the cognitive dog class 12: simple but reliable rules... note to self: did you remember to start...

The Cognitive Dog

Class 12: Simple but reliable rules...

Note to self: Did you remember to start timestamp?

Agenda

• Problem Set 2 emailed out and posted on website.

• Looking ahead

• Temporal decisions

• Representation

• Object permanence

• Means-ends

The paws that refreshes...

Oh no, not statistics again, but yes...

• Mean (average)

• Standard Deviation (measure of spread in individual samples)

• Standard Error (measure of spread in sample means, all based on the same sample size)

• Standard Deviation / SQRT(number of samples)

• Confidence interval (a range around a sample mean within which the “true” mean lies with a certain level of confidence...)

• sample mean ± t * Standard Error

• t depends on # of samples and desired confidence level

Oh no, not statistics again...

• Collect 50 samples from 50 dogs

• time to destroy old model and new model of DestructNot chew toy

• Calculate mean (tnew - told)/50 (null hypothesis mean = 0)

• Calculate standard error

• How different does mean have to be from 0 to say new toys are different than old toys...

• confidence interval = mean ± t * std error

• if range doesn’t include 0, then the results of the experiment allow you to reject null hypothesis..

Big themes coming up

• How dogs may learn about, and make sense of, a complicated world by using “simple” rules that work well enough to get by...

• Individually the rules may be simple, but the net effect is extremely powerful...

• may be especially true with respect to rules they use to learn & make use of human social behavior...

Examples we will be discussing

• Learning about objects and manipulation of objects

• Learning from others via observational learning

• Using human cues to guide behavior including attention & gaze, gesture & word

Example of a SBRE rule...

• Dog looks out window every evening as you come home?

• Look at clock on wall?

• Internal clock?

• Hear car coming up street/driveway?

• Associate smell of food prep. with arrival?

• Associate radio on with arrival?

My goals for you...

• You gain an appreciation for the power & elegance of using “simple rules & mechanisms” to make sense of the world...

• simple does not imply stupid or unfeeling...

• You gain an ability to observe your own dog and see how he/she may be using this simple, yet elegant, way to guide his/her learning and behavior

Decisions, decisions...

The big idea...

• Dogs have choices to make...

• When they make a choice sometimes it turns out better than expected sometimes worse...

• When they have to make the same decision in the future, how do they weigh past experience?

• Pay more attention to recent outcomes than less recent outcomes...

Experimental setup

Stage A: Random pail baited, dog finds food,

repeat 7 times with same pail

Stage B: Different random pail baited, dog finds food, repeat 7 times with same

pail

Test: Which pail does the dog go to?

Devenport, J. A. and L. D. Devenport (1993). "Time-Dependent Decisions in Dogs (Canis familiaris)." Journal of Comparative Psychology 107(2): 169-173.

Experimental setup: 2 Groups

• AB-T

• Test followed Stage B by 23.5 Hrs

• A-BT

• Stage B followed Stage A with 23.5 Hrs delay, & Test done immediately after Stage B

• In AB-T, choice evenly divided between A and B

• In A-BT, all dogs chose B then searched A

More recent information weighted more heavily

€

Egood =

1

tii

ngood

∑

1

t jj

ngood +nbad

∑

time27hrs ago

26hrs ago

25hrs ago

24hrs ago

good/bad

bad bad good good

time25hrs ago

24hrs ago

2hrs ago 1hrs ago

good/bad

bad bad good good

E good = (1/25 + 1/24)/(1/27 + 1/26 + 1/25 + 1/24) = .52

E good = (1/2 + 1/1)/(1/25 + 1/24 + 1/2 + 1/1) = .95

And why do I care?

• One scientist’s cognitive experiment is another dog’s foraging decision...

• Each time must balance...

• where food was last time

• possibly new information (e.g., person pointing at different bowl)

• Not a life or death decision

Decisions, decisions...

• Do I go where she is pointing, or do I go where food was the last time?

• Conservative (go with the safe choice) vs. Explorer (ok to explore)

• A dog’s willingness to experiment is a function of many things, but experience can have a big effect...

• Is the dog rewarded for trying something new, or punished?

• Rewarding what you want to see, “ignoring” what you don’t, can help make your dog a confident explorer

• Punishment based learning tends to produce conservative dogs...

Hare, B. and M. Tomasello (2005). "Human-like social skills in dogs?" Trends in Cognitive Science 9(9): 339-444

Representation...

Example: a bad storm is coming

• Immediate perception: wind and rain

• Correlated perception: cows lying down, gulls inland...

• Shallow representation: “red sky at morn, sailors take warn”

• Deep represention: “a low off of Cape Hatteras will intensify and move north hitting our area by evening.

The 64 Million Dollar Question

• When you observe behavior, what kind of mental representations underly it? Is it a representation of

• only the immediate perception?

• a weakly correlated feature of the world?

• features of the problem that captures just enough to be useful

• a deep representation that captures the problem?

The cognitive bias

• Deeper, more complete representations are “better” and indicate greater intelligence...

• My bias is to focus on cognitive bang for the buck.

• There is value & elegance in “cheap” representations that work just well enough to get the job done, mostly.

• More about the minimum needed to get by as opposed to how hard you work :-)

Object Permanence

Object Permanence...

• Commonsense about how objects behave in the world...

• presumably require building and using representations in the brain

• Have a developmental time course (aspects of OP come on line in a predictable sequence)

• Lots of work comparing sophistication of OP across species

Object Permanence

• Things we “know” to be true about the world come on line in predictable developmental order...

• Stages 1 & 2: objects cease to exist if can’t be perceived, but can track moving objects.

• Stage 3: Can pick out partially obscured object (if part of the object is there, the entire object must be there)

Object Permanence

• Stage 4: objects permanent in space and time...

• 4a: hidden object recovered if search starts before object completely obscured.

• 4b: hidden object recovered

• A-not-B error prevalent: go to last “found” location, not the observed hidden location.

• See it hidden repeatedly behind A

• Then see it hidden behind B, but still go to A

Object Permanence

• Stage 5: Mastery of Visible Displacement

• Stage 5a: A-not-B error disappears

• Stage 5b: Successive hidden displacements

Object Permanence• Stage 6: Invisible Displacement

• Subject sees object hidden in a container

• Container is placed behind a screen/box and object is removed from container.

• Container is brought out, subject is shown that it is empty and container returned to some location.

• Can the subject infer the location of the hidden object?

Development of OP

Gagnon, S. and F. Dore (1994). "Cross-sectional study of object permanence in domestic puppies (canis familiaris)." Journal of Comparative Psychology 108(3): 220-232.

Many species demonstrate object permanence fewer seem to demonstrate invisible displacement

Consistent pattern of development...

What’s the big deal?

• Object permanence and invisible displacement used as evidence for representational thought...

• Can they hold the “image” or “memory” of an object or event that they can’t directly perceive and use it to guide their behavior?

• It seems to be possible to create experiments to test for object permanence in other species, e.g., dogs...

• Researchers have an innate desire, it seems, to measure animals on the basis of what humans can do...

An aside (well a rant)...

• There is a little thing going on here that has its roots in “ontogeny recapitulates phylogeny”. The implication...

• infants are less cognitively sophisticated than adults, so other species must “think” more like infants than adults.

• No, animals think the way they “need” to think in order to solve the problems that face them in their ecological niche. There is no reason to suppose they think like infants.

Collier-Baker on visible and invisible displacement

The basic set-up...

Collier-Baker, E., J. M. Davis, et al. (2004). "Do Dogs (Canis familiaris) Understand Invisible Displacement?" Journal of Comparative Psychology 118(4): 421.

Invisible displacement protocol

• Subject sees object hidden in a container

• Container is placed behind one of 3 screens/boxes and object is removed from container.

• Container is brought out, subject is shown that it is empty and container returned to some location.

• Can the subject infer the location of the hidden object?

• G&D suggested adult dogs can solve this problem...

Gagnon and Dore suggested that adult dogs appeared to have invisible displacement

• Performed above chance

• Statistically significant difference, but not nailing it by any stretch of the imagination.

• And that is the tricky problem...

• Going from “observation” to “inference”

• They looked at the development of object permanence and invisible displacement.

Gagnon & Dore on development of Object-Permanence in Puppies...

• Tested 70 puppies from 4 weeks to 9 months

• Visible displacement seemed to come on line around 5 weeks and fully in place by 8 weeks.

• Older pups seemed to learn rapidly to check last location where object was hidden, and had more flexible search strategies (aided by learning)

• Some evidence of invisible displacement by 11 months

The big idea...

• Could Gagnon & Dore’s results be explained by the dog using “local rules” rather than a deep or even shallow understanding of invisible displacement?

• The only “slightly better than chance” performance was the “smoking gun”...

• Was there something about the experimental setup that the dogs were using to inform their choice?

Local Rule Learning...

• “Local rule learning refers to the chance discovery of an action that leads to success on a task, which is then repeated” - Collier-Baker

• In a sense, the animal is taking advantage of a regularity in the context that may be only weakly correlated with success, but beats “just rolling the dice”

Protocol for experiment one

• Starting position of ball and the target box was random, never the same more than two times in a row...

• Starting and ending position of displacement device random, but never “starting or ending in the same position more than twice consecutively”

• Started and ended in the same position

• Closed end faced dog

Experiment One...


Repeat of Gagnon and DoreDogs have

visible displacement

nailed

Better than chance, but let’s just say

they deserve a “Gentleman’s

C”

Rule out 3 local rules...

• “Use unconscious cue provided by experimenter”

• hide face & upper body via a screen

• “Go to last box visited by displacement device”

• after showing empty device go to another box

• “Go to box adjacent to displacement device”

• random ending location but never adjacent

Experiment Two


Significantly above chance

Significantly below chance

92 % of the error involved searching a

box adjacent to device

Experiment Two


Breakdown of previous results indicating success or failure when device was adjacent to target box and

when it wasn’t

Note difference in success when

device is adjacent vs. when it isn’t...

Drill down on adjacency...• “Screen may have been unduly distracting”

• have screen behind experimenter in other cases

• “Go to first box visited by displacement device”

• go to one box, reveal that it is still in device and deposit in target box.

• “Go to last box visited by displacement device”

• after showing empty device go to another box

• Explicitly control for adjacency of DD to target box

Experiment Three: adjacency matters


Significantly above chance

By itself suggests that they aren’t

using a “first box visited rule” or

cues by experimenter. May be using a

last box rule

Experiment Three: adjacency matters


“Stark difference in performance levels on adjacent and non-adjacent displacement trials across all four

conditions”

Their conclusion...

• “Thus, it appears that, in our study, at least, dogs approached the task with a bias for searching boxes next to the displacement device, regardless of adjacent and nonadjacent trial proportions...”

Experiment Four

• Children, chimps, orangutans may be biased as well to use simple strategy of searching adjacent boxes

• The lure of the displacement device may be too great a distraction...

• Remove the displacement device at the end of the trial

Experiment Four: removing the displacement device


43% of dogs in the no displacement device trials either stood in front of experimenter or stared at displacement device

Experiment Four: removing the displacement device


“Removing the displacement device did not push the dogs into using a representational strategy; in fact their performance

was diminished. These results suggest that dogs rely heavily on the displacement device to guide their search”

The bottomline on invisible displacement

• Collier-Baker :“Taken together, the results of these experiments certainly provide strong evidence to suggest that dogs do not solve the invisible displacement task in the same way as 2-year-old children do. Simple stimulus associations would seem to account for dog’s successful performance on the task”“

• Dore and Goulet: suggested that “dogs rely on the external cue of the displacement device to guide their search because it is associated with the object at the time of its disappearance”

• One way or another, the dogs seem to be using how the real world operates (DG) and simple associations to produce better than chance performance

• Its all about bang for the buck

Summary thoughts...

• In the invisible displacement tests, dogs seemed to use a feature of the environment to guide their search and by doing so they were able to do better on a task that otherwise they seemed to have no clue how to perform otherwise...

• It wasn’t a feature that people use to solve the task, or would even think of using

• The feature had nothing to do with the intrinsic problem

• In the next experiments, pulling on the string, dogs pulled on the string that exited closest to the ball, and tended to focus their attention on where it exited the box

Summary thoughts...

• While the studies referred to the dogs using “local rules”, just think of them as “simple but reliable enough” rules...

• SBRE rule: start search with boxes adjacent to displacement device

• What is “reliable enough?”

• Better than chance, or expected benefit is greater than the expected cost...

The story of the cuckoo...

Photo by Angela Anders

Osthaus

The big idea...

• Tested dogs for their apparent understanding of means-ends connections when pulling on a string to retrieve food...

• Experiment involved configurations of 2 strings, only one of which was actually attached to food.

• Tested which string dogs pulled on first...

• correct string, or

• string closest to food

Osthaus setup

Osthaus, B., S. E. G. Lea, et al. (2005). "Dogs (Canis lupus familiaris) fail to show understanding of means-end connections in a string-pulling task." Animal Cognition 8: 37-47.

Food

Red or Blue colored string

The Experiments


Do they yank the right chain?

When only one string

• All dogs able to get treat & got faster each trial & transferred from one case to next.

• Short string, some dogs just used tongue

• Generally, pawed at string, between exit & cube. Only 2 grabbed wooden cube

• In diagonal case, pawed at spot where string exited box


Two parallel strings, perpendicular to box

• Chose correct string at better than chance levels, but worse than in Exp. 1

• Better performance when strings close together?


Two parallel strings, acute angle to box

• Significantly better performance on exterior vs. overlap (11/12 dogs chose exterior string on more than 1/2 of trials)

• Proximity error/behavior seen: paw at string that exits closest to food...?


Two crossed strings

• Dogs retrieved food, but chose correct string significantly below chance

• Proximity error/behavior

• No improvement over trials


Conclusions...

• Dogs in the experiment acted as if they were following the following SBRE rule:

• Paw at string closest to food where it exits the box (closest point to food...)

• Similar behavior to what Adler & Adler saw (and you will see when you read it.

• No evidence of deeper understanding of task, or of learning to modify rule

Questions...

• A couple of assumptions are worth contemplating...

• The visual cues are salient to dog (either perceptually, or behaviorially)

• They get the treat no matter what, so what is the motivation to “ace” the test, is it only retrieval time, and pawing may feel good in any event...

• Understanding vs. performance (can’t vs. whateva)

But note...

• If one assumes that they use the SBRE rule of “pull on the string closest to the food at the closest point to the food”

• Then, the kind of observational learning seen in the Adler & Adler experiment would suffice to scaffold learning this rule.

the cognitive dog class 12: simple but reliable rules... note to self: did you remember to start...

Documents

behavior slide

sample mean t

mean t new t old

standard error t

pail stage b

desired confidence level

true mean

gesture word slide