Pattern Recognition

1
Chapter 2

• Pattern Recognition

2
The General Issue
When we look around our world we do not see
lines, colors, angles, surfaces, etc. Rather, we
see walls, desks, people, plants, etc. That is,
we see objects not the raw physical stimuli of
which they are formed. How do our cognitive
processes work to allow this to occur? One way
to answer this question is to ask what sort of
system could accomplish this feat? a question
that is asked in several contexts falling under
Machine Intelligence. One could then ask, of the
possible systems, which one does it appear humans
use the field of Artificial Intelligence.
3
Visual Agnosia - When the system fails
From Oliver Sacks book The Man Who Mistook his
Wife for a Hat What is this? I asked,
holding up a glove. May I examine it? he
asked, and, taking it from me, he proceeded to
examine it. A continuous surface, he
announced at last, infolded in itself. It
appears to have -- he hesitated -- five
outpouchings if this is the word. Yes, I said
cautiously. Now tell me what it is. A
container of some sort? Yes I said, and what
would it contain? It would contain its
contents! said Dr. P with a laugh. There are
many possibilities. It could be a change purse,
for example, for coins of five sizes. It could
4
The Challenge
In line with Chapter 1 then, the challenge is to
understand how the human object recognition
system works. Recall, our approach is this
regard is to first come up with some theory of
this system based on intuition and/or existing
data. Then we need to describe our theory, and
see if we can use it to predict new
findings. Note there is one obvious way in which
cognition research deviates from artificial
intelligence research while they actually
build the system, we tend more to describe it
theoretically although even this distinction
is fading of late.
5
Initial ObservationsTolerance to Variability
If we focus for a moment on word recognition
(words are one class of objects that are
relatively easy to experiment with), one thing we
know about the human system is that it is
verytolerant to variability The example
above doesnt even include hand-written Fs which
come in any many forms as the people that
produce them yet to us, they are all Fs Try to
create a mail sorting system with this feature
has been a major challenge for the post-office
F F F F F F f f f
f
6
Initial ObservationsTolerance to Partial
Information
Our system is also fairly good at dealing with
partial information, as when objects are
degraded or occluded.
HELLO WEIRDO
Obviously, whatever system we come up with has to
come up with some way to fill in the gaps when
we are dealing with incomplete and variable
information.
7
Initial ObservationsSensitivity to Context
The way that we see objects also depends on their
immediate context which is to say that higher
level information effects object recognition
8
A Theory Synthesis by Features
Given some of these facts (ignoring the context
issue for a moment) one theory of object
recognition is that we recognize recognize words
by first recognizing their letters, and we
recognize letters by looking for certain
features. Features like lines at certain
orientations, or curves, or the way lines
and/or curves meet (i.e., the angles that are
formed). Such a system would require
feature detectors thus one way to find evidence
for this view would be to find evidence for
feature detectors
WEIRDO
D
9
Evidence for FeaturesVisual Search
The visual search task is straightforward, you
are given some target to look for, and asked to
simply decide, as quickly as possible, whether
the target is present or absent in a set of
objects. For example, lets try a few searches
to give you a feel for this. Search 1 - Is there
an O present in the following displays?
10
Is an O present?
T T T T T O T T T
11
Is an O present?
T T T T T O T T T
T T T T TTT T T T T T T T T
T T O TTT T TT TT T TT T T TT T T T TT T
TTT T TT
12
Is an O present?
T T T T T O T T T
T T T T TTT T T T T T T T T
T T O TTT T TT TT T TT T T TT T T T TT T
TTT T TT
Q Q Q Q Q Q O Q Q Q Q
Q Q
13
Is an O present?
T T T T T O T T T
T T T T TTT T T T T T T T T
T T O TTT T TT TT T TT T T TT T T T TT T
TTT T TT
Q Q Q Q Q Q O Q Q Q Q
Q Q
Q QQ Q Q QQQ QQQ Q QQQ Q O Q Q QQ Q Q QQ
Q Q Q Q QQQ Q QQ Q Q QQQ Q QQQQ Q
14
Recap of Visual Search
These experiments show than when the target is
defined by a unique feature, it is found quickly
irrespective of the number of distractors this
so-called pop-out is assumed to support the
existence of feature detectors. Note that when
the distractors share the critical
feature, search is slow and highly dependent on
the number of distractors this kind of search
is called a serial self- terminating search. See
the text for discussion on asymmetrical search
patterns.
15
Further Evidence for FeaturesData from studies
of the brain
In addition to the visual search evidence,
investigators of the brain have shown that
certain structures in the brain (called columns
hyper-columns) appear to be sensitive to
basic visual features like lines at certain
orientations, and angles. This evidence was
acquired via cell recording studies in which the
activity of certain parts of the brain are
monitored in the presence of certain (typically
visual) stimuli if a part of the brain becomes
active in the presence of a specific stimuli, it
is assumed to be a detector for stimuli of that
type. For present purposes, the relevant part of
this is that it provides converging evidence for
the presence of feature detectors, thereby
supporting a synthesis by features notion.
16
But there must be more
As already highlighted with THE CAT IN THE
HAT example, context seems to effect what we see
thus there must be more to the story than
simple synthesis by features. Higher level
information must bias the way that percepts
are formed (Steve will now highlight the
distinction between sensation and
perception). This seems extremely puzzling how
can the word or phrase level of the stimulus
effect the way letters in it are perceived given
that one would expect that the letters need to be
perceived before the words, which must be
perceived before the phrase no?
17
Another Example of Context EffectsThe
Word-Superiority Effect (Reicher, 1969)

Until the participant hits some start key
18
Another Example of Context EffectsThe
Word-Superiority Effect (Reicher, 1969)
COURSE
Presented briefly say 25 ms
19
Another Example of Context EffectsThe
Word-Superiority Effect (Reicher, 1969)
U A
Mask presented with alternatives above and
below the target letter participants must pick
one as the letter they believe was presented in
that position.
20
Another Example of Context EffectsThe
Word-Superiority Effect (Reicher, 1969)



E
PLANE
KLANE
E T
E T
E T
Letter only Say 60
Letter in Nonword Say 65
Letter in Word Say 80
Why is identification better when a letter is
presented in a word?
21
A Return to the Problem
Once again, the word-superiority effect
demonstrates the puzzling effects of context (see
the text for other examples). One would think
that letters are identified before words
are. But, if this is so, why would the lexical
status of the item (i.e., whether it is a word
or nonword) effect the accuracy with which the
letters are identified? Clearly, information
about the word is being acquired as the letters
are being identified, and is actually effected
the way in which the letters are identified how
could this happen? Before we try to answer this
lets consider some other data that will help us
to form a theory.
22
Other Relevant Findings?
One typical procedure for studying the
recognition of words is to present a word briefly
(say 20-30 ms) followed by a post-stimulus mask
(some row of letters or symbols presented in the
same location as the target to prevent further
processing of it) - this is called a masked
word-identification task. Using this task,
researchers have found that humans can identify
more common words (e.g., TRUCK) better than less
common words (e.g., YACHT) - frequency
effects Also, they can better identify words
that have been recently experienced (a so-called
repetition priming effect), and this is
especially true for low-frequency words - priming
effects
23
Other Relevant Findings?
Participants tend to misidentify words with
uncommon spelling patterns as words with more
common spelling patterns (e.g., BOUT as BOAT),
and will misidentify nonwords (e.g., SALID) as
words that are like it (SALAD). Also, they have
more difficulty identifying nonwords
with irregular spelling patterns (e.g., ITPR)
than those with regular spelling patterns (e.g.,
PIRT). Collectively, these results are viewed as
indicating that the word-recognition system has a
bias towards well-formed stimuli, where a
well-formed stimulus is one that has
letter combinations consistent with words in the
language.
24
A More Complete TheoryA Feature-Net for Word
Recognition
Our previous simple feature net cannot explain
all these findings, but a more complicated
version can. While building this net, we keep
all these findings in mind, and try to design the
net such that it will be sensitive to the things
that humans appear to be. Thus, as a first step,
lets consider how to incorporate the frequency
and priming effects into a feature network
to do this, we need we need to introduce the
notions of activation, resting levels, and
thresholds.
25
Activation, Resting Levels, Thresholds
The job of a feature detector is to signal (or
fire) whenever its feature is present, how does
it do so? Here is a current view. Each detector
starts with some resting level of activation, and
some threshold level which, when exceeded, will
cause the detector to fire. When a feature is
present in the environment, it sends activation
to its detector this activation adds to its
resting level, increasing it if the increase is
sufficient to bring the level above
threshold, the detector fires.
Threshold
Rest
26
Bringing in Frequency Priming
When a feature occurs and its detector fires, the
new resting level is slightly higher than the old
one the detector is primed. This means it will
need less activation from the environment the
next time that feature occurs. Features that
occur frequently acquire chronically high
resting levels and therefore, in a sense, are
always primed. Thus, just like a primed
stimulus, a detector for a frequently
occurring feature will require less activation to
fire. This all means that some feature detectors
(those associated with frequent or recently
experienced features) are more ready to fire than
others. OK so back to the feature net.
27
Step 1Feature net that can handle frequency and
priming
28
What does this get us?
This explains why words that are more common or
primed can be better identified because words
that require less activation should be easier to
recognize. It also explains why uncommon words
are sometimes misperceived as more common words,
and why nonwords are misperceived as words since,
again, the network would be biased to perceive
items that require less activation. However, it
does not explain why regular nonwords
are perceived better than irregular nonwords as
there is no reason to believe that regular
nonwords should require less activation (i.e.
neither type of nonword has been seen - ever).
To explain this finding we need to add another
level to the network.
29
Step 1Adding a bigram level to explain
well-formedness
By adding a frequency-sensitive bigram level, we
can account for the findings of well-formedness
along with the others.
30
And so ...
Based on all of this, we are left with the claim
that human word recognition is based on a
feature-detector system that is biased to
perceive common or recently occurring
features. Based on this model, we can make
explicit predictions about situations where the
system will do well, and others where it will
make errors thus the system can be further
tested and refined. Steve will now use this as a
further example to explain the difference between
machine and artificial intelligence and the
book has a good section on considerations of the
pros and cons of this system you should check
that out as well.
31
Beyond Word Recognition
To what extent should this be viewed as a general
models of human pattern recognition as opposed to
a model of word recognition? There are reasons
to believe that other recognition systems work
in a similar manner as this one gt book
discusses geon theory of object recognition gt
Steves skunk versus cat example gt Steves
occlusion example However, there may also be
aspects that are specific to certain recognition
systems gt see the discussion on face recognition
in the text