In this commentary, the thought experiment discussed by Jarvilehto (1998) is reviewed and revised. The methodology of the thought experiment that investigates the efferent influences on receptors in knowledge formation actually investigates the opposite. Moreover, it is invalid, in that it uses conditions that are impossible in principle. I will discuss the methodology of the thought experiment and propose another approach.
1. In his target article, Jarvilehto (1998) argues that knowledge formation cannot arise from sensory input only and that the influence of efferent signals cannot be ignored. Jarvilehto discusses two experiments with rabbits, followed by a thought experiment. My comments will mainly concern his thought experiment.
2. I agree with Jarvilehto that efferent signals can contribute considerably to knowledge formation, but they are not a necessary condition, as has been shown, for example, by Steels (1996b). I will discuss how the method Jarvilehto introduces in his thought experiment is fundamentally wrong in its approach, and how the use of the term knowledge in the target article is questionable.
3. In the following section I will comment on the thought experiment, arguing that it is wrong. I will also propose an alternative approach to investigating the efferent influences on receptors in knowledge formation.
4. In the thought experiment, Jarvilehto describes an organism that is able to move in an environment under four conditions. The hypothetical organism has only two types of effectors. so it can move in two different directions. In the first two conditions, the organism (or "agent," as I will call it) has no receptors whatsoever. The agent's environment is a cube closed by walls and filled with a homogeneous energy-field.
5. In the first part of the thought experiment the agent has only two types of effectors, each with a certain amount of energy; it has no sensory equipment. The agent can use only one of the effectors at a time for going in a certain direction until the energy level of the effector has been used up, whether or not the agent hits a wall. When the energy of this effector has been depleted, the other effector takes over and the agent will move in another direction until the energy level of this effector has been depleted. While one of the effectors is used, the energy level of the other is restored. I am not sure whether this first type of organism is able to accomplish the task without any afferent connections to detect its own energy levels, but the organism from the second part of the experiment certainly cannot accomplish the task it is supposed to perform.
6. In the second part of the thought experiment, the agent is somehow able to sense the walls of the cube without sensors! The agent adapts its interneuronal connections in such a way that it optimizes the energy so it can turn before it hits the wall. Jarvilehto concludes (par. 29):
"Thus, it seems that the organism comes to know the structure of
its environment in the sense that it can anticipate the walls and
the instant of hitting them."
Jarvilehto informs us (par. 31), correctly, that such an agent must have some kind of a sensor (e.g., to detect the presence of a wall from a change in the energy flow of the effector). But for the sake of argument, Jarvilehto has us suppose that this agent does not have sensors. Unless Jarvilehto presupposes other conditions, which have not been specified or clarified, this makes the whole argument invalid. What we have is an agent that is impossible in principle. Jarvilehto argues that this agent somehow has knowledge of its environment, while it only has efferent pathways. But the agent does have afferent pathways, otherwise it cannot exist; so there is a contradiction, which makes the argument invalid.
7. Another thing that struck me was the use of the term "knowledge." Although Jarvilehto makes a comment about this (par. 39), I wonder whether such a simple agent does have knowledge. One can easily build an organism on two wheels that drives towards a light source by simply connecting a left light sensor to a right motor and a right light sensor to a left motor, all with excitatory connections (cf. Braitenberg, 1984). The question we should ask is: does this agent have knowledge about how to drive towards a light source? Does an ant that follows a trail of pheromone know what it is doing? I think this is reactive behavior and that the term knowledge should be reserved for a higher cognitive level.
8. In paragraph 30, Jarvilehto makes another mistake: that this organism could be simulated in a computer environment. This is a misinterpretation of the cited article by Tani and Nolfi (1997), who do not mention a simulated agent without afferent pathways; their agent does have afferent pathways. A simulation of of the agent in the first part of the thought experiment may be easy to implement, but the hypothetical agent from the second part, like Tani and Nolfi's physical agent, cannot be implemented without a sensory apparatus.
9. In the third and fourth part of his thought experiment, Jarvilehto proposes an organism that is possible. Such an organism already has artificial look-alikes: Steels {1996) shows us how robotic agents can learn dynamically how to survive in a (closed) ecosystem with a charging station and several competitors by driving to the charging station (or energy spot) and attacking the competitors.
10. Although the last two parts of the thought experiment are possible, I have problems with the hypothetical methodology Jarvilehto uses to show the influence of efferent pathways on knowledge formation. I think there is a better way to show this. What Jarvilehto actually illustrates (if we assume that his thought experiment would be valid) is the influence of afferent pathways. He first looks at an organism with only efferent pathways, and then considers what happens when the organism also gets afferent pathways. So we see an agent that is better in knowledge formation if it has afferent pathways. Hence we see the afferent influence of knowledge formation, not the efferent influence.
11. In order to investigate the efferent influence on knowledge formation, one would do better to start with an organism that only has receptors, rather than effectors, and then investigate what happens when it suddenly has effectors to use. Agents with only receptors are capable of knowledge formation (Steels, 1996b; Vogt, 1998a). In these experiments agents create knowledge (or meaning) from perceptual information only. This knowledge can then be used in, say, lexicon formation. Vogt (1998b) has shown that two agents have difficulties in developing a coherent lexicon from perceptual information alone. When the lexicon is instead formed while performing a coordinated task such as the hearer's providing feedback about the topic through pointing, then the agents do quite well in developing a coherent lexicon. We accordingly can see that the influence of pointing (i.e., a task using efferent pathways) increases the usefulness of lexicon (or knowledge) formation.
12. I understand why Jarvilehto introduced his thought experiment in connection with results from the experiments on rabbits. In those experiments a rabbit first learns to gain food by pushing a pedal. Later, when the eyes of the rabbit are covered, it learns the same task. The optic nerve is activated when the rabbit pushes the pedal in the same way as when its eyes are open. But as Schmid (1998) rightly notes in her commentary on Jarvilehto's target article, the rabbit could have correlated the control of the motor signals with the optic signals beforehand. A better approach would be to have a born blind rabbit learn the task and see whether this will also activate the optic nerve. Yet another approach would be to use a rabbit whose vision is functional but who first learns the task with eyes covered; the activation of the optic nerve is then monitored.
13. I agree with Jarvilehto that efferent information is important for knowledge formation, but it is not necessary. The methodology of his thought experiment, however, is rather strange and has errors. It would be very interesting to see thorough experiments that can show what the influences are of efferent knowledge formation.
I would like to thank Edwin de Jong for proofreading this commentary and for some useful suggestions.
Braitenberg, V. (1984) Vehicles, Experiments in Synthetic Psychology. The MIT Press, Cambridge MA.
Jarvilehto, T. (1998d) 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/psyc-bin/newpsy?9.76 ftp://ftp.princeton.edu/pub/harnad/Psycoloquy/1998.volume.9/psyc.98.9.76.efference-knowledge.4.schmid
Steels, L. (1996a) Discovering the competitors. In: Journal of Adaptive Behavior 4(2). 1996.
Steels, L. (1996b) Perceptually grounded meaning creation. In: Tokoro, M (ed.) (1996) Proceedings of the International Conference on Multi-Agent Systems. AAAI Press, Menlo Park Ca. pp 338-344
Tani, J. and S. Nolfi (1997) Self-organization of modules and their hierarchy in robot learning problems: A dynamical systems approach. Sony Computer Science Laboratory Inc., Technical Report SCSL-TR-97-008.
Vogt, P. (1998a) Perceptual grounding in robots, lecture notes on artificial intelligence. In Learning Robots, Proceedings of the EWLR-6, Lecture Notes on Artificial Intelligence 1545, eds. Birk, A. and Demiris, J., Springer. 1998. http://arti.vub.ac.be/~paul/ground.ps.gz
Vogt, P. (1998b) The evolution of a lexicon and meaning in robotic agents through self-organization. To appear in: Chris Knight and J.R. Hurford (eds.) The Evolution of Language (selected papers from the 2nd International Conference on the Evolution of Language, London, April 6-9 1998). Forthcoming. http://arti.vub.ac.be/~paul/evol.ps.gz