Hello,
I am very pleased that Stevan invited me to contribute to your
Skywriting exercise. I am sorry I have not been able to contribute
much. This is really a busy period for me (a new baby has arrived
recently, and these are also the closing weeks for our teaching
semester).
Here is a short reply to the commentary on the Harnad 2 - Symbols
paper. I hope it will help to clarify some of the misunderstandings.
1) MODELLING OF BEHAVIOUR AND LANGUAGE. The proposed modelling
approach tries to "ground" organisms' behaviour (motor and
linguistic) into the "ecology" of mushroom foraging tasks. Here,
due to the complexity of the computational approach, some behaviour
are modelled in detail (e.g. with the movement to approach a
mushroom). Other aspects are more simplified (e.g. actual exchange
of calls between neighbouring organisms, or eat/mark/return
behaviours). In related studies the latter aspects are modelled in
more detail. For example, in the simulation with Smith and Valenti
(cited in a footnote) organisms receive an input call only when there
is a close organism that produces it. The choice of the level of
detail depends on the major aims of the study. In Valenti's
simulation we were studying the effects of food scarcity and
population size, therefore it was essential to model in detail the
actual exchange and use of calls. In this study, our focus was on the
Theft/Toil distinction, and we preferred to reduce the complexity of
the availability of calls, to concentrate on the theft/toil
comparison. I believe that a more detailed modelling of the
eat/mark/behaviour and of the availability of calls
improves the model (as in Valenti's work), but it does not invalidate
the results on the adaptive advantage of theft strategy for
category learning and language evolution.
2) TOIL/THEFT DISTINCTION. In the second stage of evolution, the
distinction between a toil learning strategy and a theft learning
strategy is essential, and it's important that readers understand it
clearly.
Toil organisms have to learn the new behaviour "Return"
(i.e. activation of the output unit corresponding to this new
behaviour - in fact, this unit was never used during the previous 200
hundred generations) by having in input only the mushroom FEATURES
(i.e. ABCDE).
Theft organisms learn to activate the behavioural output unit for
Return by having in input only the CALL.
Table 1 shows the major difference between input and
backpropagation output for these two strategies. It is not
correct to say that theft organisms receive the same input as toil
organisms except for the CDE features.
3) GROUNDING TRANSFER. It is it important to say that theft organisms
actually recognise a Return mushroom (1) from the call that describes
it (as expected after the explicit backprop learning) AND ALSO (2)
when they see its features (for a study of the "grounding transfer"
phenomenon, please see also the paper Cangelosi-Greco-Harnad). The
learning of the call and behaviour Return will also ground them in
the perceptual categories.
Ciao,
angelo
-------------------- Angelo Cangelosi, PhD --------------------
Senior Lecturer, Centre for Neural and Adaptive Systems
School of Computing, University of Plymouth
9 Kirkby Place
Plymouth PL4 8AA (UK)
E-mail: angelo@soc.plym.ac.uk
WWW: http://www.tech.plym.ac.uk/soc/research/neural
tel. (office) +44 1752 232559 (fax) 232540 (home) 791214
This archive was generated by hypermail 2b30 : Tue Feb 13 2001 - 16:23:06 GMT