Chapter 10: Learning & memory

From: Harnad, Stevan (
Date: Fri May 09 1997 - 13:40:05 BST

Chapter 10: Learning & Memory

Levels of Explanation

You will have a whole course on learning and a whole course on memory
next year, so this Chapter can only skim the surface. This chapter
is also the best one in the book, so although it has more concepts, the
chapter presents them in a way that makes sense (even without my lecture

Is memory something real, or is it something that psychologists
attribute to us by analogy with the computer and its ROM (read-only
memory), RAM (random access memory, and disks?

There is a way to look at what we ordinarily call learning and memory
without using either word or concept: We could say that throughout their
lives organisms interact with their worlds -- whether the world of
trees, clouds, and water, or the world of predators and prey, or the
world of "social" interactions with other organisms of their kind.
(For our kind this would refer mainly to interactions via language.)

According to this learning-free and memory-free view, what some people
want to call "learning" and "memory" is just an expression of our
ignorance about what the daily actions and interactions of organisms
with their worlds does to their brains. Most events during an organism's
waking day CHANGE its brain in some way. A child who has touched a hot
stove and gotten burnt does not have the same brain he had before this
happened. The experience has changed his brain so the child is less
likely to touch a stove again.

We can say the child "learned" not to touch the stove, and a "memory"
has been stored in his brain, but if we really knew all the details
about this interaction and what change it caused in his brain, we would
not need to refer to "learning" or "memory." Knowing the brain and the
brain change would explain everything. Is this correct?

It is certainly true that if we knew exactly what changed in our brains
at every instant we could describe the change and leave it at that.
But if we did that, we would not be explaining the mind but just
describing the brain. Nor would we be understanding the brain, any more
than we would be understanding a toaster if we described every change
that occurred in the toaster from the moment we inserted the bread to
the moment we removed and munched it.

Description is not enough. Understanding how a system works requires
a CAUSAL explanation: What is going on inside a toaster that makes it
capable of taking a white piece of bread and turning it into a warm
dark piece of bread that we like to eat. To explain the toaster we would
have to explain how the lever lowered the bread, how the heat went on,
how the timer worked, etc.

Does this mean that we have to be toaster engineers in order to
understand how a toaster works? Not entirely, but you have to understand
an engineering explanation, otherwise you cannot understand the toaster.
On the other hand, you need not understand the PHYSICS of toasters
right down to its atoms to understand how they work.

By the same token, an understanding of how the mind works does
require a causal explanation of how the brain is able to do what it can
do, but this does not necessarily require understanding every neuronal
change that occurs in it.

Classical and Instrumental Conditioning.

You all know it, but this is a good time to recall what classical and
instrumental conditioning are:

When a dog sees, smells and eats food (Unconditioned Stimulus, US), it
salivates (Unconditioned Response, UR). If it hears a bell (Conditioned
Stimulus, CS) often just before it gets food, it will start to salivate
when it hears the bell (Conditioned Response, CR). This is classical
conditioning. It occurs with animals as well as people. It also seems
to require that the person or animal be CONSCIOUS of the pairing of the
bell and food; this is clearest in classical conditioning with human
beings: If the person is prevented from noticing the pairing of the
bell and the food because they are doing a distracting task such as
mental arithmetic, then the conditioning does not occur.

This is a controversial area and not everyone agrees that classical
conditioning is not possible without the subject being aware of the
connection. For example, electroconvulsive therapy (ECT) (shock
therapy) causes "retrograde amnesia": patients don't remember the shock
or what came before it. So if you ask them what it was like, they have
no idea. But if they are brought to the room for another ECT, they
start to feel frightened, even though they cannot say why: The room is
like the bell (conditional stimulus, US) in the classical
conditioning, and their fear is a CR.

Of course the patients ARE conscious just before the shock, even though
they forget afterwards. Maybe that's enough to cause the conditioning.
Some have reported classical conditioning during sleep (but we are not
unconscious throughout sleep; we just forget most of our dreams).

In any case, memory turns out to be an important factor in classical

Classical conditioning is thought to arise purely from association:
Pairing the CS with the US and UR produces an association so the
CS alone can produce the CR (salivation, the same as the UR, but
in response to the bell rather than the food, which is absent).

Instrumental or Operant Conditioning.

The way operant conditioning works is by rewarding Responses: The
responses become the "instrument" for getting the reward.
Operant conditioning is explained by the Thorndike's "Law of Effect":
Organisms tend to repeat any action that has rewarding effects.
Behaviourists have suggested that operant conditioning explains not only
the lever pressing and key pecking of rats and pigeons, but everything
we do as well, from recognising patterns to using language.

As you will see in this chapter, what REALLY needs to be explained
in human behaviour -- HOW association works and under what conditions
it affects later behaviour -- is not explained by association or
the law of effect. This chapter looks at the conditions under which an
association will "stick," and help you in a later task, and the
conditions under which it does not "stick." "Sticking" is the
pretheoretical, intuitive notion behind "memory."

After an event, there are many ways you can test whether it has left a
mark in memory: Suppose it's a word: Can it be RECALLED, with no cues
or help? Can it be recalled if you give a CUE (e.g., it starts with "M")?
Can it be RECOGNISED if you see it on a list with other words? And even
if all these tests fail, can you RELEARN it more readily than new words?

There are thought to be many kinds of memory. First, there is short term
memory (STM) and long term memory (LTM). Short term memories are
the ones you can recall immediately after something has happened:
He hear a word and are asked a few seconds afterward what it was,
and you can say what it was because it is still "echoing" in your mind.
The same is true if you see a picture and have to say what it looks like
very soon after you've seen it: You can still see the "image" or "icon"
in your mind.

This kind of short-term memory is called "echoic" memory for hearing and
"iconic memory" for seeing. There is an important difference between
iconic and echoic short-term memory, because you can "rehearse" echoic
memories, but not iconic ones. This means you can deliberately say a
word over and over to yourself (and to a lesser extent, you can
"imitate" a sound over and over in your head) to keep the short-term
memory "alive" for longer. You can't do this with iconic memory, because
whereas a sound is something you can DO, a picture is not (except for
artists, but even for them, it would take far too long to "draw" the
picture of what they have seen to "rehearse" it over and over in their

In fact, because of this difference between hearing and seeing
(namely that you can DO sounds, especially words, but you cannot
really DO sights) even visual images are rehearsed and remembered using
words and echoic memory. Recall the Chapter 5 and the many
ways that speech is "special." When you study language in later
years, you'll hear some of the reasons why language took the form of
speech rather than, say, gesture (although gestural languages
do exist).

Iconic and echoic memories are very brief (unless rehearsed), and that
is why they're called short-term memories. If you need to remember
something a few hours, days or years later, you have no short-term
memory "trace" left, because short-term traces "decay" within a few
hundred milliseconds. You can still remember, though, on the basis of
a different, long-term memory.

There are many different memory "dichotomies": Many memory theorists
will say that they study 2 kinds of memory; the trouble is that
the pairs are not all the same ones: One possible way to put them
together is as in figure 10.4:

Two Kinds of Memory: STM & LTM

First, there are two kinds of memory: short term memory and long term

Two Kinds of Long-term Memory: Declarative & Procedural

Then there are two kinds of long-term memory: "declarative" and
"procedural" memory.

Procedural memory is memory for HOW to do something. Your memory for
riding a bike is procedural: You know HOW to do it, but you're not
very good in EXPLAINING (or "declaring") how you do it.

Declarative memories are the ones you can "declare." Declarative
memory is memory THAT something is the case. (I know that Tony Blair is
the new Prime Minister. I remember seeing the announcement of
the last poll.)

Two kinds of Declarative long-term Memory: Episodic and Semantic.

Episodic declarative memory is memory for the episode, the experience.
My memory for the moment of the election announcement is episodic.
I know where I was, what I saw, etc. at the time.

Semantic declarative is memory for facts: You can usually describe
semantic memory as a proposition: I know ( = I remember) that
Tony Blair is Prime Minister. I also know (remember) that Bill
Clinton is President, but I can't remember the episode of when I first
learned that. There are many things we know, hence remember, without
remembering when and where we learned it; without remembering the

Nonsense Syllable Studies of Memory:

Ebbinghaus's simple experiment of having people hear strings of nonsense
syllables and then testing their memory for them under different
conditions has provided a lot of interesting information about memory.

First, there is the well-known primacy/recency effect: You
remember the first and last items better than the middle ones.
The last items are probably remembered better because they
are still in short-term memory. The first items may be remembered
because they could be rehearsed longer. The middle items
are the one that have most interference between items both in
front of and behind them.

In his famous 1956 paper on the magical number 7 +/- 2
<br>George Miller
found that we cannot hold more than about
7 items in short-term memory. This would apply to the
number of digits we could recall, or the number of nonsense syllables.
For example, if I said to you out loud:

one - seven - two - nine - four - eight - two - ...

You could manage to repeat it back after I said it. But if I added more
digits, your recall would go down.

However, if I RECODE the digits as two digit numbers, you can remember
twice as many:

seventeen - twenty-nine - forty-eight - twenty-six - thirty-nine...

A lot of cognition consists of this kind of recoding of information into
bigger "chunks" as Miller called them. Language itself is a means
of recoding complicated information into compact. portable words and
sentences. You will learn more about this when you study language
directly. You can try to relate it to the language chapters in
this book as well.


It is thought that information in short-term memory has to go through a
period of "consolidation" before it becomes a more-or-less permanent
item in long-term memory. Some theorists think that consolidation
occurs during sleep. Some of it clearly occurs while we are awake,
though, otherwise we would not be able to remember something for
30 minutes!

Memory consolidation is thought to involve a part of the brain
called the "hippocampus." Evidence for the hippocampus's role in
consolidating long-term memory comes from patients like the famous
"HM" who had his hippocampus on both sides of his brain removed in an
operation for epilepsy.

HM has no problem with his short-term memory -- he could remember
7 +/- 2 digits, just as we can. But he couldn't remember them longer
than the scope of short-term memory, which decays quickly, so that
after a few seconds, it's gone. Now life is continuous, rather than
a string of discrete digits. HM does manage to maintain continuity
as the day goes on: If you are introduced to him and he is told
your name is Tony Blair, HM will say "Hi Tony" and will continue talking
to you, but a minute later, he will no longer remember your name,
and if you leave for some hours and return, he will not recall ever
having seen you before.

HM has been saying, ever since his operation in 1954 that "Everything is
coming back to me now..." He says it today, as he did the day after
his operation, because for him, every day IS the day after his
operation. If things happen gradually, like the aging of his mother,
he will not be shocked to say an old lady 30 years older than
her age the day after his operation. He continues to recognise his
mother and is not troubled by the way she looks. This is in part
"anosognosia" the unawareness of one's own neurological illness.

If an attempt is made to counter HM's anosognosia by drawing attention
to the aging of his mother by asking him what the date is, and how old
she is, he will date her as 30 years younger than she really is. If you
then say "Doesn't she look a lot older than that?" he will ponder
and say "Yes, now that you mention it" and he may be troubled about it
for a few seconds, but no longer, as by then he will have forgotten.

A much more moving sign of his amnesia is the fact that his father
died a number of years after his operation, and when he was told, he
naturally burst in tears. But since then, every time he is told
his father has died he bursts into tears as if he is hearing it for the
first time -- because he IS hearing it for the first time, insofar as
his memory is concerned.

HM's injury was mostly in the hippocampus, and that is why he has
"anterograde amnesia" -- amnesia for anything new happening since his
operation -- except for the short span of his short-term memory.
Other patients can have the reverse problem: Long-term memory
intact, but a much shorter short-term memory span (e.g., able to
recall only 1 digit instead of six).

When you have injuries in two different areas, like the hippocampus and
the cortex, and two different deficits, e.g. long-term memory or
short-term memory deficits, and injury in one area and not the other
produces the one deficit only, and the reverse pattern of injury
produces the second deficit only and not the first, this is called a
"double dissociation." This is a compelling way to provide evidence
for the existence of two independent functions that would otherwise only
be theoretical hypotheses.

HM lacks long-term declarative memory (both semantic and episodic)
but, apart from short-term memory, he also has near-normal
Procedural memory. If you taught him the computer game of tetrus,
he would play it for, say an hour with no problem; he would also
get better at it, as we all do, during that hour's practice.

If you returned the next day, he would not remember either you
or the game, so you would have to identify yourself and teach him all
over again, but he would quickly reach the level he had the last time
and go beyond it. In other words, he improves with practice, just as we
do, but he has neither episodic memory for the practice, nor semantic
memory for the fact that he knows how to play tetrus. But his procedural
memory for tetrus would be near normal.

Procedural memory in certain amnesias is not the only one spared:
Semantic memory may still be good -- the memory for new facts. What is
typically the worst in such amnesias is episodic memory: You know new
facts, but you don't know how you know them: You cannot recall the
episode from which they originated.

Now you may no doubt be diagnosing yourselves with amnesia, because for
most of your knowledge you also have no idea what the original episode
was. So it is true that we are all PARTLY amnesic: We remember the
fact or skill, but not the experience. Recall Funes the Memorious and
Luria's Mnemonist, the stage memory artist "S." They had much better
memories for every episode of their experience, but as a consequence
they had problems learning, abstracting the invariant feature or the
semantic fact that was shared by many different KINDS of experience.
To be able to categorise and to generalise from experience, some parts
of it have to be selectively noticed and stored, and other parts -- most
of it, usually, must be ignored of forgotten. Much of learning consists
of this kind of selectivity, ignoring the irrelevant part:

Well, almost all of the episode in which you learned, or began to learn
a fact is typically not only irrelevant to the fact, but actually masks
it with all the irrelevant details. To extract the semantic fact (or the
procedural skill), most of the episode must be suppressed.

In the early stages of learning, a lot of attention and conscious
control is needed, because you do not yet know what features are
relevant. This is called the "controlled" phase of learning. But as you
learn to ignore the irrelevant features and notice and use only the
relevant invariant features, the skill becomes more automatic.
This is similar to learning to recode information into bigger "chunks."

When you are learning to drive, in this "controlled" phase, it is risky
for you to talk at the same time; but once you have learned how to drive
and it becomes automatised, you can speak while you are driving.


As the story of Funes the Memorious suggests, it would be a handicap
to remember every episode exactly: You could never learn at all.
There are many different theories about forgetting: One mechanism
would be interference: a later episode "scrambles" the trace of
the preceding one in short-term memory. Another theory is that new
memories "over-write" prior ones in long-term memory. Experiments
show that the timing is important, because for the first quarter
hour after an episode interference effects can be shown, but by
a half-hour after the episode, memories are immune to retroactive

Read the experiments described in the section on forgetting
(Pp 299 - 302) so you can follow the logic of the experimental analysis
of normal memory.

Models for memory.

Neural nets are being used to model memory. In backpropagation, as you
will recall, units are interconnected and there is feedback from
correct and incorrect outputs in response to inputs: The connections
that lead to a correct output are gradually strengthened and the
connections leading to an incorrect out are weakened. There are
parallels to this strengthening and weakening of connections in the
hippocampus. The brain also appears to use "distributed"
memories, rather than memories localised in one place.

At this point memory theory starts to join with general cognitive
theory of concepts. This chapter suggests that concept learning
experiments with human beings have supported the "prototype theory" of
representation, because of such effects as better memory for a
"prototype" that you have never seen than for less "typical" examples
of it that you have seen. This effect is similar to the caricature
effect in the chapter on face recognition. There are complications
about these theories of representation, but they can only be
understood in an upper-year course on mental representations.
If you want to think about it, here is a hint: What is the difference
between recognising on the basis of a stored "prototype"
versus an invariance-detector?

This archive was generated by hypermail 2b30 : Tue Feb 13 2001 - 16:23:52 GMT