Re: A Touch of Thermo

Glen M. Sizemore wrote:


Behaviorism could be regarded as the view that holds that behavior can be
treated as a subject matter in its own right and not a "symptom" of
activities in the "mind" or brain. The connection with thermodynamics is,
I think, quite clear in that respect. The history of behavior analysis is
comprised largely of the discovery of relations between exposure to
particular sorts of environments (where "environment" does not simply mean
"the stimuli present," but refers mostly to the REALTIONSHIPS that prevail
between responses and consequences and between stimuli) and behavior. The
relations are discovered, generally, by directly empirical means, and
little use is made of the hypothetico-deductive method because both the
independent and dependent variables are macroscopic and easily observed
and manipulated. Again, the connection between this endeavor and
thermodynamic is quite clear.

Olea has a large repertoire concerning extensions of the notion of entropy
to the sorts of things studied by behavior analysis, but much of what he
says is somewhat beyond me. He is, however, a rather careful thinker, and
I don't doubt that he could make a contribution to the understanding of
behavior.

A few weeks ago I reread Ilya Nemenman's "Fluctuation Dissipation Theorem
and Models of Learning". Much of it is still somewhat over my head, but I
got more out of it this time around than before, largely because I had read
a little booklet: "Noise and Fluctuations", by D.K.C. MacDonald. From
MacDonald:

"If any observable system is in "passive" thermal equilibrium with its
surroundings [...] then Eqs. 12 or 16 should suffice to describe the
fluctuations. Put otherwise, from fluctuation measurements on a system
essentially in thermal equilibrium we can expect to determine a generalized
viscosity or mobility (and if we so desire, how it varies with temperature,
and with frequency or time response, which may provide valuable
information), but we *cannot* expect to discover details of the atomic
structure of our system from such measurements.
On the other hand, when a system is *not* in adequate thermal equilibrium,
generally speaking because significant energy is passing to or from the
system, we require more detailed knowledge of the system to predict the
over-all fluctuations; alternatively, we may then expect to derive from
fluctuation measurements more detailed information about te structure of
the system we are observing." [the emphasis is MacDonald's]

One of Nemenman's main points is that the transient response of an organism
to an abrupt change in a schedule of reinforcement provides information
that is not available from the steady-state response "at equilibrium". In
particular, the analogue of the "atomic structure" about which the
transient response gives added information, is not physiology per se, but
the structure of the learning algorithm implicit in response profiles.
Alternatively, it probes the structure of the hypothesis space underlying
the response. He proposes specific experiments, and derives specific
predictions. I only know of a few experiments along those lines, but the
predictions, so far, seem to bear out. I have mentioned this one a few
times:

Corrado, G.S., Sugrue, L.P., Seung, H.S. and Newsome, W.T. (2005).
Linear-nonlinear-poisson models of primate choice dynamics. J. Experimental
Analysis of Behavior 84:581-617.

Nemenman's logic runs like this: given some plausible (and testable)
assumptions about the complexity structure of ambient contingencies in
natural environments, there is an ideal, an optimal strategy (learning
algorithm/hypothesis space) that, on average, maximizes fitness - what
would Bayes Do? The behavior of organisms, by hypothesis, comes close to
realizing this ideal (and if it does not, we have probably misconstrued the
optimization problem being solved). We can, given our assumptions, predict
what Bayes would do, and compare that to what organisms actually do. All of
this is independent of any theory of the underlying physiology. But, given
a match of the behavior of an organism to the behavior of a Bayseian agent
- an algorithm - there is now a principled approach to elucidating the
physiology that mediates it. And that is what Corrado et al. did.

Ambient contingencies do not all belong to the same comlexity class, so we
should expect specializations (e.g. conditioned taste aversion), but there
is, it seems, by some plausible assumptions, a preponderance of nested
hierarchies, and so a default algorithm, out of which it should be possible
to deduce general features of habituation, Pavalivian, and Operant
conditioning. Still, "math is hard" - Barbi.

Cordially,
Michael


.

Relevant Pages

  • Re: Fundamental theorems, dilemmas, fitness, and information.
    ... Let me explain that an organism in responding to a stimulus can ... particular symphony of other impacting mutations, ... as an internalization of anything. ... is another word that suggests willful response, ...
    (sci.bio.evolution)
  • Re: 50 years later, Marvin Minsky still doesnt get it
    ... I ascribe intelligence to the organism that can create ... The behaviorist notion of "response class" is IMO a starting point ... tactic put in the same bag rats and humans. ... If there were no novel responses, ...
    (comp.ai.philosophy)
  • easter at all Elizas coffee
    ... For Casper the trail's domestic, in response to me it's precious, whereas ... responding in connection with stingy, throughout well-known, ... in relation to the organism because of the suburb. ...
    (sci.crypt)
  • Re: 50 years later, Marvin Minsky still doesnt get it
    ... most of them apparently turn out to be better at solving the problems presented to them than humans are. ... the behavior of the organism, but how this behavior was produced. ... After all, these cameras do show two aspects of intelligent behaviour: a) they accept a wide range of inputs, which vary along several dimensions; and b), their own behaviour is produced in response to these inputs. ... The notion of a response class subsumes the fact of novelty: an organism capable of conditioning to a response class will respond to novel stimuli, ie stimuli that have not been encountered before, but that are "sufficiently like" the prototype of the response class. ...
    (comp.ai.philosophy)
  • Re: Linear Phase via Phase Cloning New Method
    ... reasons why i wont just post the algorithm. ... definitions the response doesnt really apply. ... is the response of your filter? ... linear phase. ...
    (comp.dsp)