Sea Lions and Parrots: Smaller Brains, Equivalent

Abilities

Psych 1090Lecture 15

A few years after Herman began publishing his work on the

dolphins, Schusterman decided to see what

a smaller-brained marine mammal might do….a sea lion

Schusterman was also trained in the behaviorist tradition

And, to make sure that his techniques would match those of

Herman,

he went out and actually studied with Herman for awhile

And decided to use the gesture language because it required less

hardware

He trained a sea lion, Rocky, to do most of the same things that Ake and Phoenix had learned to

dowith basically the same grammar

to act on various items— “frisbee over” meant jump over the

frisbeeAnd also taught some relationals

So, like the dolphins, Rocky could understand when to take the

frisbee to the surfboard

and when to do the reverse action

Unlike Herman, Schusterman also taught Rocky relational terms…

Bigger/smaller, lighter/darker

Relational terms are really important because they are totally context dependent

As we talked about in earlier lectures, what is bigger on trial 1

can be smaller on trial 2

so the subject cannot answer with respect to absolutes;

the subject has to understand the concept that is being tested

And, of course, the subject has to be able to transfer to new

exemplars without any training

And Schusterman and Herman were different in other ways…

They got into quite a battle about ‘grammar’…

Herman insisted that what his dolphins were doing was

grammar…In fact, one of the reasons he

started the research was to show that a nonhuman could learn

grammar

But Schusterman argued—correctly—that what had been

learned

was rule-governed behavior…

The first step toward grammar, but nothing quite as complex as

what humans use

As an aside, birds engage in such rule-governed behavior in their

complex songs

and any animal, bird or mammal, and like early humans,

that combines a vocalization with a gesture in a standardized

manner

is engaging in a primitive grammar:

A standardized rule-governed behavior that enables efficient

transmission of information

They then continued to “duke” it out, arguing over what is or is not

grammar in their subjects,

Using “anomalous” commands

Now, we talked about the anomalous commands for

dolphins awhile backand mentioned that Lana would

‘erase’ anomalous sentences from her computer

and we saw that Kanzi would carry out semantically stupid

actions that were grammatical…

Only Schusterman tried to make sense of all this

in terms of conditional discriminations and something

called equivalence class formation

The latter is, according to some researchers, the basis for human

linguistic competence

Now, we’ll get back to equivalence when we discuss the

results of Schusterman’s experiments

But now is probably a good time to introduce the concepts BRIEFLY

and why it is thought to be important

And, as usual, a source of controversy!

In some ways, equivalence is just a special form of transitive

inference

Only instead of using, e.g., “bigger/smaller”, it uses

“equal to”

So, if A=B, B=C, A?C

Which suggests that any animal that can do TI should very easily

be able to understand equivalence

Of course, as with TI, a lot has to do with how the subject is trained

or exposed to the stimuli

And, of course, there are fancy terms for the various

relationships

We’ll get to all that later…for now all we need is to

understand what Schusterman means by “functional

equivalence”…Which is basically that Rocky understood that some signs related to objects, some to

actions and some to attributes

In human language terms, nouns, verbs, and adjectives

Note that Schusterman is looking at a very simple form of

equivalence here…

Just hierarchical categories of how the signs relate to the sea lion’s

world

And, an aside…at one point, Schusterman used two-handed actions for one type of sign and

one-handed actions for another…

I’m not sure that he kept to this, but at least in the beginning it would have been an important

cuesimplifying what the animal had

to learn…

So, in the paper on anomalous sentences, the issue was how

Rocky would respond

If the sign from one functional class was placed in the spot

normally reserved for another functional class

or if something was missing in the sentences she was given

Basically, she would balk if given something that didn’t make

sensethough sometimes she’d do what

she could with what she was given

Suggesting that she understood enough to know that these

weren’t right

Interestingly, she balked appropriately if, for example, an

object was missingbut did what she could when a

modifier was missing

Of course, she had learned to work with objects before they had been

modified…

which may have mattered…

Also is not clear if her dropping first anomalous modifier had to

do more with memory…

But anyway you look at the data, Rocky had some idea about what was going on…

She never performed randomly on nonlabeled items

But whether she had learned a real syntax was still not clear…

Was she balking when an action was missing because of syntax

or just because she had been trained to station until she saw

a verb?

In terms of conflicting modifiers, Rocky did about 3/4 with respect

to recency…

Given that only a fraction of trials were reported, this was probably

statistically relevant…

But not as strong as argued in the discussion

And, of course, as noted by Schusterman, all that Rocky did

was to enable her to get the most fish per session

So, were these actions that she performed meaningful in any real

sense?

Or just actions to be performed in particular ways for food?

Such questions can, of course, never really be answered….

But they were asked by everyone who argued against the results of these ‘animal language’ studies

And, of course, the arguments between the researchers didn’t

help

Now, such arguments were, in one sense, very good and

importantBecause a field can’t grow and we

can’t learn from our mistakes if the mistakes aren’t found

The issue is whether there should be a chance to fix the

mistakes and continue

I didn’t choose to assign the papers where Herman and

Schusterman go after each other

They are referenced in the papers that have been assigned for those

who are interested

I’ll just briefly go over the bit we did read…

In many ways, Rocky did not differ all that much from the

dolphinsAlthough Schusterman does

see differences….

Primarily with respect to reinforcement issues…

As stated, dolphins were ‘paid’ to guess and the sea

lion was notwhich again brings into focus how different types of training

regimes result in different outcomes

And the importance of experimental design

While I don’t see either set of experiments as really telling us all that much about the actual

ways in which the animals process information…

I do see that there isn’t too much basis for a pure linguistic

interpretation

Just like the starling-recursion discussion …

IF you argue that failure on a task demonstrates lack of linguistic

knowledge

Then you have to accept that success does demonstrate

linguistic ability…

But the real issue is whether the task really was the crucial

one…Note, the issue is not whether the CONCEPT was the crucial one…

It may still be….

But the task chosen to represent the concept

AND exactly how the task is taught to the subjects

e.g., the tamarins were exposed to the task and the starlings

taught…And these are the same issues here for the dolphins and sea

lions

And this brings us back to the equivalence issues….

And whether one can actually analyze language from the standpoint of equivalence

Or if equivalence is one of several issues

So, let’s talk about what all the terms in equivalence

mean

and how they might relate to language….

for humans and nonhumans alike

So, we start with reflexivity, A=A

Which is just another fancy term for identity match-to-sample

And, we have seen, most animals are pretty good a learning that

and transferring it to totally novel stimuli

Although for some animals what is shown during transfer is just a savings in trials needed to ‘get it’

compared to the number of trials needed to learn the original task

But, basically, it isn’t rocket science

The next part is symmetry and that is a bit more difficult

Because not only is it a conditional relationship…

that is, “if A do B”, but also “if B do A”

Many nonhumans have trouble with symmetry

But nonhumans can indeed be trained to respond to

symmetryIt doesn’t however simply

emerge from their initial if A then B knowledge

And it really isn’t entirely clear that it emerges for children

either

Although various studies suggest that it emerges much more easily for children than

for nonhumans

And, of course symmetry is what underlies a lot of what language is

about….

Because comprehension-production require symmetry….

And, as we saw, sometimes animals taught only production couldn’t comprehend and vice-

versa

The animals’ abilities may have had more to do with the way

they were trained than with their innate abilities

But you can see why the issue of equivalence is so critical to

languageA subtle point that is not

addressed except by Premack, however,

Is whether even knowing that A and B are related by symmetry

means that A is a symbol representing B

You could see how just accepting that A and B are interchangeable doesn’t mean they are symbolic

representations

Think of it this way…I can give you a token for a cup of coffee, or

vice versa

But the token doesn’t represent that cup of coffee

Unless I know otherwise, the token can instead represent $1

And that’s one of the big issues in the idea of training many

equivalences….

Because many things can, for example, cost $1, but the $1

doesn’t really represent them…

Except in an odd, basic sense

But let’s go on to transitivity

that is the equivalence form of TI…

So that the subject, trained on A=B and B=C can, without

training, understand that A=C

But, again, are we talking about real symbolization or not?

We can see how it might relate to language…

in the sense of “key” (the vocal label) = KEY (the written label) =

= “key” , etc

But, as with the $1 example, you can see how the issue becomes

clouded…

Now, some researchers argue that equivalence is a prerequisite

for language

Others argue that language is a prerequisite for equivalence…

If you could show equivalence in a nonlanguage trained

animalOr only in a language-trained

animal

You might answer the question one way or the other…

But it is never easy

Now, initial studies on various animals suggested that they

couldn’t understand equivalence relationships

If they had been trained, that meant that equivalence didn’t

emerge spontaneously

Was it their abilities or the training?

Schusterman argues that training is ok…

As long as the animal can transfer to novel situations

Which makes sense, because the animal has to learn what it

is that we want it to learn in the first place

Of interest in this paper was that the previously language-trained

animalwas the one that had the most trouble and had to be dropped

Suggesting some kind of dissociation between

language and equivalence…

Or between what was previously taught and language….

Quite possibly, Rocky didn’t understand why there was a

major switch

Hand signals used to be the command and now pictures

were being used

And for Rocky, the hand signals connected to objects and

actionsAnd now all that were available

were the pictures

The set up was what we saw in the videos with Rio on Sci Am

A B

If Rio was given A, she had to pick B

All As were familiar so she had to choose novelty

B A

Now she was given B and expected to choose A

If not, she was given training to do so

Now, they used arbitrary symbols and not letters or

numbers

And the sets differed on different trials

So A could be a circle and B could be a plus sign

The point was that no particular sample was right or

wrong,But depended on what was

being trained

So she learned 6 sets of two items so as to understand the

pairings between them

But for 24 sets, she was not trained on the symmetry

relation

So there were a bunch of pairings for which she knew

only to choose novelty

Now she was given some additional training involving some symbols she had not

before seen….

B C

And, again, at first she had to choose novelty

But then she was trained if she saw the relatively novel

itemShe had to choose its pair

And she was trained on 6 such pairs

Then there was training for A-C and C-A

With review training on the A-B and B-A sets

But only for the samples that had been set aside for training

She could then be tested on the sets that had not be trained for equivalence

To see if she had somehow understood what should go in

which group….

Remember, she had been trained on all objects in one direction

So what was being trained was whether she understood the relationships with respect to

their categories

It is as tho’ I trained you

Q 5

V 9

For lots and lots of sets like that

And then I also trained you on

Q

And then tested you on

Q V

Could you link the correct objects?

Obviously, Rio had a large number of these types of trials

And lots of reminders of what was connected to

what, at least in one direction

And, when tested, she did quite well…

The assumption was that she had learned the

concepts of symmetry and transitivity

And not just paired associations

Because the pairs were always in one direction and she was tested in opposite directions

One can argue that by the end of all this training that the associations were just

lumpedSo that radio-tulip-crab was just some grouping distinct

from car-rose-fish

But that is what is meant by an equivalence class

That the objects do become connected and the

interconnections are solid

Even if the interconnections are not specifically trained

But, of course, what this has to do with language is still kind of iffy

What it does show is that Rio had some kind of representation

of the setsShe was, for example, never shown that car and fish were

relatedOnly that radio and crab were

related via tulip

And had to deduce that rose did the same for car and fish

The suggestion is that such representation abilities underlie

language

But, of course, these abilities may simply be needed in order

to survive

And have been co-opted for language

The same kind of issues are involved in the mutual

exclusivity article with my birds

Is the ability really language related

or just something that is used much more basic that is co-opted

to help with label acquisition?

I start by arguing that ME is considered a human strategy

mainly because the folks studying children argued for ME

to be uniquely human…

The question is to what extent is ME a basic strategy for learning

So, let’s look at what ME generally is….

When children are first learning labels

Their initial rule seems to be not only that every item has a

labelbut each item has only ONE

label

If this is a If this is a “turtle”, it can’t “turtle”, it can’t be an “animal”be an “animal”

That is, they see the labels “turtle” and “animal” as

mutually exclusive

Of course, if children were stuck in this stage,

they’d never learn appropriate hierarchical concepts or categories

And thus they outgrow it fairly quickly

But, in the early stages children—like Premack’s apes—will use

ME to help them learn

And not just object labels but also, later, attribute labels

So, initially, if you show a child a familiar item and an unfamiliar

one

And you say “take the dax”

Well, the child definitely knows that a shoe is “shoe”

Probably hasn’t seen a hole punch like that one

And, by exclusion, will take the hole punch and assume it is the

dax

When then given

And asked to “find the dax”

They won’t go for novelty again but take that hole punch

Thus they engage in what Susan Carey initially called “fast

mapping”The child might not understand

that “dax” refers to hole punches in general

But at least was able to associate the novel object and the novel

word from pairings with familiar ones

Now, the issue is really important because a dog seems to have done something similar

But children seem to be able to do this type of behavior for ‘dax’ and ‘wug’ at the same time (i.e., for two essentially novel items

And the dog wasn’t tested that way

And when Schusterman’s sea lions were tested after

supposedly learning a label-object connection via ME

that is, by being given a mix of novel and familiar items and told

“take the dax”

They failed…needing 200 trials to get it right

Now, for children, ME may sometimes help them learn that

the strange label is for an attribute

So, if they are playing with an oddly colored ball

And told that it is chartreuse

They may infer that chartreuse is the color of the ball

But that part of ME comes later in development

and I called that bit “attribute deductivity”

to separate it out from ME proper

But that’s what I really wanted to study…

Essentially, if you look at the ball-chartreuse issue

You are indeed giving the child an alternate label for the same

itemNow, that’s not what I had done

with Alex…

Alex Alan Alda

“woodwood” “keykey”

I had trained him that several different UNCOLORED things

could be called “wood” or “keys”

““green wood”, “green green wood”, “green key”key”

But then when it came time to train him on, say, colors

I simply used the color as an additional label

Not as an alternative label

And I used two different items that had the same color

And actually trained a novel category label as well…

So Alex saw me, using the M/R technique, work with a student

and say things like

“That’s green wood; the color is green and the matter is wood”

We structured the training so that ‘green’ was unlikely to be

considered an object label

An, interestingly, we used the same question “What’s here?”

all the time

Only if an item had color did he have to give the color

He didn’t seem to have any ME type issue

He learned the label just about as quickly as he learned any

other object labelAnd although he sometimes

made generic errors…

calling a colored object by just its object label

He didn’t make too many such errors

And, he very quickly learned to label other items with the

appropriate colors

and learned several other color labels and used them appropriately as well

So he seemed to understand how objects could have multiple labels

But did that mean that, unlike children, he didn’t have ME?

Or did his behavior have more to do with his training…

that is, using the attributes as additional labels, so he didn’t see

them as alternative label?

And thus had no problems…

What would happen if we trained birds in a way that more closely

resembled the ways in which children experienced items?

So that they saw lots of differently shaped (like Alex) but also

differently colored (unlike Alex) items

all given the same labels….

“truck”

“key”

“yellow” “blue”

That is, once they labeled all these colored keys as “key”,

Then I wanted them to label them just “yellow”

Like the children, an alternative rather than an additional label

I’ll stick to Griffin, because Alo was having behavior problems (had been abused before coming into

the lab)We trained Griffin on colors separately from objects…

via a computer screen

That in and of itself made it more difficult for Griffin to label

the colors

Because he didn’t get anything he could handle, just the ability

to request a treat

But he did eventually get them

But when it came to getting Griffin to transfer his color

labels to his objects

He had major issues

He couldn’t immediately transfer and actually had to be trained,

slowly and painfully

And, when given a new object…plastic cups…that came in

various colors…

He refused to learn the label “cup” and insisted on giving its

color

Even to this day he won’t label “cup”

And, interestingly, children who are actually given lots of

information like “This is a dog; a dog is a kind of animal”

like Alex, rarely exhibit ME

Whereas most children respond like Griffin

So training is extremely important in terms of how

learning occurs…

And might help children with dysfunctions who have trouble learning labels for attributes

But what does this mean for the concept of ME?

Most likely, ME is a special form of general information

processing

Even pigeons and rats, given tasks that have nothing to do

with language

exhibit something called ‘blocking’

Which, in brief, means if they have learned to make a particular association

say a tone and some food,

they find it more difficult to make the association between light and

the food

One way of figuring things out would be to see if Griffin could learn a new set of categories

with training like Alex…

So far, he is resisting….

But we keep trying…

Now, there are other parallels between nonhumans and

humans with respect to label acquisition

And these get into even more complicated areas

Such as brain maturation

Greenfield (1991) posited that control of such Greenfield (1991) posited that control of such parallel development initially resides in a parallel development initially resides in a

single neural structure (roughly Broca’s area) single neural structure (roughly Broca’s area) that differentiates as a child maturesthat differentiates as a child matures

In children, object and vocal label In children, object and vocal label combinations arise at about the same time in combinations arise at about the same time in developmentdevelopment

Want Want more more cookie…cookie…

In language-trained great apes, In language-trained great apes, object and signed or computer-object and signed or computer-

generated combinations also arise generated combinations also arise at about the same time in at about the same time in

developmentdevelopmentWant more Want more tickletickle

So let’s look in some detail at what Greenfield was

considering combinations

and what such combinations actually

meant

Greenfield’s definition of pairing strategies

Pairing is equivalent to “more + Pairing is equivalent to “more + X” or “want + X”X” or “want + X”

According to Greenfield, According to Greenfield,

Potting is equivalent to “want + more Potting is equivalent to “want + more + X” but the child could also say “X + + X” but the child could also say “X + want + more”want + more”

Subassembly implies understanding Subassembly implies understanding that “want + more + X” is made of that “want + more + X” is made of “want + more” and “more + X”“want + more” and “more + X”

Greenfield and colleagues (Johnson-Pynne et al. 1999)

argue for a homologous neural substrate in apes

and suggest the possibility for monkeys…

They trained Cebus monkeys to do the cup-stacking task

The training occurred painfully and slowly, but did occur

Interestingly, Cebus have some limited combinatory behavior in

the wild

They combine two actions and sometimes two calls

So, the researchers argued that all this had relevance for the

evolution of language…

A primate system, in which humans and apes were

language-ableand monkeys showed some

precursor ability

That is, one might conclude that parallel development of communicative/ physical combinatorial acts initially controlled by a purported single neural center is

likely unique to primate evolutionary lineslikely unique to primate evolutionary lines provides support for the motor/gestural provides support for the motor/gestural theory of language evolutiontheory of language evolution

not unique to humans, not unique to humans,

is the beginning of primate language-is the beginning of primate language-like syntactic developmentlike syntactic development

And basically everything else could be ignored….

All the combinatory work with the dolphins and sea lions

And, of course, no mention of creatures so evolutionarily

distant from primates as birds

But Huber had shown some data for action

ordering in parrots, and then there was

Griffin

A student had been cleaning up the lab and put lots of bottle caps on the counter where Griffin was sitting…

She heard a bunch of clanking and looked up to find

Griffin putting one cap inside another and dumping them on the

floor

Now, Griffin was already saying things like “want nut”

and “want corn”

He had learned these from watching and listening to Alex

And seemed to understand “want X”…

So we couldn’t examine the emergence of 2 object and 2

label combinations…

but what about 3 object and 3 label combinations?

He wasn’t yet doing either of those things

So we gave him bunches of things to combine….

Or bottle caps of various sizes

Griffin’s data for combining objects:Griffin’s data for combining objects:

Method AttemptsMethod Attemptsaa SuccessesSuccessesbb

Pairing 233 217 74 38 29Pairing 233 217 74 38 29

Pot 12 7Pot 12 7Subassembly 6 0Subassembly 6 0

a a Any effort made to combine objects; Griffin had 16 failures. Overall number of Any effort made to combine objects; Griffin had 16 failures. Overall number of trials was 251; 244 of these were with three objects.trials was 251; 244 of these were with three objects.bb A stable assembly. A stable assembly.cc Griffin picked up and put down a successful assembly; no further manipulation Griffin picked up and put down a successful assembly; no further manipulation

dd Successful assembly tossed off the counter. Successful assembly tossed off the counter.

ee Any other manipulation of assembly (e.g., destruction); includes trials in which he, Any other manipulation of assembly (e.g., destruction); includes trials in which he, for example, flipped internal cap over while it was inside external cap).for example, flipped internal cap over while it was inside external cap).

UnitUnit Pick up Throw Pick up Throw

ManipulateManipulate

Note that Griffin also frequently

picked up objects side-by-sideplaced caps over bits of foodplaced objects into large plastic rings balanced small objects on large ones

Now we had to train him form some 3 label combinations….

So we chose 2-corner paper/wood and 5-corner

paper/wood

In the space of the study, he learned only one of those

But we taped him when he was playing on his counter

And found that he could produce a large number of different combinatory vocalizations

Most of which we—nor Alex—never used nor trained

Two-label combinationsTwo-label combinations Three- or more label Three- or more label combinationscombinationswant pawant paaa green cornergreen corner want pop cornwant pop cornwant walnutwant walnut green nutgreen nut want corn nutwant corn nutwant corkwant cork green beepergreen beeper want grain nutwant grain nutwant grapewant grape green water green water want cork nutwant cork nutwant grain want grain green woolgreen wool want a nutwant a nutwant corner want corner green birdiegreen birdie wanna cork nutwanna cork nutwanna go-backwanna go-backbb go back go back wanna corn nutwanna corn nutwanna come-here?wanna come-here?bb go chair go chair want some want some corn(er)corn(er)wanna chair wanna chair here corn here corn 2-corner wood 2-corner woodcork nut (an almond) cork nut (an almond) walnut grape walnut grape cork nut showercork nut showerwhat’s here? what’s here? five corner five corner go back chairgo back chairwant ringwant ring do you wanna do you wanna grate?grate?

do you want grape?do you want grape?you wanna go-back?you wanna go-back?bb

aa “Pa” is Griffin’s term for pasta. “Pa” is Griffin’s term for pasta. you wanna nut grapeyou wanna nut grapebb “Come-here” and “go-back” seem to function as single units, i.e., unanalyzed wholes. “Come-here” and “go-back” seem to function as single units, i.e., unanalyzed wholes.

And splitting up the data like this

Doesn’t get into the ways that Greenfield examines children,

on the basis of phonemes

Dealing with “want cork” and “want corn” as three item

combinations

Issues of note:Issues of note: Griffin was not trained on object-based tasksGriffin was not trained on object-based tasks We looked for spontaneous manipulative behavior We looked for spontaneous manipulative behavior as vocalizations developed complexityas vocalizations developed complexity

He rarely combined >2 items, but percentages of He rarely combined >2 items, but percentages of 3-object and 3-label combinations were comparable3-object and 3-label combinations were comparable

We limited training on 3-label combinations; we We limited training on 3-label combinations; we then could document spontaneous utterancesthen could document spontaneous utterances

We trained “x-corner wood/paper”, but “2-corner We trained “x-corner wood/paper”, but “2-corner wood” was only 1/14 recorded 3-label combinations wood” was only 1/14 recorded 3-label combinations (i.e., labels Griffin chose to combine were rarely those (i.e., labels Griffin chose to combine were rarely those that he was taught)that he was taught)

And, probably the most interesting,

The timeline!

Now, parrots don’t have a Broca’s area

Although new data suggest striking parallels between avian

and mammalian brains….

particularly with respect to areas involved in vocal

learning…

Researchers now think that the functions in mammalian Broca’s area are spread over more than

one avian area

And, of course, if you remember all the material on MNs, you begin to wonder about the

connection between MNs and these data

Other birds demonstrate related combinatorial behavior: Male marsh wrens, Cistothorus palustris

engage in complicated woven nest building at the same time that they construct and memorize hierarchies of neighbors’ song repertoires so they can serially order their own song responses (i.e., reorder or recombine their own songs in new ways) to best defend their territories

Does the same neural substrate initially mediate concurrent vocal and physical combinatory behavior in any given avian species?

So….

Does a single substrate, if found, differ across avian species?

Neural structure of vocal control pathways differ between oscine songbirds and psittacids; the supposition has been that vocal learning arose independently in these groups…

but might the responsible structures may have been inherited from a common ancestor, and been lost in some descendants?

if one structure is indeed initially responsible for object-based and vocal combinations and

ThusThus….

even if this structure differentiates as an animal matures,

then psittacine and oscine combinatory behavior either

likely arose independently or both evolved from a much earlier brain structure

in both wrens and parrots, physical and vocal combinations involve beaks and tongues;

Note, too….Note, too….

such data, along with arguments that spoken language derived from gestural/motor forms without major neural restructuring, add credence to a motor control theory of song origin (Nottebohm, 1991) and possible involvement of a related functional neural substrate

Arguably, emphasis should not be on a specific neural

substrate, but on the types of

coordinated neuronal activity linking various brain areas—

and the likelihood that alternative neurological means

can achieve the same end