Transcript Oral presentation slides

The Word
Superiority
Effect
OR
How humans use context
to see without really
seeing and how
can it help the field of
computational vision




Humans are not simply detectors of patterns of
light.
We infer interpretations of the physical stimulus
from the context.
The effect of context is made clear in a
fascinating phenomenon called the word
superiority effect.
The effect was first discovered by James Cattell
(1886).
First major breakthrew –
Reicher’s experiment in 1969



Reicher presented strings of letters – half the time
real words, half the time not – for brief periods.
The subjects were asked if one of two letters were
contained in the string, for example D or K.
Reicher found that subjects were more accurate at
recognizing D when it was in the context of
WORD than when in the context of ORWD.
How Reicher excluded factor like
memory and guessing?
Asking the subject only about one letter
(and not a ‘whole report’).
 Doing so immediately after the display.
 Using a forced choice task (rather than
identification), when both choices would
make sensible words.

Reicher’s findings


‘Word Superiority Effect’ : comparing four
letter words against non-pronounceable nonwords
(e.g. WORD vs. ORWD), there was an advantage
in reporting single letters from words than from
non-words.
‘Word-Letter Effect’ (WLE): report of one letter
from a 4 letter word was more accurate than the
report of a single letter presented alone.
Since there are four times as many
letters to 'perceive' with the four letter
words, this result is counter-intuitive,
and therefore quite striking.
So what is the subject of
my project?
The project’s goals
1.
2.
To conduct an experiment that will verify the
“word superiority effect”.
To try and examine if the effect, and the
hypothesis the researchers presented to explain
the effect, has any ramification that can help in
the computational vision field.
First experiment

The original experiment, similar to the way it was
conducted by Reicher.
Second experiment

An experiment which examines the effect of the
words and letters size on the effect. This is a
different version of the first experiment.
Third experiment

An experiment which is based on an
experiment which was conducted in 1979
by Adams. This experiment checks if the
subjects can recall details on the letters they
just seen.
The two major obstacles


Finding a way to present the words for a very
brief time (about 30 milliseconds), but in a way
that the words will be sufficiently seen on a
computer screen.
“Translating” the experiment to Hebrew. Since
the effect can work mainly on mother Tongue, it
was necessary that it be conducted in Hebrew.
The results
first experiment
Words
Letter strings
Letters
Success
Success
Success
percentage
percentage percentage
3 letters 0.738
3 letters 0.6
0.741
4 letters 0.725
4 letters 0.56
5 letters 0.786
5 letters 0.381
Second experiment
Size
Words
Nonwords
Letters
allinclusive
8
0.589
0.619
0.524
0.643
14
0.929
0.571
0.714
0.786
36
0.651
0.714
0.476
0.651
48
0.607
0.857
0.929
0.75
Third experiment
Only in 40% of the words, the fonts were
reported in 60% accuracy or more.
conclusions
I’ve found evidence for the word
superiority effect but not to the word
letter effect
Bottom-up and top-down processing



Bottom-up or data-driven processing: Processing
which is driven by the stimulus pattern, the incoming
data.
Top-down processing: Processing which is
influenced by the context and higher-level
knowledge.
To obtain such a result, one must postulate
interacting bottom-up and top-down mechanisms
which process information in parallel.
The Interactive Activation Model


McClelland and Rumelhart (1981) asked Exactly
how does the knowledge that we have interact with
the input?
The Interactive Activation Model is a system that
includes both bottom-up and top-down processing.
Interactive Activation Model
Word level
Letter level
Feature level
The Interactive Activation Model:
1. There is a node for each
word and each letter (in each
letter position).
2. The nodes are organized into
levels.
3. The nodes are connected to
all other nodes within levels
or betwen adjacent levels.
4. Connections may be
excitatory or inhibitory.
WORD
LEVEL
LETTER
LEVEL
FEATURE
LEVEL
1. The word "word" is presented.
2. The activation of "W" is shown
3. The features send activation to the letters.
3. The features send activation to the letters.
4. The letters send activation to the words.
4. The letters send activation to the words.
5. The words send back activation to the letters.
5. The words send back activation to the letters.
6. The same process occurs simultaneouly for "O", …
6. The same process occurs simultaneouly for "O", "R", and
…
6. The same process occurs simultaneouly for "O", "R", and
"D".
7. "D" in context is easier to recognize because it receives
activation from…
7. ...the letters …
7. … and the words.
Back to computational vision
What did we see so far?
Top-Down view: Appearance-based
recognition .
Network structure: Relaxation labeling
What can we add?
Combining the two kinds of processes
together. Not only comparing the stimulus
to our database and not only collecting the
stimulus from bottom up, but a
combination of both.
What can we add?
The top down process doesn’t work exactly like in
Appearance-based recognition :
 The process activates all the nodes connected when it
is starting to receive information. In fact, the process
activates many nodes that are barely connected. You
can even say that it wastes resources because for
every stimulus , many nodes are activated.
 It is a circular process, that means that if E activates
EAT than EAT will also activate E in reaction.
 The connections are not only excitatory, but also
inhibitory.
So what is the problem?
In order to get a useful tool, the database
should include a lot of data.
 In humans, this database is built by our
experience over the years. If want the
computer to have a big database will have
to teach him how to build this database
himself.

So what is the problem?
It also requires a lot of memory, to recall
the data on each node. If there isn’t enough
memory it will only extend the amount of
time which is needed to process the data.
 If there is inconsistent data (like the word
red written in green), the Interactive
Activation model will only delay the
decision.

The End