Lecture 20 - Schemas and Scripts

Question 1

Schemas and Scripts general description

Answer 1

things that help us understand and encode things in memory for later use

Question 2

understanding and memory

Answer 2

memory is really for understanding: taking past experiences to understand what’s going on right now

Question 3

ACT

Propositions

Answer 3

basic unit of memory in ACT theory

there has to be a truth value that you can assign to it

Question 4

How many propositions does the following statement contain?

“Sally thinks her neighbor, Bob, is left-handed.”

Answer 4

three:

1) Bob is left handed
2) Sally thinks
3) Sally thinks Bob is her neighbor.

Question 5

How many nodes (concepts) does the following statement contain?
“Sally thinks her neighbor, Bob, is left-handed.”

Answer 5

five:

Sally

thinks

neighbor

Bob

left-handed

Question 6

The ACT model assumes that more activation of associated

nodes will lead to

Answer 6

stronger links (they build up over time)

Question 7

Adaptive control of thought makes an interesting

prediction based on the spread of activation to related links.

Answer 7

• The amount of activation leaving a node is divided by all the links exiting that node.
• More links should lead to a dispersion of activation, slowing reaction times.

• Fan effect: it takes longer to recognize sentences that include
concepts included in many other sentences. (The more distractors
you learn, the slower you are to retrieve the right information.)

Question 8

more links =

Answer 8

less activation

b/c so many links that could possibly be acitvated which leads to slower reaction times when determining a truth value

Question 9

Fan effect

Answer 9

it takes longer to recognize sentences (or make a truth statement) that include
concepts (many paths, many nodes) included in many other sentences. (The more distractors you learn, the slower you are to retrieve the right information.)

Question 10

Adaptive control of thought

But don’t we get faster at recalling a series of facts as we have more experience or expertise in that area?

Reder & Ross (1983)

Answer 10

had subjects learn facts with different fan sizes - the number of links coming off of a node (i.e. more or fewer related propositions).

• Subjects were then presented with a sentence and asked if it 1) was something they recognize appearing before or 2) if it was plausible it appeared before.

 − Participants were slower to recognize sentences with bigger fans (same fan effect as before).
 − However, they were faster to rate as plausible sentences with bigger fan size.
 − Knowing more facts may strengthen links and give you the gist of the information but there's still that span of activation

Question 11

This ACT model provides an important examination of fundamental
processes in memory and makes key predictions:

Answer 11

1. Memories are stored as meaningful propositions. The propositions contain nodes (the concepts) and links (associations to other concepts).
2. When a concept is activated in working memory, spreading activation is sent to associated concepts via the links. The more
   activations over time, the stronger the links become.

Question 12

ACT

Relevance in a current situations is determined by

Answer 12

historical activation patterns.

Question 13

ACT

Computational complexity is reduced because

Answer 13

not all details will be activated equally and working memory is limited.

you just need the gyst

Question 14

ACT

Retrieved information may be ____ and open to
____ based on associations.

Answer 14

incomplete

confabulation

(with spreading of activation not all of it will reach enough activation to reach working memory)

Question 15

. In the original form (McClelland &

Rumelhart, 1986), PDP has the following characteristics:

Answer 15

• The form of representation is neurologically (inherently biologically plausible) inspired and presumes that meaning is distributed over layers of many units.

• Processing is accomplished by changing the strengths (weights) of excitatory and inhibitory (valence) links between
units. (how much does one unit activate another unit?)

− Networks of neurons (concepts) that fire together, wire together.

− The broader term is connectionist model.

Question 16

Adding a concept in ACT adds a ___; adding a

concept in PDP ______.

Answer 16

local node

changes a network of existing units (change organization or linkage between units)

Question 17

excitatory

Answer 17

encourages activation across the board

Question 18

input layer

Answer 18

links can be of different strengthens (tells you how weighted the signal is)

can be excitatory or inhibitory

Question 19

A simple network is made of

Answer 19

input and output units. In more complication designs (not in book), you can have hidden units that add computational flexibility.

Question 20

In PDP models, episodic information builds

Answer 20

the semantic network.

Question 21

As each event (episode) provides inputs to the

network, it

Answer 21

changes the weights of connections between units.

Question 22

Multiple encodings of the same (or similar)

events

Answer 22

strengthen the connections, forming a conceptual network (semantic).

overtime episodic info accumulates to semantic categories

Question 23

Contex

Answer 23

is automatically stored with the episodic content. Partial input should also activate
(provide cues to) the conceptual network

encoding everythings!! (including background stuff)

Question 24

This kind of encoding is

Answer 24

very resilient to damage
(compared to the localist ACT model).

very distributed vs. ACT discrete model

Question 25

Importantly, the same network of units can represent different semantic content.

Answer 25

• Weights for each link change in response to input and feedback.
• You might see similarities in activation for categories.

at first the weights are pretty equal and overtime the weights of the hidden units change with exposure and similar things will be activated more similarly in the pattern of activation (that’s semantic knowledge): built up overtime with training

Question 26

semantic memory is the value

Answer 26

of the weights

Question 27

Should PDP models also have fan effects?

Answer 27

yes

because you strengthen weights overtime with more exposure

links will lose strength too

Question 28

PDP models make some of the same predictions as the ACT model.

Answer 28

Fan effects should also be seen: A PDP network trained with many different instances may have weak connections between
units. This would lead to slow (and incorrect) responses.
• With increased inputs (i.e. expertise), the weights
between links should change and the responses should get faster.

Question 29

But there may be some problems. For example, PDP may have overwriting (new memory erases old memory), but this doesn’t seem to exist.

Answer 29

− Not a problem for ACT, but may be an issue for PDP. If you change weights between units, you are essentially erasing the original memory.

− Perhaps the shear number of units makes this less a problem?

Question 30

PDP seems to offer many of the same advantages as ACT, but uses a more flexible (plastic) and resilient architecture:

Answer 30

1. Memories are stored over many units, with layers of a network. The loss of some units does not seem fatal to the maintaining the
   concept.• This is a biologically inspired system, with operation analogous to neurons within layers of cortex.
2. Processing is accomplished through activation of links
   (connections) between units.

• Links can be trained over many instances/inputs (connection between episodic and semantic memory). The
resulting semantic content is built from episodic information.

• Weighted links can be excitatory or inhibitory.
• Computational complexity is reduced because contextual cues (mainly) activate the relevant units. (environment becomes part of the training)

Question 31

So far, study of memory (STM/WM and LTMM) has focused on lists or words or simple sentences. Is this what memory is typically used for?

Answer 31

• Not so much.
• Outside the laboratory, our memory encodes current information to help us understand situations and act within them.
• Memories of similar /relevant instances in the past reduce the cognitive complexity of the present situation.
• Memory is a constructive process, using current input + topdown (previously encoded) knowledge.