Langston, Cognitive Psychology, Notes 3 -- Attention
 
I.  Goals.
A.  Where we are/Themes.
B.  Detection.
C.  Filtering.
D.  Search.
E.  Automatic processing.
F.  Concentration.
 
II.  Where we are/Themes.
A.  Here are some situations:
1.  You're driving in a strange neighborhood looking for "Long" street.  You accidentally turn on "Lone."
2.  You're thinking about a quiz that's coming up in your next class as you walk there.  Someone calls your name, but you don't hear them.
3.  You arrive late at a party and try to find your friends.
4.  You're driving home and want to stop at the store.  Suddenly you find yourself at home and you didn't stop.
5.  You're trying to think about the research paper you're working on, but you keep thinking of the great first date you had last night.
What do these have in common?  Attention.  The topic for this unit is to try to explain each of these facets of attention.  It plays a role in detection (the first situation), filtering and selection (the second), search (the third), automatic processing (the fourth), and concentration (the fifth).  We'll have something to say about each of these as we go.
B.  Where we are.  Remember, we're working our way through this box model of the mind.  We've talked about the sensory register and pattern recognition (last time).  The register holds information briefly, and pattern recognition figures out what the information is.  Today we're looking at the "filter" and "selection" components.  Both of these are controlled by attention.  Attention allows you to filter out some information, and select from what's left.  Overall, attention helps pass information from box to box and provides the mental energy for each box's tasks.
 
Information processing model
 
Attention:  (1)  highlights parts of the environment and blocks other parts, (2) primes a person for speedy reaction, and (3) helps you retain information.  The basic processes are focalization (selection) and concentration (staying on a subject).
C.  Themes.
1.  Is attention's effect early or late?  In other words, do you process to a high level (like after pattern recognition) and then select, or do you select early?  The model we're working from is fudging a bit.  It puts the filter early and selection late.  There is some debate about where each of these goes (maybe they're the same thing).  Putting some at both places is probably a safe compromise.
2.  What is attention?  It's treated as a magical property that can transfer information around, choose information to attend to, improve memory for information, block unwanted information, and prime desired information.  But, what is it?  You're eventually going to see it equated with mental energy.  You have a certain amount, if you're using it to do some task, there's not much left over for other things.
3.  What does attention do?  We'll see lots of proposals, some practical, some theoretical.  Hopefully, at the end of the unit we'll be able to make a chart outlining the role of attention in processing.
 
Top
 
III.  Detection.
A.  Is there a threshold:  Fechner.  A threshold is the minimum amount of stimulation required for you to have a sensation.  Some sample thresholds:
1.  Vision:  On a mountain in utter darkness on a clear night you can see a candle 30 miles away.
2.  Hearing:  A watch ticking 20 feet away.
3.  Smell:  A drop of perfume in a three room apartment.
4.  Touch:  The wing of a bee on your cheek.
5.  Taste:  One teaspoon of sugar in two gallons of water.
It would be nice if a threshold would work like a step function.  Before you reach it, you get nothing.  After you reach it, you always see something.  Why?  Since the threshold is the minimum, and that should be based on physical features (like the limits on vision), it should be a constant.
Unfortunately, the threshold looks more like an ogive.  In the middle, you sometimes see something, sometimes not.  This makes a threshold a statistical concept (point of 50% detection).
What's the problem?  Partly attention.  What you're looking for can have an impact as early in processing as your ability to even detect that something is there.  Here's an example:  You're getting ready for a date.  You expect a phone call to make final arrangements.  When you get in the shower, it's pretty noisy, and some of the sounds could have similar frequencies as a telephone ringing.  So, in this condition, listening for the phone could make you think you're hearing it ringing when it isn't.  You're a lot more likely to dash out of the shower by mistake when you're listening for the phone to ring than when you aren't.
B.  Signal detection:  Swets, Green.  Get rid of thresholds, talk about detection.  We're going to separate the two features of this situation: sensitivity and bias.  Sensitivity is the limit of your perceptual system.  Some differences are easy to detect, and some are quite difficult.  Bias is your willingness to say you detect something.
The situation when you're trying to detect something always presents you with the four options in this box:
 
Situation
 
The two states of the world (stimulus present, stimulus absent) combined with two possible responses (see something, see nothing) yield four situations.  Hit (H) = something, you see it.  False alarm (FA) = nothing, you see something.  Miss (M) = something, you see nothing.  Correct rejection (CR) = nothing, you see nothing.
Sensitivity has to do with the strength of the signal (thing in the environment).  Signals are always perceived against a background of environmental and/or neural noise.  The stronger the signal is relative to the noise, the easier it is to discriminate something from nothing, and the higher the sensitivity will be.
Bias has to do with the perceiver's willingness to say that they see something.  They have to place a criterion somewhere along the dimension of evidence for a signal.  Above the criterion, say "I see it."  Below the criterion, say "I don't see it."  In a noisy environment, sensitivity is low, so you always have to make some mistakes (M or FA).  Placing the criterion in different locations changes the likelihood of these errors.  If misses are expensive relative to FAs, make the amount of evidence required for a "yes" very low.  If FAs are expensive, make it harder to say "yes."  I can manipulate your bias by changing the costs of these two mistakes.  The whole situation can be characterized as below:
 
Distributions
 
Why is this good?  I can manipulate these and see what effect your bias has on what looks like attention.  The automatic part (detection) is separate from the attention part (bias).
 
CogLab:  We'll discuss your data from a signal detection task.
 
How does this relate to driving in a strange neighborhood?  You're looking hard for "Long" so your bias is low (it won't take much evidence for you to detect "Long").  When you see "Lone," there's a lot of evidence for "Long," it exceeds the criterion, and you mistakenly turn.

Application:  This section has a lot of application to human factors.  When detection is crucial and you can't increase d', you have to set the criterion based on the situation.  I have some examples of medical images that can make this more clear.
 
Top
 
IV.  Filtering.  At some point, you have to choose what to attend to.  A filter is the place where the choice is made.  Before the filter, everything comes in.  Only attended stuff comes out of the filter.  (If we decide a filter is even necessary.)
A.  Early filter models.  Broadbent places the filter early:
 
Filter model
 
This is based on two sorts of evidence.  The basic task is called dichotic listening.  You put on headphones and a message comes in each ear.  The messages are different.  The first evidence for a filter comes from a task called shadowing.  You listen to two messages, and repeat out loud (shadow) the one that is your target message.  What we look for is how much of the unattended message gets through.  If you have an early filter, little (or nothing) should get through.  Some examples of this:  I have four samples.  In one, there's a change in speaker on the unattended ear.  In another, there's a change in language on the unattended ear.  In the third, the message in the unattended ear switches to being played backwards.  In the fourth, nothing funny happens (this is called a "catch" trial, because if bias is making you cheat, we'll catch you).  Broadbent says "if it was being processed, people would notice this, but they don't, so attention operates early."
Another sort of evidence for filtering also comes from dichotic listening (Broadbent, 1954).  I play numbers in both ears.  Here they are for the first sample:
 
Left ear Right ear




5
 
Then, I ask you to tell me the numbers.  What happens is most people report all of one ear, then all of the other (3,2,5,7,4,1).  It's like a filter allows in one set of information, flips to the other channel (ear) and lets the rest in.  Test?
If I ask you to report them in sequential order (3,7,2,4,5,1) it's really hard.  Why?  You have to flip the filter back and forth.  If I speed up the transmission rate, it gets even worse.
B.  Attenuation models.  Problems:  Some stuff can get through.  Cherry (1953) describes the cocktail party phenomenon.  You're at a party in a conversation, and hear your name from across the room.  Even though you're shadowing your conversation, your name gets in.  Treisman (1960) did a more empirical version of this.  She played two stories, one in each ear.  At some point, the stories switched ears.  People who were shadowing briefly followed the story to a new ear even though shadowing was supposed to be to an ear, not a story.
What's the explanation?  Treisman goes for attenuation.  It's kind of like turning down the volume on the radio.  What you're attending has a high volume, the rest has a low volume.  The stuff in memory has some threshold.  If the volume exceeds the threshold, you'll be aware of it, regardless of what you're attending to.  But, since the volume is louder on the attended channel, there's a better chance of it getting through.  For unattended stuff to get in it has to be pretty special (like your name) or very relevant (like the next word in a story you're repeating).
So, it's not really a filter, just less in terms of amount.
C.  Capacity models.  A different way to look at filtering is to think in terms of capacity.  You have x attention to spend, spread it around to all of the tasks that need some.  The more capacity one task takes, the less there will be for other tasks.  This is like an extended version of the attenuation model.
The basic paradigm for testing this is called dual-task.  You give a person two tasks and look at the way they allocate attention.  Usually, one task is called the primary task, and the person is supposed to do as well as they can on this task.  The other task is called a secondary task.  The person is supposed to do this too, but focus on the primary task.  If capacity models are a good way to think of attention, then the harder the primary task gets, the worse people should do on the secondary task.  (The more attention you devote to the primary task, the less you have to give to the secondary task.)
It's not really wise to talk about filters when you take this view.  Instead, the person processing information can select whenever they want, but the more attention they have to devote to each item, the harder this will be.  Example:  It's easy to make a feature discrimination (a blue 'S' surrounded by red 'X's and green 'T's).  No attention is required for an early filter that picks out blue things before recognition happens.  But, if you have to decide based on meaning, that takes attention (the energy that allows recognition to happen).  A task that requires more processing before you can select will take more attention for each item, and will be really hard.  Let's put all of that to the test with an experiment.
Johnston and Heinz (1978):  The primary task is shadowing.  Two lists of words are read, one in each ear.  The easy version of this is to shadow a female voice when a male voice is in the other ear.  You can choose what to shadow before recognition, and it won't take much attention.  The hard primary task is to shadow fruit names when the other ear has animal names.  To know which you're shadowing, you have to process up to the meaning.  That takes a lot of attention.
The secondary task is to respond to a light.  You don't know when it's coming, but when a light comes on, you push a button.  The amount of attention you have available to devote to the secondary task depends on how hard the primary task is.  With male and female voices you have lots left over.  With categories, you don't have much.
Results:
 

Primary task Secondary task
Shadow one list (control) 
Easy primary task (voices) 
Hard primary task (categories)
1.4% error 
5.3% error 
20.5% error 
310 ms 
370 ms 
482 ms 
 
So, making the primary task harder takes more capacity and reduces performance on both tasks.  Still, people can move the filter from early (voice) to late (category) if they have to.
To tie it in with our example at the beginning, you don't hear someone calling for you because you're devoting your capacity to thinking about the quiz.  The filter cuts out new inputs.  Note how this is an internal thing that's taking attention.  What you think about doesn't always have to come from outside.
 
Top
 
V.  Search.  Look for one item in a set of unrelated items.  Treisman (1988) presents a model of this.  She sees two kinds of search.  Feature search is looking for a unique feature.  This doesn't require attention.  Conjunction search involves combining features.  This does require attention.  Note how this relates to capacity models.  Feature searches are before pattern recognition and are easy.  Conjunction searches require higher processing and are hard.
Example of conjunction and feature search (stimuli as in Treisman & Gelade, 1980):
Conjunction search:  You have to look for a combination of features (in this case, green and "T").  (Find the green "T")
 
Hard:  Harder:
X T X T  
X T T X  
T X X X  
T T X T
X T X T T T X T  
X T X X T X T T  
T X T T X X T X  
X X T X T X T X  
T X T T X T X T
 
Feature search:  You can look for a single feature (blue or "S").  (Find the blue "S")
 
Easy:  Just as easy:
X T X T  
X T S X  
T X X X  
T T X T
X T X T T T X T  
X T X X T X T T  
T X S T X X T X  
X X T X T X T X  
T X T T X T X T
 
A target with a unique feature will "call attention" to its location ("pop-out").
What do we know about these search tasks?
 
Conjunction searches  Feature searches 
1.  Take attention (the more distracters there are, the longer it takes to search) 
Location stuff: 
2.  Helped by cueing the location of target (know where in maps to conjoin). 
3.  If you don't know the location, you can't tell what it is.
1.  Don't need attention, "pop-out" (no increase in time with more distracters). 
Location stuff:  
2.  No help from cueing (don't need it). 
3.  Know what without knowing where
 
Why does it work like this?  Attention.  If you need attention, it's going to be hard.  So, back to our party, if your friend is the only one with dark hair in a room full of blond people, he/she will be easy to find.  But, if you have to conjoin features (short, dark hair, male), that will be harder.

CogLab:  Change Detection.  This is related to search, but it also has a more real-world application.  One thing for us to consider as we look at your data is whether certain things might be more likely to be detected than others.
 
Top
 
VI.  Automatic processing.  Some tasks don't seem to require attention.  These tasks are ones that have been performed many times, and appear to happen on their own when the appropriate stimulus is present.
A.  How do we know a process is automatic?  Three criteria:
1.  Occurs without intention.
2.  No conscious awareness/Can't be introspected.
3.  Doesn't interfere with other activities.
3'.  Fast.
One way is to test this is to ask a person to perform two processes simultaneously.  If doing two things at once is no harder than doing one thing, one of the tasks must be automatic.  For example, if you can talk to me and do mental addition, talking or mental addition must be automatic.
B.  Example:  Logan.  Algorithmic processing is step-by step, requires attention.  For example, do alphabet arithmetic.  A = 1, B = 2, ...  At first, problems like A + M = ? are hard.  But, with practice, they get fast.  Logan says it's because you've stored the answer after practice, so it's retrieval instead of an algorithm.
C.  Classic automatic process:  Reading.  Two examples.
1.  Reading by the whole word method.  You don't do well at counting letters because you normally don't process letters when reading.  As an example, count the 'F's in the passage.  Most people miss the "of"s, so they say two instead of six.
2.  Stroop interference.  It's hard to say "green" to "RED" written in green ink when you're supposed to say what color the ink is.  Automatic reading (or word identification) causes trouble.  Try this list (naming the color of the ink):
 
RED  
BLUE  
YELLOW
RED
BLUE  
YELLOW 
YELLOW 
RED 
GREEN 
GREEN 
BLUE
GREEN
 
CogLab:  We did a Stroop task as a demonstration of methodology, but we can look at the data again here.
 
So, why do you go home instead of stopping at the store?  Driving home is automatic and happened without attention.
 
Top
 
VII.  Concentration.  The last topic has to do with your ability to pay attention.  There are two sides to this coin.  First, sometimes you have to attend to the same task for an extended period of time.  For example, you might be writing a paper.  This can be hard.  Other times, an unwanted thought keeps intruding.  You don't want to pay attention to it.  This can also be hard.
We can bring this into the lab with a suppression and expression task (Wegner, Schneider, Carter, & White, 1987).
Suppression:  Try not to think of a white bear.  This can be really hard.
Expression:  Try to think of a white bear.
I have each group ring a bell every time they think of it, they go five minutes.  That's the suppression and expression curves.  Note that both groups start out thinking about it a lot, and then both fade.
After suppression or expression, I reverse the task.  So, if you were suppressing, now you express.  These people are almost obsessive in their thinking about it.  If you were expressing, it's about the same when you switch to suppressing as suppressing alone.
So, why can't you stop thinking about that date?  It's like not thinking of a white bear.  Thought suppression is hard.  Simultaneously, keeping on a task is also hard.  We'll see some applications of this when we look at human factors psychology later.
 
Top
 

Cognitive Psychology Notes 3
Will Langston

 Back to Langston's Cognitive Psychology Page