Newsgroups: comp.ai.philosophy
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!news.mathworks.com!zombie.ncsc.mil!news.duke.edu!news-feed-1.peachnet.edu!insosf1.infonet.net!internet.spss.com!markrose
From: markrose@spss.com (Mark Rosenfelder)
Subject: Re: What's innate? (Was Re: Artificial Neural Networks and Cognition
Message-ID: <D2xozo.CKH@spss.com>
Sender: news@spss.com
Organization: SPSS Inc
References: <3fosrd$2if@mp.cs.niu.edu> <3fq5ih$hkb@mp.cs.niu.edu> <D2vEL5.9u8@gpu.utcc.utoronto.ca> <3g169u$s8o@mp.cs.niu.edu>
Date: Tue, 24 Jan 1995 23:52:35 GMT
Lines: 86

In article <3g169u$s8o@mp.cs.niu.edu>, Neil Rickert <rickert@cs.niu.edu> wrote:
>I don't think enough is known about the details of prewiring of the
>brain to definitively settle this.  But let me put this into
>perspective by commenting on how I see evolution.  As I see it,
>evolution is an incremental engineer.  It operates by making small
>changes.  [...] Small genetic changes
>can proliferate existing structures, and/or use them for new
>purposes.  A wholly new structure, however, would require more
>substantive genetic changes [...]

>Our adaptation for language required some changes in throat
>structure, and motor neurons for the musculature of the vocal chord.
>It most likely required new higher level structures in the brain.
>But the new musculature and neural tissue could be very similar to
>existing such structures dealing with other parts of the anatomy.  If
>that were the case, these might result from relatively simple changes
>to the genotype.  But if the language specific neural tissue is a
>near copy of other neural tissue, then language learning must uses
>processes similar to those used for learning in other domains.  That
>is, language acquisition must be part of a general learning facility.

The last sentence doesn't follow.  We don't know that there *is* a "general
learning facility"; there might be a number of specialized learning 
systems.  And it could be that a language facility developed out of
some existing *specialized* subsystem (one dealing with primate calls,
for instance).

>Chomsky explicitly denies that language acquisition can be part of a
>general learning facility.  Thus he must be arguing for a completely
>new structure built on completely new principles.  

Again, this doesn't follow.  It might be a new structure built on 
old principles; it might be a re-use of an old structure; it might be
a relatively simple new system built on top of and exploiting various
existing structures.

>The argument for UG is based on the poverty of stimulus.  It is
>argued that a child, using induction, testing, etc, cannot learn the
>correct grammar.  Therefore the grammatic ability must be innate.

The poverty of stimulus argument is not the only basis for UG.  Other
important grounds include

a. Children seem to avoid whole classes of mistakes; one would rather
expect to find more *kinds* of mistakes if they started out with no clue
as to the structure of language
b. Language (unlike say cooking or hunting) has been linked to particular
areas of the brain
c. All natural languages share some grammatical properties; a fact easily
explained with UG, but requiring some other explanation without it
d. Children of almost all mental levels can learn language, even if 
other cognitive abilities are severely impaired
e. Unlike other cultural complexes such as technology or religion, languages 
are found in every human culture in roughly the same level of complexity
f. The ability to learn a language fluently virtually disappears around
puberty, although general learning skills remain and are even enhanced.
The picture of language learning *after* puberty is more in line with what
one might expect with general learning abilities: highly variable by 
individual.  Apart from the disputed case of first-language acquisition,
there is proof that the vast majority of humans *can* learn a language 
fluently using general learning skills
g. Simple languages are elaborated into complex ones naturally by children
who grow up speaking (or signing) them; a thoroughly unnecessary refinement
if children were merely learning their parents' language.

As for the poverty of stimulus argument itself, it's worth pointing out that
the perceived problem is not that the children don't hear enough language
(which is why Colby's 20,000 hours of language experience is a red herring),
but with what is *not* heard: explicit negative feedback.  (Parents sometimes
correct children grammatically, but some do not, and those children learn 
language just as well, and sometimes better.)  As ANN researchers have found,
it's more difficult to learn a task with only (or largely) positive feedback.

>But the poverty of stimulus argument can be turned upside down.  If
>the child, with a process of trial and error testing, cannot find the
>grammar, then evolution, also using trial and error testing, cannot
>find it either.  Given the slowness of evolution, and the relatively
>short time in which homo sapiens evolved from other apes, the number
>of trial-and-error tests that could have been performed by evolution
>is smaller than the number of trial-and-error tests that a child can
>perform during the period of language acquisition.

The tasks are not comparable.  Consider the class of grammars as powerful
as that of a natural language.  To oversimplify drastically, the child 
has to determine which of these possible grammars is actually in use;
evolution need only pick one randomly.
