Timo Jarvilehto (1998) What is a Machine?. Psycoloquy: 9(79) Efference Knowledge (5)

Volume: 9 (next, prev) Issue: 79 (next, prev) Article: 5 (next prev first) Alternate versions: ASCII Summary
Topic:
Article:
PSYCOLOQUY (ISSN 1055-0143) is sponsored by the American Psychological Association (APA).
Psycoloquy 9(79): What is a Machine?

WHAT IS A MACHINE?
Reply to Schmid on Efference-Knowledge

Timo Jarvilehto
Department of Behavioral Sciences,
PB 222, 90571 Oulu
University of Oulu,
Finland
http://wwwedu.oulu.fi/homepage/tjarvile

tjarvile@ktk.oulu.fi

Abstract

Schmid (1998) brings forward an interesting aspect of the interplay of afferent and efferent processes from the point of view of machine learning. Equating these processes with bottom-up and top-down processing, respectively, Schmid on this basis supports the importance of efferent influences in learning. Schmid has difficulty, however, in seeing why traditional concepts of cognitive science would not suffice in the treatment of such problems and why they should be modelled by an organism-environment system. Although I do appreciate the viewpoint of machine learning -- something quite new for me in this context -- I cannot agree that afferent and efferent neural processes could be simply equated with bottom-up and top-down processes or that such a conceptual distinction would help in understanding real problems of knowledge formation or learning in machines, let alone in organisms.

Keywords

afference, artificial life, efference, epistemology, evolution, Gibson, knowledge, motor theory, movement, perception, receptors, robotics, sensation, sensorimotor systems, situatedness
1. Schmid (1998) is quite right that environmental and organismic characteristics may be manipulated independently by an experimenter and that hence they can also be described independently. It would even be odd to deny this basic mode of analysis in traditional psychology and neuroscience. However, what is at issue in my target article (Jarvilehto, 1998) is not this, but the problem whether such an analysis will ever lead to understanding the real behavior of the experimental subject. The characteristics varied in the experiment are those which the experimenter has defined in his theory. Do they exist as such also for the experimental subject? It is quite possible (and even probable) that the environmental and organismic characteristics constituting the behavior of the subject will differ, at least partly, from those specified by the experimentalist's theory. A stimulus, for example, is a stimulus for the organism only if the organism is sensitive to that kind of disturbance in the environment (see para. 41 of the target article) and if the organism has an organization into which the environmental influence may be fitted. Even if the experimental stimulus constitutes part of the perception of the subject, its meaning is not related to the experimental manipulations as such, but to the behavioral situation as a whole, and to the history and future of the subject.

2. As to the machine learning we should first ask: What is a machine? Is it only the set of elements we put together and then allow to work in some environment? When we build a machine, what exactly do we build? When we build a car, for example, we must simultaneously construct its environment: roads, bridges etc. The car without the possibility to drive from one place to another is hardly a car. Thus, the construction of the specific action environment is as much a part of building the machine as much as is the construction of its "inner" parts. Usually this is not so conspicious, because the action environment is already there before the planning is started and it is the planner (and later user) who connect the machine and its environment. Thus, a functioning machine is also an "organism-environment" system which consists of the machine, its environment and its planner/user.

3. The hypothetical environment of the machine is hence a human environment, described in words. In the construction of machines we explicitly separate machine and environment and use language to describe the parts of the machine and the parts of the environment related to it. It is the planner who contemplates what will happen with a certain set of elements in an environment that HE knows. However, as with the experimental subject so with the machine, it is there may be features of the world that are not (and perhaps cannot be) parts of the environment of its planner. This could be a reason machines break down eventually: It is impossible to take into account the whole universe in building a machine, some unknown factors are always at variance with the the construction of the machine.

4. What if the constructor of a robot, for example, does not know the exact environment in which the machine will act? Such a robot should probably be build to be able to take into account unknown environmental factors and to interact with the environment at points which are unpredictable to its constructor. It cannot be completely preprogrammed; its structure should rather be modifiable to cope with new circumstances. As indicated by the thought experiment in the target article, such an interaction would be possible even on the basis of movement and its prevention alone, but the learning process gets much more efficient with receptors and efferent influences. Thus, such a robot should be a system that can randomly (?) vary the characteristics of "receptors" to make new contacts with the environment when the original points of contact are lost or new ones appear. I do not know whether it is really possible to build such machines, but my impression is that they may be on the way (see Clark, 1997).

5. Schmid seems to think I am restricting learning to the system with efferent influences only. This is not the case; I am only suggesting the superiority of such a system in comparison with one lacking such a capability, especially in a dynamically changing environment. I define learning as a process of widening and differentiating the organism-environment system, rendering new results possible (Jarvilehto, 1994). Learning may occur even in a relatively simple system with such a capability, as in the thought experiment on the organism without receptors. (It is true that even this organism is not really without any receptors, as Schmid points out (p. 6), but this "energy receptor" is not a receptor in the sense that it cannot specify any structure for the environment.) From my definition it follows, however, that all learning is related to doing, to the development of skills. Thus, learning by doing is not a separate category among different of modes of learning, but the basic characteristic of all learning. The possibility of conceptual/perceptual learning separate from action exists only in a theory that abstracts organism from environment as two distinct systems, thereby entailing that learning can only occur within the organism (in the brain).

6. According to my definition, learning cannot be limited to the organism. Consider learning to play a piece on the piano: When the pupil learns the piece electrodes on his scalp reveal changes in brain activity (hence learning does take place in the brain). Suppose, however, that we connect the electrodes not to the scalp but to somewhere inside the piano. We will again record systematic changes during learning; in the movements of the hammers, in the succession of activated strings, in the vibration amplitudes, etc. (hence learning occurs in the piano). This problem arises only because we are localizing the results of learning to separate components of the system that changes during learning; only the whole system can produce the learned results. When one learns to write, it is arbitrary to say that one's hand learned the writing or one's brain did, because neither of them learned it: The severed brain can no more write than can the severed hand.

7. The finding of efferent influences on receptors is interesting, because it shows that the basic innervation of the muscles and senses is identical: for both we have centrifugal and centripetal nerves. We can accordingly speculate that their functions are also similar: muscles are for sensing as much as the senses are for moving. A change in the sensitivity of the receptor means a new kind of structure in the organism-environment system exactly in the same sense that its structure is changed by the contraction of the muscle. In both cases, it is the change in the availability of new parts of the environment that makes new results (e.g. percepts) possible.

8. From this point of view there is no top-down or bottom-up processing; such concepts are possible only if we ASSUME different functions for the muscles and senses a priori, thereby separating movement and perception on an anatomical basis. In organism-environment theory, perception is not a top-down or bottom-up process; neither does it consist of a sequence, such as perception->action->change of environment->perception->etc. When the subject perceives, he acts together with his environment; when the subject acts, his action is not something that belongs to his body alone; it necessarily involves also those parts of the environment which constitute his behavior. Perception leads to another perception, and action to another action, because perception and action are only different aspects of the same process.

9. Last, I must stress that the existence of efferent influences on receptors does not OBLIGE us to change the traditional basis of our theorizing about knowledge formation and the functions of senses in favor of something like the theory of the organism-environment system. Although such phenomena may be puzzling for the traditional investigator, he can deal with efferent influences with ad hoc assumptions about attention or top-down processing, just as it was possible to maintain the geocentric view by postulating epicycles. The basic difference between traditional cognitive science and organism-environment system theory, however, is that the former can handle such findings only through ad hoc explanations, whereas in the latter they are an essential factor in explaining the functioning of the system.

REFERENCES

Clark A (1997) Being there. Putting Brain, Body, and World Together Again. Cambridge, Mass.: MIT Press.

Jarvilehto T (1994) Learning as formation of man-environment system. In: Proceedings of the International Workshop on CLCE, ed. by Levonen, J.J. and Tukiainen, M.T., p. 7-8. Joensuu: Joensuu University Press. http://wwwedu.oulu.fi/homepage/tjarvile/learn.htm

Jarvilehto, T. (1998) Efferent influences on receptors in knowledge formation. PSYCOLOQUY 9(41) http://www.cogsci.soton.ac.uk/cgi/psyc/newpsy?9.41 ftp://ftp.princeton.edu/pub/harnad/Psycoloquy/1998.volume.9/psyc.98.9.41.efference-knowledge.1.jarvilehto

Schmid U (1998) Bottom-up and Top-down Processes in Learning. Commentary on Jarvilehto on Efference-Knowledge. PSYCOLOQUY 9(76) http://www.cogsci.soton.ac.uk/cgi/psyc/newpsy?9.76 ftp://ftp.princeton.edu/pub/harnad/Psycoloquy/1998.volume.9/psyc.98.9.76.efference-knowledge.4.schmid


Volume: 9 (next, prev) Issue: 79 (next, prev) Article: 5 (next prev first) Alternate versions: ASCII Summary
Topic:
Article: