information for possible multiple uses at different times, including describing, planning, answering questions, etc. Referring to these as "where" and "what" functions betrays a deep but common failure to understand requirements for working systems of different sorts. The more recent replacement of these labels with the labels "action" and "perception" betray a failure to appreciate the variety of functions of perception, including its role in online intelligence. See On designing a visual system
Examples: a robot, like Boston Dynamics' BigDog produces very impressive behaviours. But it does not know what it has done, what it will do, what it hasn't done but could have done, why it did the one and not the other, what would have happened if it had selected a different option, what options might be available in a few seconds time, what the consequences of those various options are, and many more.
Many questions need to be answered: What sorts of evolutionary transitions led to such counterfactual-metacognitive capabilities in humans? Which other animals have them? At what stage do they develop in children, and how?
How does all that relate to the "proto-mathematical" ability to look at a triangle and ask what would happen to the area if one of the vertices moved relative to the opposite side, as illustrated here?
Note: there is much more to be said about "offline intelligence" and how almost all the research inspired by a concern with embodiment, dynamical systems, enactivism (etc.) fails to address some of the deepest aspects of biological intelligence (a recurring theme here).
So the ordinary concept of "language" like the ordinary concept of "tool-use" does not pick out a well defined scientifically useful class of phenomena, and diverts attention away from deep similarities between things for which we do not use the same label in ordinary speech. (Compare the unobvious similarity between graphite and diamond.)
Infinite competence behind finite performances. (Inserted: 22 Jun 2014)
A feature common to both (a) human languages (spoken, signed, or written) and
(b) internal forms of representation required for percepts, goals, preferences,
plans, questions, puzzles, explanations, ... including some of the contents of
non-human animal minds, is that a type of form of representation is required that
(i) structural variation in what is represented,
(ii) variation in complexity of what is represented,
(iii) no fixed limits on the complexity that can be accommodated,
(iv) some sort of compositional semantics insofar as the content represented
in a complex structure is systematically related to the contents of the
parts of the structure and the current context.
Condition (iv) is essential for coping with structural novelty, e.g. perceiving, or thinking about, or intending to construct, something never previously encountered. If everything perceived, intended, hypothesised, explained, or predicted could be adequately represented by a collection of N scalar measures then all variability would be accommodated by the possible set of values those measures. But many animals evolved abilities to cope with environments in which objects, processes, and actions exist with much more complex forms of variation, including structural variation (illustrated by the variety of plant and animal forms and the variety of ways in which different collections of plants, animals, non-living matter can be assembled and interact -- for instance the variety of processes involved in assembling a nest from twigs, or weaving a nest from leaves).
An observation due to Chomsky half a century ago (although he did not express it like this) is that any economical solution to the problem of being able to cope with the kinds of variation that can occur in human experiences will have the potential to cope with infinite variation. We now know that infinite potential (or competence) can exist in a computer, for example insofar as it includes and can execute a recursive definition of the factorial of a number) despite it having performance limits due to memory limits, or memory addressing limits. Those performance limits can co-exist with the infinite competence in the factorial procedure, which works in principle for any integer input, no matter how large. How biological information processing systems actually implement that infinite competence is an empirical question. How many forms of solution to that problem and by what evolutionary pathways is a problem for the M-M project.
Eventually a deep theory of evolution of forms of information processing will need to account for the various intermediate stages that can occur, both in evolution and in individual development, both in humans and in other intelligent species. A particular requirement that follows from (iii) is that no matter what the size limits of individual brains, the organisms need forms of representation and mechanisms of operation, that in principle have infinite generative competence. As Chomsky noticed in connection with human languages, actual performance may be limited by various aspects of the implementation.
Some of the issues are discussed in connection with human language in
Ian Roberts, 2009,
The Mystery of the Overlooked Discipline: Modern Syntactic Theory
and Cognitive Science,
Unfortunately, most linguists who understand this point don't seem to realise that it is a special case of requirements for animal intelligence encountered long before the evolution of human intelligence and human language. See also Chappell and Sloman (2007)
Immanuel Kant had similar ideas in relation to human mathematical competences.
Sometimes educational policies that try to emphasise 'understanding' at the expense of 'memorising' miss an extremely important function of memorising as an aid to altering the level of complexity of what the learner can understand.
Presumably there was some sort of evolutionary transition between being able to work out plans or solutions to problems and having mechanisms for storing results of such computations for future use.
A closely related, but more subtle development is the ability to remember discoveries about what does not work, so as to reduce the risk of following false trails in planning, reasoning, designing, doing mathematical reasoning, etc. Compare the work of Sussman on 'compiling' critics mentioned above.
Transitions of this sort occurred several times during the development of programming languages in the 20th century. A tutorial introduction to use of patterns in programs manipulating list structures is here.
Grammars can be used not only for linear structures, like sentences, but also for things like networks. Some of the early AI research in vision (in the 1960s) made use of "web-grammars", i.e. grammars for networks or graph-structures, to express the contents of visual percepts.Many researchers seem to assume that grammars are relevant only to languages used for communication, ignoring requirements for internal information processing in animals and machines.
For more complex species, evolution seems to have "discovered" the advantages, especially as life-spans increase, of more powerful ways of enhancing the genetically specified design, to cope better with threats, opportunities and constraints in each individual's environment, i.e. replacing pre-configured with meta-configured competences.
In some cases, e.g. in humans and some other altricial species, evolution also seems to have discovered the advantages of not only slowing down physical development while information processing mechanisms adapt to each individual's circumstances, but also staggering the onset of various kinds of later learning that build on the products of earlier learning: delayed activation of a meta-cognitive learning mechanism allows it to start looking for patterns in what "lower order" mechanisms have discovered when the patterns are richer and more stable, instead of wasting effort analysing patterns that are spurious, because based on too few instances and tests. This may be specially important in cases where learning cannot easily be undone. These ideas are developed in a little more detail in Chappell and Sloman (2007).
These are crude analyses: far more details, based on far more examples, are needed.
All living things have semantic competences insofar as they can use any information at all, whether with external or purely internal referents. A subset seem to have meta-semantic competences regarding themselves or others. These competences may be genetically fixed for some species and in others may develop under multiple influences (meta-configured competences, mentioned above). In humans many kinds of social/cultural education, and in some cases therapy can enhance meta-semantic competences, whether self-directed or other-directed (e.g. getting better at telling whether your actions are upsetting someone).
Human infants (and perhaps the young of some other species) need to develop a variety of meta-semantic competences, some self-directed some other-directed, some combined with counterfactual reasoning (e.g. "what could I have done differently?", "How would A have responded if I had not done X?", "Can A see this part of X?", "Can A tell what I can see?", "What does A think B did?"). Psychologists have used the label "mind-reading" for this sort of capability, but mostly restricted it to a small set of competences involved in working out what another believes or thinks, especially in situations where they don't have up to date evidence. This is just one of many cases where a fashion for a particular kind of research has spread because it is easy to vary experimental details, without ever thinking about the kinds of mechanisms required to make any of the competences involved possible at all.
For example, meta-semantic competence requires an architecture that supports referential opacity as well as referential transparency -- the usual default. Referential transparency refers to properties of representations where replacing item A in a larger representation referring to object O with item B also referring to O makes no difference to what is represented by the whole structure, and whether it is true or false. For example, if Fred is chairman of the club, then if it's true that the chairman of the club is a cricketer, then it is also true that Fred is a cricketer. But in a referentially opaque context, e.g. "Joe believes that the chairman of the club is a cricketer" replacing "the chairman of the club" with "Fred" can turn a true statement into a false one, or vice versa.
Some researchers favour trying to model such effects by extending the language used with a new operator (e.g. "believes that") and modifying normal inference rules. I suspect that what is really needed is a change in the architecture, to support a separation between information structures accepted as true (beliefs) and the information structures that represent possibilities that are not accepted. This is essential for planning, and for perception of affordances.
"John is in the kitchen and Mary is in the pantry"Moreover if he is in a corridor looking for his mother and shouts "Where are you?" and hears an answer from the end of the corridor "I'm here" then he knows she is either in the kitchen or in the pantry because those are the only two rooms leading off the end of the corridor and he can see that she is not in the corridor. If he then goes and looks in the kitchen, and finds that she is not there, he correctly infers that she is in the pantry. In order to do that he does not need to follow some rule:
P or Q Not-P Therefore QRather he merely needs to reason that as there were only two possible locations and one is ruled out the other one is the correct one. He may use the word "must" to express this "She must be in the kitchen". That is the "must" of logical necessity, but it doesn't come from a rule -- it comes from constraints linking the alternatives and the observed facts in this situation. The rule may later be formulated as a generalisation of the inferences made in such situations. But the articulated rule need not be involved in understanding the necessity that makes the inference valid.
A. Sloman, 1968/9, Explaining Logical Necessity,
Proceedings of the Aristotelian Society, 69, pp. 33--50,
A reverse process seems to me to be far more common and far more important: some feature or competence is produced in members of a species by natural selection. Then later, because the genetically-specified competence is too specific to be useful in enough different situations, the competence may be split into some general framework, provided by the genome, and context-specific details acquired by some sort of adaptation to the details of the environment. (E.g. locomotion that evolved for relatively flat terrain might be replaced by a general competence to acquire locomotion suited to the individual's environment, which may be rocky, or on a mountain slope, etc.
In some cases this could lead to an inherited group of partly similar competences being split into a number of inherited sub-competences that can be combined in different ways, using learning mechanisms to find the combinations that are useful for an individual's environment. (EXAMPLES NEEDED. REF Deacon?).
In more sophisticated cases, instead of learning (e.g. by experimentally finding out what works), a process of creative problem-solving or planning may enable individuals to work out new ways of combining fragments of old (learnt or inherited) competences. This could have the effect that parts of the genome specifying a particular combination of competences might become redundant because individuals who need that combination can synthesise it through planning or learning, when needed, and perhaps synthesise a combination better tailored to the particular environment than the previously evolved version.
In particular, running a model could not in general prove that something is
impossible. And while running a model could show that particular values
of variables are possible, more than running is required in order to partition
possible and impossible sets of configurations. Likewise running a model cannot
show that some relationship necessarily holds. E.g. exploring a variety
of particular triangles by
deforming triangular shapes cannot show that the internal angles of a (planar)
triangle will necessarily always sum to half a rotation (180 degrees). For more
on this see:
This covers a vast mixture of types of process, mechanism, form of representation, information content, and uses of information, on many scales, for many purposes.
The vast majority of successful organisms on this planet, whether measured by individual numbers, variety of species, or biomass, lack brains. Brainless organisms provide both the base of food pyramids for others, and in some cases essential forms of symbiosis (e.g. bacteria in the human gut.) Lacking brains does not stop them processing information, e.g. in controlling their reactions to their immediate environment and internal processes, including reproduction and growth.
Even in organisms that have brains there is a vast amount of sub-organism control (including homeostasis) and learning (adaptation) that does not use brain mechanisms, e.g. in metabolism, reproduction, growth, brain development, immune reactions, and many more -- but I suspect that only a tiny subset has so far been identified.
The majority of such cases, and certainly all the earliest cases, historically and developmentally, seem to rest on molecular information-processing, for example processes required for building brains, which, at least initially cannot use brains, though later on in life that can change in various ways as discussed briefly in [*].
It is commonplace to ask how the physical changes (e.g. construction of new complex molecules, or changes in the availability of oxygen) occurred.
But it is also important to ask "Where does all the information come from?" -- e.g. the information specifying complex organisms used in their reproduction.
Compare Paul Davies, The Fifth Miracle: The Search for the Origin and Meaning of Life, 1999We should not assume the information all came in one large dollop when the earth was formed, or when the universe was formed: for it is possible that the structure of matter and the space it occupies provides a platform for certain kinds of interactions and (positive and negative) feedback loops to create novel information (not in Shannon's sense, but in the sense of content that refers).
Exactly what information emerges may not be totally determined in the initial state, if some of the interactions leading to new physical structures and new types of information processing are physically unpredictable.
Another possibility is that external perturbations (e.g. asteroid impacts, changes in radiation reaching the planet, could significantly alter the environment in which already evolved organisms continue evolving -- including changing the physical properties of the environment or eliminating or reducing other relevant species, e.g. prey or predators.
The latter would be a special case of a process that happens continually, namely the environment for any species can be changed as a result of evolutionary changes in other species in that environment, including prey (food), predators, parasites, symbiants, etc.
Memes are self-reproducing information structures that move between information users with capabilities for communication, imitation, teaching, learning, and related competences.
A full discussion would need to include the transitions that led to production of information-processing machinery capable of supporting meme construction and reproduction -- very different from the mechanisms involved in encoding, copying, transmitting, using, interpreting information in genes.
Note: learning by imitation can be seen as a special case of this more general kind of learning by external provocation and detection of new affordances. (The information-processing requirements for learning by imitation are ignored by many who regard that as an explanatory category.)
Russell Foster, Professor of Circadian Neuroscience at Oxford University, is obsessed with biological clocks. He talks to Jim al-Khalili about how light controls our wellbeing from jet lag to serious mental health problems. Professor Foster explains how moved from being a poor student at school to the scientist who discovered a new way in which animals detect light...
"An important part of the "learning" required to deal with the three dimensional world of objects, processes, and other beings was done by evolution. Each child need not do this learning itself.
"Evolution solved a different problem than that of starting a baby with no a priori assumptions."
"Animal behavior, including human intelligence, evolved to survive and succeed in this complex, partially observable and very slightly controllable world. The main features of this world have existed for several billion years and should not have to be learned anew by each person or animal."
"In view of evolution, one would expect the fact of being hungry to be represented both chemically and in the language of thought."
"... we may have been lavishing too much effort on hypothetical models of the mind and not enough on analyzing the environment that the mind has been shaped to meet."
NOTE: an online PDF version of Hayek's The Sensory Order is available here:
(The 'BW/PDF' version is smaller and slightly more readable.)
Jump to CONTENTS list
Other documents introduce the general project and discuss conjectures about overlaps between mechanisms originally used (pre-historically) to produce the mathematical knowledge accumulated in Euclid's elements and mechanisms involved in non-human animal intelligence and types of discovery pre-verbal children can make ("toddler theorems"), which I think have unnoticed connections with J.J.Gibson's claim that a major function of perception is discovery of affordances. [*] Some very sketchy theoretical ideas about the nature-nurture issues related to toddler theorems are presented in this paper published in IJUC in 2007: http://tinyurl.com/BhamCosy/#tr0609 Jackie Chappell and Aaron Sloman Natural and artificial meta-configured altricial information-processing systems There's more on toddler theorems here
And many other local colleagues and students, including: Jeremy Wyatt, Achim Jung, Dean Petters, Nick Hawes, Richard Dearden, Ales Leonardis, Rustam Stolkin, John Barnden, Peter Hancox, Sebastian Zurek, Manfred Kerber, Veronica Arriola-Rios, Jon Rowe, Mark Ryan, Peter Coxhead, William Edmondson, Ela Claridge, Peter Tino, Marek Kopicki, And various members of the Theory group who have, from time to time, tolerated my wild speculations...