Computers and Ape Language

1
Computers and Ape Language
Psych 1095 Lecture 8
2
        Gill, T.V., Rumbaugh, D.M. (1974).
Mastery of naming skills by a chimpanzee. Journal
of Human Evolution, 3, 483-493.
3
Well start with the summary piece about the
controversy over ape language
I wont go over all the details, as weve been
therebut will address some of the newer issues
We havent discussed the philosophical issue of
how achieving something like language, thus
blurring the human-nonhuman distinction
formed an undercurrent to the controversy
4
For people who discount evolution in general, the
issue of language evolution is incendiary
And, of course, if we cant really define
language, we cant really separate what humans
doincluding primitive tribesfrom what the apes
do.
The first article, written just before the huge
NYAS witch-hunt conference
depicts the issues separating the labs
5
The ASL labs on one side, arguing that
language/communication is a social act and
1) must therefore be taught in a social setting
2) and that lack of a social setting is why the
Premacks and Rumbaughs didnt get real language
3) and who cares if the data are anecdotal so
are the data for children
6
The chip/computerized labs argue that
1) ASL is cued and uncontrolled
2) ASL data are anecdotal and total corpus
material is unavailable
3) ASL work is heavily overinterpreted to support
what the researchers want to see
4) The social setting is not just irrelevant but
bad
7
The interesting bit was that, as we have seen,
Terracewho used ASLwas not supporting ASL
labsor anyone
Gardner and Fouts argue, as we saw, that Nims
training was not really social or referential
(and well see more of the problems later)
And that operant techniques gave just what was
expected.lack of transfer and no innovation
8
You can see, too, the acrimony of the
argumentsdevolving to legal, not scientific,
battle
Unless one knows sign and see the original films,
one finds the data difficult to evaluate
One of the upshots of all this was that the
Rumbaughs decided to include social behavior
But between apes, not ape and human
9
The idea was that having the apes interact would
reinforce the use of the system as a
communicative tool
But would avoid having the humans contaminate
the system with any cues
Well discuss their experiments shortly
The point of this article was to give you a
flavor for the internal warfare that was
splitting the field apart
10
And papers like that by Thompson and Church
didnt help, either
They constructed a simple computer program that
took the initial condition (e.g., food in or not
in the machine Tim present or absent)
And then spit out a relevant sentence
Of about 6500 of Lanas actual formulations, only
about 500 were not appropriately modeled
11
That is, less than 10but they still involved
most of the trained sequences
So, although Lana did put together some really
interesting lexigrams
the argument could be made that these were no
more common than the anecdotes reported by the
Gardners
Except that a computer, rather than a person,
logged them onto data sheets
12
So, as noted above, the Rumbaughs decided to look
at a different aspect of communication.
They used Sherman and Austin and instead of
dealing with the now-discredited sentence
concentrated more on the use and meanings of the
lexigrams themselves
And how the apes could exchange information via
these lexigrams
13
The first thing that the Rumbaughs did was
separate requesting from labeling
which turned out to be a LOT harder than they let
on in this paper
And actually led to a breakdown in food-symbol
correlations
Initial labeling was easyas you might expect,
the apes simply associated the symbol with the
food reward..
14
Initially, Savage-Rumbaugh tried to show the
chimpanzees that a label was not necessarily a
request by substituting a different food for the
one named
or, as in many operant paradigms, a single food
for all responses
These substitutions, not surprisingly, broke the
referentiality of the labels
And retraining was NOT as simple as this article
states
15
These problems are described in separate
papers(and her 1986 book)
Several steps were needed to get back on track
First they tried switching between providing the
labeled food and the single food reinforcer
But that didnt work at all
16
Then they tried providing a plastic replica of
the labeled food, followed by the single
reinforcer food
And that didnt work at all, either
Finally, they provided both the labeled food and
the common reinforcer food at the same time
And then slowly cut back on the portion size of
the labeled food,
Adding hugs and good job as well
17
What was ALSO critical was that they added two
gestures.one involving showing and one
involving offering
That is, physical cues..
To separate out the two situations
Just as an aside.we simply required Alex to add
want for requests
And he learned to identify stuff he didnt want
in order to get what he did
18
In any case, the training involving the 102 and
201 trials was on the last step only
So much more difficult than they imply
And other aspects were a lot easier
Lets look at the statement stating to the
animal that a sealed container held one of a
variety of foods. If the animal could decode this
statementand request the correct food, we
replied yes.
19
So what the ape saw was the equivalent of
The ape then had to request the food
And it took 5 trials for the apes to learn to do
this kind of match-to-sample
20
At the time, too, foods were likely to have had a
different color background than the locations.
Possibly the difficulty was more in getting them
to ask for things they didnt really want
so as to then get rewarded with something they
did want
In any case, decoding wasnt an issue at least
here
21
Initially, the animals hadnt been asked to work
with any sort of delay.
Now one ape was asked to wait a minute after
seeing a container baited before he could get to
the keyboard and state
Note that he didnt yet have to comment about the
container
22
The second ape now had to state
And they could shareagain, not exactly rocket
science
The first ape did have to label, but the second
only had to match-to-sample
Giving blind tests was a good idea, but the task
was still pretty simple
23
It wasnt all that much different from the
training that they had received, except for the
short delay
The Rumbaughs did make sure that the ape was not
using a position cue, and really could label
which was absolutely necessary
But didnt ensure that actual information was
transferred, only the need to match the symbol
24
The next step, involving photographs, is really,
really odd.
Here the Rumbaughs state that the apes had never
been taught to identify photos corresponding to
food names
But in the 1980 paper we also read, they describe
in detail how they trained the SAME animals to
label photographs
by taping photos to objects
25
What are we to believe????
If they could already identify via a photo the
food they had requested via the lexigram
Why did they need training 2 yrs later???
These are the kinds of inconsistencies that never
were answered
26
The next step was intriguing mostly because it
got the apes to share food willingly
Here the apes were given what looks like possibly
four food items
Tis difficult to see more than the one that is
involved in the study
27
And, of course, the apes didnt want to share
things like chocolate
So I suspect that they centered on moderately
desirable stuff the apes would share
In any case, the communication via the symbols
was not much more than one ape labeling X as x
And the other ape decoding x as X
28
The question that lots of critics asked was
whether the apes really were communicating per se
Because it didnt matter if it were an ape, a
human, or a machine doing the actions
And this paper triggered a really major response
by Skinner and his students
who did see the project as just match-to-sample
and simple association
29
There are lots of problems with this pigeon
paper, but it isnt just a parody of the ape work
Lets look at the keyboard
Jack on the left has actual colors
Jill on the right has only symbols
30
First, the symbols didnt move around so that
Jack and Jill could have learned position cues
something significantly different from the apes
Also, Jack and Jill did not trade roles, so that
they were even less likely to be communicating
than the apes
And they were trained individually
31
So, Jill was taught, after seeing a special key
hit (What color?)
to poke her head through the curtain, look at a
color, and associate the color with R, G, or Y
If correct, she was given some grain
Jack was taught to hit the What color? key,
look at a lit button (R, G, Y) and hit the
corresponding colored button
32
If he was correct, he also got grain
The birds were actually taught all this via a
backwards chain of events, but that isnt the
real issue
other than it divorced the meaning of the actions
from the actions a bit more
The experimenters then put the birds together, so
that they took the place of the machines
33
After they habituated to each other, they began
to do what they had been trained to do
Waiting for each other to act in place of the
machines
The researchers also disabled the symbol keys on
Jills side to control for extraneous cues
So what are we to make of this?
34
You also need to know that the Rumbaughs
criticized the Premacks, arguing that the
latters animals just did matching
But did Sherman and Austin really engage in
symbolic communication?
The pigeons, because the animals didnt switch
roles, did not
They just meshed independent behavior patterns
35
But did the apes do that much more?
Had they simply meshed the behavior patterns they
were taught vis-à-vis their machine
to working with one another?
And even if all that was what they did, was it
still symbolic because they had to use symbols?
Is association learning truly symbolic?
36
Lets see how the Rumbaughs replied
First, they argued that the pigeons likely simply
had positional, rather than symbol, associations.
And that was absolutely true
And although weve seen how tricky it is to
separate out even symbolic association from real
reference
the issue for the paper was not use of the
lexigrams as symbols
37
But rather whether the apes were really and truly
communicating with each other
or had learned a series of associative steps to
get what they wanted
The Rumbaughs, as we noted, argue that the
pigeons didnt exchange roles
which, again, was true we dont know if they
could learn both roles
38
But, again, we dont really know what the apes
understood
The Rumbaughs are correct when they argue that
Jack didnt care how Jills light was
activatedand vice versa.
But one has to wonder if it mattered to Sherman
and Austin, either.
What I dont understand is that they state that
the apes could NOT match symbols
39
Match-to-sample is one of the easiest tasks
possible
it would be really bizarre if Sherman and Austin
could not learn that task easily
Too, the Rumbaughs argue that Sherman and Austin
have much more sophisticated use of their symbols
Which may be true, but those data were not
presented in the published paper
40
The Rumbaughs argue that their lexigrams are much
more sophisticated than the R, G, Y keys.
which, again, is true but has little to do with
the issue of the paper we read
As for the apes understanding of gestures versus
that of the pigeons
They are basically arguing that the apes can
learn nonlinguistic cues, not so different from
natural behavior
41
The additional controls that the Rumbaughs
describee.g., touching nonsense syllableswere
important
So why werent they in the paper?
The Rumbaughs argue that the symbol enabled the
ape to recall what was hidden
But they never tried (say, 5 of the time)
touching an incorrect food symbol to see what the
apes would do
42
The Rumbaughs argue that they did not train their
apes
But if the ape did not do what it should, the
food was taken away
which is a form of training
And, again, tis amazing that they claim no
training on photographs here but do so in a later
paper.
43
Tis also true that apes have great object
permanence and that a short delay should not
affect how they respond
So they shouldnt have needed specific training
on these tasks
Pigeons might need such trainingbut thats an
issue of intelligence, not language
In sum, there are significant differences in the
two studies.
44
But one couldnt tell what all of these were
from what Rumbaughs published
And there are some intriguing anomalies
Particularly given the blistering critique that
follows in the next article concerning the sign
language studies..
What is particularly interesting is their
argument about words
45
They argue that Sherman and Austin, because they
both produced and comprehended their labels,
understood words
But, as well see, Lana did not
And the huge amount of training their animals
needed to separate labeling and requesting
suggest otherwise
They suggest that comprehension/ production
skills must be taught separately
46
I believe such is not the case and was a
consequence of their procedures
And, tellingly, they backtrack immensely over
what Lana had learned.
Now agreeing that it was simple operant
conditioning.
Not realizing that Sherman and Austins tasks
were only a bit more sophisticated
47
They are, however, absolutely correct in stating
that studies with apes led to much more detailed
investigations with children
and to a much deeper understanding of what we
mean by a word, a name, a label and a
sentence
and that these controversies were likely to harm
the field overall
but they did not back off
48
But, before we go continue with the internal
critics
We need to look at what their criticisms
engendered in scientific communities in general
The Sebeok paper is from the popular press, but
sums up the witch-hunt very nicely
First, it compares the complex signals the apes
learned to Hans hoof-tapping
49
Which we know is a bogus generalization..
Sebeoks claim is that these subtle cues
influence ALL such studies and make them outright
deceptive
What is incontrovertible is that we all emit the
kinds of subtle, nonverbal cues that Sebeok
describes.
These facilitate normal human communication
50
The issue is that there is a huge difference
between waiting until a horse stops tapping
and cuing an animal to make one of 100 possible
signsor choose one of 4 chips
Would it have been better if Sarah had been given
a box of chips and the experimenter didnt know
the placement?
Probably.
51
But apes do not follow eye gaze (work by
Tomasello) and generally the chips were so close
together that eye gaze wouldnt help
A box would have solved the argument that Sarah
was watching for the experimenter to relax as she
chose the correct chip
A serious problem is that communication is a
SOCIAL act
Take away social interaction, and what do you
have?
52
An animal that is trained operantly, knowing only
associations
Weve never been able to get the correct balance
to avoid criticisms of cuing but keep the
communication system intact
Where Sebeok goes off the deep end is to compare
the work w/ psychics
53
Sebeoks criticism of the Gardners double-blind
is also off-base
Arguing that the experimenters could guess what
was being shown.
by ESP??
Even if Washoe gave evidence, by lip-smacking,
that the slide was of food,
MANY food slides were being shown
54
Other criticisms involve how the observers might
have seen something or changed their answers.
which basically amounts to calling the Gardners
liars
As for cuing with the Rumbaughs system
Could the apes have watched for a cue as they
searched for a lexigram?
55
Possibly, but supposedly the experimenters
couldnt see the keyboard, just the final choices
And supposedly the keyboard had lots of lit
buttons, not only the 11 food buttons
But we are never told that in the journal
articles
Is there a kernel of truth in the Sebeoks
arguments?
56
Yesand had they stuck to that kernel, their
criticisms would have been constructive
And might have led to the appropriate controls.
But they went sooooo overboard that they were
dismissed
except by the large number of folks who wanted to
discount the studies
57
So lets look at the Rumbaughs paper that is
their equivalent of Premacks word paper
A sort-of response to critics
Here they are talking about categorization
Can an ape learn a hierarchical system such that
all xs are X and all ys are Y
Even if xs are keys, sponges, rakes, etc. and
ys are bananas, bread, cheese
58
Of course, the issue is edible vs nonediblea bit
simpler than other categories
And what they only briefly allude to is that, for
Sherman and Austin, the tool labels were
functional, not object labels
That is, the lexigram
was what you hit when you saw a lock it didnt
necessarily mean key
59
So that they couldnt look at a key and choose
the correct lexigram
Lana, however, was taught the objects devoid of
their use, so she had trouble figuring out that
connection
All the animals were taught this task in many
very small steps
First to sort physical foods and tools
60
Then to associate one lexigram with the foods and
another with the tools
Not exactly rocket science for an ape
Then they were taught to place the food and tool
lexigrams appropriately
Then shown novel foods and tools and asked to
signify if stuff were edible or not
Getting such a concept shouldnt be too difficult
61
Lana, however, couldnt do this.
She seemed, actually, to have exchanged the
labels, given her results(3/10)
She was tied to the labels, in a sort of mutual
exclusivity.
Its an
Its always been
Why should I label it as
62
Lanas second round is even more interesting,
when she got 1/10
Such data do seem odd
Then we are treated to the photograph work.which
contradicts the earlier paper
Again, the apes had to be trained to sort the
lexigrams into food or tool use
We dont know how long this training took
63
Eventually, they passed a test showing they could
sort novel food and tool lexigrams appropriately
Into something that represented an edible or not
The data show good comprehension of the
individual lexigrams, but not necessarily food
and tool
Red and green bins would have worked, too
64
Does the paper show some neat behavior?
Yesbut maybe not as neat as the Rumbaughs would
like us to believe..
And Lanas inability to transfer from labeling to
comprehension is telling
We see it with our parrot Griffin, who wasnt
taught properly.
And its not the issue of Grifs intelligence but
our deliberate training
65
So lets end up with the Rumbaughs criticisms of
Nim
They start by, correctly, identifying
pre-representational use of symbols from
representational use.
the early associative connections that occur even
in young children
But then, interestingly, they argue for a
controversial step
66
That of the signaler making sure that the signal
has been received
Something Sebeok would argue involves cuing
Can you have it both ways ???
They make an issue of referential pointing.
something apes dont seem to use in the wild but
do learn in captivity
67
The underlying issue is joint reference
Something that is critical for child, chimp, and
parrot, and I suspect other critters
Baldwin and others have shown that if you jointly
attend to an object with another individual
and label that object in joint attention
the individual learns the label.
68
The child, however, has to be about 15 mos old to
engage in joint attention on a routine basis
And one would expect that such behavior would
have to develop in other creatures
For younger humans, the caretaker does the work
Figuring out the focus of the childs attention
and labeling, etcand by 9 mos the child will
maintain the attention
69
So, again, why wouldnt such behavior develop in
apes?
We see it develop in our parrots
The Rumbaughs arguments that many apes find it
difficult to expand their labels
is central the issue of mutual exclusivity that
we just discussed
70
Category labels need to be added as additional,
not alternative labels
This is a key, its a green key
This is a key, its a tool to use this way
Such is true for children as well as nonhumans
As for adding labelswell.
71
A young signing child probably wouldnt do much
different from Nims you me Nim sweet drink
gimme
except maybe to eliminate the Nim
particularly if there was a choice of things to
request
The arguments about comprehension hold a bit more
water
72
We didnt see too many tests that required Washoe
to choose an object
But neither did we see it for Lana
With my birds, we found that we had to be very
careful when we did tests to see if a request was
meaningful
If you request an apple and I offer you
chocolateyou might take it!
73
And if each time you requested a banana I offered
you an apple,
youd start to conflate the symbols
So you have to put in only a few probes and
balance the desirability of the choices
And even then, you cant be sure of what is
happening apple, banana.maybe you dont care
74
I hope you all carefully read the transcript of
the training that Nim had
Because if this was representative, Im amazed
that Nim learned anything
The trainer asks him to sign to get an apple
slice he could care less
When he finally asks to eat apple, she doesnt
respond
75
Hes learned that signing doesnt get him what he
wants
She then tries to teach him the label slice
So he mimics what she does she doesnt accept
his sloppy attempt
She tried to mold his sign to improve it she
ignores his attempts
76
He responds sorry when he sees her sign things
that he knows are reprimands
Maybe no contrition, but he has some associative
understanding
Now she gives up on slice and makes him place
the apple next to the other bit of apple
He still doesnt get anything but good
77
If you were Nim, would you understand that this
is a matching task?
Now she tries to get him to label the banana
slice
He asks to eat it and she ignores him
Then he sorta signs banana and she gives him a
piece
78
The trainer does not consistently demonstrateor
labelthat she wants Nim to match.
Sometimes she lets him eat the fruit, sometimes
she doesnt.
How is he to understand what is going on??
Think about training in which there had been a
clear demonstration of the task
79
With the use of a label matchand no use of
slicewhich muddied things even more.
The Rumbaughs training is at least quite
consistent
But we really dont know if, given a choice of
foods, one of their apes would indeed choose what
he requested
And share as long as he doesnt want it himself..
80
I could not find video of Sherman and Austin
performing these tasks
I hope you also read the transcripts closely
In 7/11 of the first trials described, the apes
use pointing gestures to show what they want.
Not good statistically to show symbol and
keyboard use.
81
In the next set, 6/14 trials also either do
involve sharing or involve points
Better, but still not exactly what youd expect
for the system to be working
Note that these apes were trained on these
behavior patterns
And the symbols used here probably were not novel
ones
82
The Rumbaughs here are trying to avoid the issue
of experimenter cuing by letting the ape subjects
interact
which is a neat way of designing the experiment
when the animals cooperate
But, again, we cant be sure that all the labels
function appropriately the apes may or may not
really want what they request
Tho they do know how to respond to a lexigram
83
There are other papers that we wont read that
show that the apes can bring the appropriate tool
to one another
First the apes were taught the function of
several tools
for example, a key to open a box, a wrench to
open a pipe with a wingnut on the end
Next they were trained to request the tool from
their trainer via a lexigram
84
Next they had to give a tool when shown the
lexigram
Note that each step was trained independently and
took a lot of training
What might have happened if these steps were
trained together?
The Rumbaughs didnt figure the apes could
understand such complexity
And, by this time, why didnt the apes connect
tool and lexigram easily?
85
Now they separated the apes so they couldnt
communicate at first
While separated, one ape is shown a food being
hidden in a container
And the container requires something like a key
to open a lock or a stick to push things through
And that ape requests the needed tool from the
other ape
86
If the second ape does indeed decode the lexigram
and bring, e.g., a key
They both get to share the food
But what is actually happening in terms of their
understanding of their actions is unclear
Could one ape be replaced by a computer?
87
Im not being facetious.the issue is whether
this is real communication
And its not a simple issue to solve
In yet another study, the ape had to look at a
food in one room,
Identify it via lexigram in another room
And then comes back in and points
88
He gets the food only if the lexigram and the
point match
This would be indicative of displacement IF the
apes had done this on their own
But they needed significant amounts of training.
Even though the delay was quite short
What did they make of the task?
89
Were they confused that they had to do the
additional point?
How can we tell?
How might Nim have reacted?
How many of Nims failures were due to
incompetent training?
Or maybe the possibility that Nim wasnt a
particularly intelligent ape?
90
We havent really discussed the possibility of
individual variation in the different animal
subjects.
One would expect that animals vary just as
humans.
That most of them are averagebut that some are
Einsteins and some are pretty stupid
And researchers couldnt choose their subjects
91
Colleagues of mine at the Portland Zoo tried
signing as enrichment with one of their apes
The animal barely learned anything
If that had been the only ape to be trained, the
field would never have gotten going
No one gave an ape an intelligence test before
starting to work with it
92
I also want to talk about a somewhat less
appealing set of reasons for all the back-biting
in the area
After the 1980 election, a huge shift occurred in
the amount of funds available for research
NSF funding rates, for example, which had been
about 25 (25/100 grants funded)
dropped to about 10
93
Thus everyone was really scrambling to get a
piece of the remaining small pie
And the mentality was that clobbering others in
the field might improve ones own chances
What happened, however, was that everyone lost
funding..
Panels decided that we didnt know what we were
doing at all and other areas were more worthy
94
The Rumbaughs survived from funds from Georgia
mental health subsidies
because, as we saw in the video last time, their
work was directly applicable to help children
with disabilities
Premack decided he didnt want to deal with all
the craziness and shut down
I got small grants from a private foundation
thanks to Don Griffin
95
Another issue was the crazy publicity
The huge initial excitement was well beyond what
you could imagine
Even tho we didnt have immediate internet
access to the material
Every magazine and TV ran stories about the work
And that engendered a lot of anger and envy
96
Imagine yourself as someone who had spent 20
years doing operant conditioning on rats and
pigeons
and you might be considered the best in the area
But your face and animal were not the lead story
on TV nor was your work published in Science
routinely
Nor were conferences inviting you to keynote
97
In such an atmosphere, any mis-step at all would
not only be observed
but would be the source of a lot of gloating and,
again, major publicity
Articles like the Sebeoksones that damned the
research because the researchers could
conceivably find a way to cheat
were accepted and praised
98
Thats not to say that there werent problems.
But had folks gotten together and had conferences
to talk about how they could cooperate to FIX the
problems.
The story might have turned out differently
At one point, interestingly, Terrace was put on
the Rumbaughs payroll
99
And suddenly he starts writing articles that are
relatively complimentary
He rightly argues that the study we read about
food and tool should be expanded to include
location and drink
So that its not edible vs non-edible
But accepts Rumbaughs claims that the work is
proceeding without seeing any data (which hasnt
been highlighted anywhere)
100
Terrace argues that the Rumbaughs apes do not
demand things of each other spontaneously.
But its not clear that they ever needed to do
so.
Would have been interesting to see if they fed
only one apethe dominant one
And if the subordinate begged using a symbol
rather than a natural gesture
101
Terrace was extremely positive when discussing
the food-sharing experiment
Even tho we saw that initially the apes pointed
and didnt use the lexigrams
He is still somewhat critical that apes do not
use any symbols simply to comment on their
environment
But one wonders how hed react if they did.
102
For example, Alex may comment that his birthday
cake is yummy or that his corn is cold or
beans are green
Obviously appropriate comments.
But would these comments be considered just
chatter because they differ from childrens
comments,
which tend to direct a caretakers attention to
something new
103
And he is extremely laudatory about the work with
a pygmy ape, Kanzi
And well spend much of the next lecture
discussing how Kanzi differed from the regular
chimpanzees
Not only with respect to species differences but
also with respect to the different type of
training he received
104
with successes and attendant critics
Not too many papers have come out in the last few
years
So well stop about a decade ago
At least in terms of assigned reading
Ill briefly discuss some newer material in the
lecture