School of Computer Science THE UNIVERSITY OF BIRMINGHAM

CoSy project CogX project


Commentary on:
Can Computer Models Help Us To Understand Human Creativity?
http://nationalhumanitiescenter.org/on-the-human/2010/05/can-computer-models-help-us-to-understand-human-creativity/

By Margaret Boden



Last updated: 19 May 2010 (Re-formatted: 21 May 2016)
Installed: 19 May 2010

This was written very hastily shortly before the final deadline, in response to an invitation to comment on Margaret Boden's article.
Because it was written in such a hurry it is rather messy. I may try later to rewrite the version here.


This was previously at a now defunct web site:
http://onthehuman.org/2010/05/can-computer-models-help-us-to-understand-human-creativity/

CREATIVITY AS A RESPONSE TO COMBINATORICS

We inhabit a richly articulated environment, with potential for astronomical combinatorics.

It wasn't thus for all our forerunners, such as microbes in chemical soups, where the only variations possible that could affect individual control decisions are changes in concentrations, concentration gradients, flow rates, and illumination, and perhaps amounts of neutral matter (e.g. sand particles) in the soup. The comparative simplicity, and paucity of information in such environments made it feasible for natural selection to produce pre-compiled answers to all the possible control problems that could arise.

But many increases in complexity, in environments and in the organisms themselves, provided increased demands on the control mechanisms, which, as we'll see, tended to shift the burden of decision making from the genome to capabilities and knowledge developed or acquired by the individual.

Many different changes in the environment could increase the diversity of decision contexts for primeval soup-dwellers, including the presence of persistent structures on a relatively large scale -- for instance a large rock with different required nutrients on different parts of the surface.

For some feeders, randomly slithering or crawling around the surface and constantly grazing might suffice.

However, if different nutrients are needed at different times, and the direction towards a required nutrient cannot be detected in local chemical gradients, and the food cannot be sensed at a distance, then our microbe needs the ability to take in and store information about how places are related, so that it can choose appropriate routes to the food it needs, when the need arises, starting in different locations.

Some robots can do things like that -- moving around and building re-usable maps of terrain or buildings (i.e. using SLAM mechanisms: for Simultaneous Localisation And Mapping),

If the food is also mobile, and it attempts to escape being eaten by detecting an approaching grazer and moving away, or if there are competing grazers, then decisions about direction and speed of motion towards food become more complex. For instance if several competitors are already close to a source of the currently required nutrient then it may be better for a grazer to head for a more distant source of the required food, where there are no competitors.

Of course, whereas previously the grazers could use local sensors and SLAM capabilities, they now need to be able sense what is happening at a distance and use that information when deciding which food source to head for. The greater the distance at which they can perceive things the more varied the combinations of information items on which to base decisions.

If the food evolves a defensive shell so that a grazer can no longer simply approach and consume, but also needs to remove the shell, that would favour the evolution of beaks, teeth or claws that are capable of holding and breaking open the food.

But now a hungry creature has more complex control decisions than merely selecting directions and speeds of movement to get to immediately consumable food: it also needs to control the motions of beak, teeth or claws in relation to individual food items. That can involve selecting a direction of approach if there are intervening hard obstacles and more detailed control of motion if the shell has recognisable weak points that need to be attacked. (Try searching the web for videos of parrots eating walnuts.)

Such weapons that provide manipulation capabilities can also acquire other uses, e.g. damaging competitors for the same food, or removing obstacles that prevent access to the food.

In many situations, this will increase the variety of options for action available to individuals and the variety of sequences in which actions are performed. E.g. in which order should one (a) move towards food, (b) look for competitors for the food, (c) move toward competitors, (d) threaten or attack competitors, (e) start eating food, (f) decide whether to retreat, .... and more.

Evolution has at least the options of either producing innate control systems that produce behaviours that successfully achieve consumption of food, or developing some sort of learning mechanism that enables the best control decisions to be associated with various sensed conditions to be learnt by each individual. The former is feasible if conditions remain static for a long time, so that relatively fixed (though not necessarily simple) designs can be evolved. If the world changes too fast, evolution will not be able to catch up, and species whose individual members can learn will have an advantage.

There are many more scenarios that can be imagined, which are not necessarily entirely fanciful -- in view of the enormous diversity of life forms we find on earth. Some scenarios involve the food evolving to live on dry land out of reach of its aquatic consumers. If the aquatic food source runs out then evolution could favour variants of the grazers that develop the ability to move onto dry land to meet their needs.

If the food organisms then develop stalks supporting the edible portions out of reach of the grazers, the result could be evolution of jumping, or climbing abilities, or long necks, or long limbs supporting the manipulators, or abilities to build climbing supports or to throw things to bring down food.

All changes in morphology of the food-seekers are worthless without corresponding changes in control mechanisms -- requiring more and more varied kinds of information in more and more complex situations to be used in acquiring food.

Some of the foods may develop camouflage, by evolving towards an external appearance that resembles that of other inedible objects. In that case the feeders may have to develop more sophisticated perceptual abilities, perhaps taking account not only of the appearance of the edible items but also their location and perhaps how they respond to various tests, such as the sound produced when they are tapped with something hard. Alternatively, instead of directly evolving those abilities, evolution can produce mechanisms that learn in an individual's life time how to do it.

However, if individuals are not born or hatched with all the competences they will need, then they will take some time to learn, during which they require help from adults -- protection, feeding, and opportunities to learn. This will add to the information processing demands on the adults, who then need to detect and cope with not only the environmental situations that are relevant to potential harm and benefit to themselves but also those that can affect their offspring. They need to be able to detect vicarious affordances. This adds to the learning requirements of the offspring, if they have to learn to look after their offspring.

The point of all this is that as the environment poses new challenges, evolution can produce responses that involve various combinations of change of behaviour using the original morphology, and change of morphology allowing new behaviours and requiring new control mechanisms, which can indirectly lead to yet more changes requiring even more complex behaviours and more complex control systems.

The idea of an "evolutionary arms race" is not new. My examples are presented only in order to emphasise the increasing demands on control mechanisms and the need for increasingly sophisticated information-processing systems involved in such control.

These are not easy for biologists to study because internal information processing mechanisms and behaviours and not accessible to normal methods of observation and measurement. They also do not leave fossil records.

There is much more that can be said about the changing requirements for (a) forms of representation in which information is acquired, stored manipulated, interpreted and used; (b) the changing ontologies required as more complex types of information about states and processes in the environment are acquired and used; (c) the changing architectures in which more and more varied information-processing competences are combined, a process that itself requires more control decisions e.g. about which competences to activate and how to deal with conflicting control decisions generated internally; (d) the changing trade-offs between evolutionary changes in the genome and the processes of learning and development, partly under the control of the environment, in individuals.

I return to my starting observation: We inhabit a richly articulated environment, with potential for astronomical combinatorics. When the wind blows there are many possible combinations of strength, direction, temperature, humidity, whether the motion is linear or curved (as in tornadoes), and many different kinds of airborne objects possible, e.g. sand, leaves, twigs, discarded rubbish, smoke, pollen, fumes from chemical plants, etc.

When you stand in a forest surrounded by trees, bushes and other objects there are many possible directions in which you can move which will vary your relationships to other things in significantly different ways: bumping into a tree-trunk, bumping into a blade of grass and bumping into a wasp nest will have very different effects.

Even if you remain still and do nothing, other people, other animals and things moved by wind or water can produce changing relationships that you may need to take into account.

For an animal with mobile manipulators there are even more possibilities that can be combined in different ways with the possible changes and movements in the environment.

Some numbers

If there are a hundred different things in your environment each of which can change or remain as it is, then the possible number of distinct combinations is 1267650600228229401496703205376. (i.e. 2 to the power 100).

If your learning procedures allow you to experience each of those possible combinations, and you try one every second, then the time required would be approximately 40196900000000000000000 years (dividing the number of combinations by the number of seconds in a year).

If each of the changeable items has more than two options the numbers will be even larger. If we consider not just instantaneous changes, but sequences of changes over a period of a few seconds the numbers of possible sequences of changes will be even larger.

An implication of all this is that neither evolutionary time scales, nor individual learning times available in the life of an individual can cope with exhaustive learning.

Some reduction in the number of learning episodes required can be achieved by 'chunking', i.e. grouping together things that do not differ very much. But even that will not tame the combinatorics, as has long been evident in the case of language (as Chomsky, among others, pointed out).

I am completely certain that you have never previously encountered the sequence of words in the paragraph you are now reading, unless you deliberately reread it. But that novelty does not stop you understanding what I am saying. You have ways of WORKING OUT the meaning, on the basis of your understanding of individual words and phrases and the ways they have been put together.

Similar things can be said about many actions a child performs: some of them are trial and error behaviours, or copying behaviours, but many are novel solutions to problems (novel as far as the child is concerned) where the child is able to work out what will happen by combining previously acquired knowledge and reasoning about it -- for example when a child works out for the first time that the edge of a rigid circular object can be used as a screwdriver, or who invents a new practical joke to play on a sibling, or who first devises an efficient strategy for nesting cups, after finding that randomly choosing a 'next' cup leads to frequent wasteful back-tracking.

One such strategy is to start with the largest cup, then always seek the largest remaining cup to insert next. You probably worked that out for yourself long ago, without even being aware that you had done so.

All this is a brief introduction to the study of the many ways in which biological evolution was under pressure to provide humans and other animals with information-processing mechanisms that are capable of acquiring many different kinds of information and then developing novel ways of using that information to solve any of millions of different problems without having learn solutions by trial and error, without having to be taught, and without having to imitate behaviour of others. I.e. they are P-creative solutions.

I conjecture that these highly practical forms of creativity, which are obviously important in crafts, engineering, science, and many everyday activities at home or at work, are closely related to the mechanisms that also produce artistic forms of creativity. But for that we have to go into the complex topic of where motives come from, and how alternatives are evaluated, about which I have said nothing so far.

One of the problems of AI researchers is that too often they start off with an inadequate understanding of the problems and believe that solutions are only a few years away. We need an educational system that not only teaches techniques and solutions, but also an understanding of problems and their difficulty -- which can come from a broader multi-disciplinary education. That could speed up progress. It might even be a creative solution to an educational problem.

Of course, none of this will impress people who don't WANT to believe that machines can be creative. They just need to learn to think more creatively.

I have more on these topics in presentations here.


Note added after the above had been submitted.

I forgot to make it clear, though I hope it was obvious, that my response is merely an elaboration of some of the points Maggie had already made -- emphasising the importance of studying cognition and creativity not merely in the context of humans and future machines, but also in other products of biological evolution, taking full account of the features of the environment that helped to define the problems evolution solved, some of them rather unobvious problems.


Maggie's reply

Maggie replied
"Yes, P-creativity (as well as learning) is going to be needed if the environmental/behavioural combinatorics are astronomical. As for creativity in arts vs. science, it seems to me that the same broad principles of creativity are at work in both cases, with respect to the **origination** of new ideas-although of course there will be detailed differences from cases to case. But the **evaluation** pf the new idea is very different indeed in the two cases. E.g. an artwork doesn't necessarily have to 'match' the external world in any detail (although many do: think of fifteenth.sixteenth century Dutch interiors and still lifes)."
28 May 2010: My (partial) response
I think saying that the evaluation is "very different indeed"
focuses on the differences, which can be very great. But there are
also similarities in the evaluation processes, especially in the
kind of creativity in coping with the environment that also provides
the basis of mathematical competences and creativity, which I have
recently been discussing in presentations and papers, e.g.
    http://www.cs.bham.ac.uk/research/projects/cogaff/talks/#talk29
    If learning maths requires a teacher, where did the first teachers come from?
    Why (and how) did biological evolution produce mathematicians?
and the accompanying workshop paper:
    http://www.cs.bham.ac.uk/research/projects/cogaff/10.html#1001

In particular, I believe that our ability to do mathematics, which
drives and is partly driven by empirical science, is based on a kind
of productive laziness which requires perception of powerful
re-usable patterns. I think there are deep connections with music
and poetry.

I shall add some more detail later, distinguishing different aspects
of creativity in

    a. identifying a new problem

    b. working towards finding a solution

    c. finding a good way to present the solution (e.g. to make it
       clear, compelling, as easy as possible to understand)

    d. presenting the solution to others in a social interaction

    e. being part of the audience for such a presentation whether
       live or on paper

    f. teaching others about a proof, theory, technique, machine,
       building, poem, composition, painting, style, ... even if you
       are not the originator.

In each of those processes there are aspects of P-creativity that
involve evaluation, but the evaluations are of different kinds. In
each case, if we compare mathematics, engineering and science, with
music, poetry and painting, there will be both deep similarities and
deep differences.

There are interesting overlaps when the activities come together in
a single new bridging creation -- the engineer who finds a new way
of making flutes, violins, pianos, or saxophones, for example,

More later.

Maintained by Aaron Sloman
School of Computer Science
The University of Birmingham