Recognition

1
Recognition

• Chapter 5
• Identifying and Classifying
• patterns, objects, faces

2
Ch 5 Identifying and Classifying Concepts
COGNITION AND CONSCIOUSNESS Subliminal Visual
Priming COGNITION AND INDIVIDUAL DIFFERENCES
Right Hemisphere and Self-Recognition
COGNITION AND NEUROSCIENCE Causal Similarity,
Perceptual Similarity, Childrens Categorization

• I. Bottom-up Views of Pattern Recognition
• A.Structural-Description-Based (SDB) Approaches
• Pandemonium Model Marr RBC
• B.View-Based (VB) Approaches
• Template Matching and Object Recognition
• II. Top-down Processing in Pattern Recognition
• A.The Word-Superiority Effect
• B.A Scene-Superiority Effect
• C.Two-System Views of Object Recognition
• Structural Description and Episodic Systems
• III. Recognizing Faces
• A. The Face-Inversion Effect
• B. Wholistic Processing of Faces
• C. Dissociations/Associations in Face Recognition
• IV. Concepts and Categories
• Similarity Prototypes Exemplars
• Explanation-based view on Concept Representation

3
Identifying and Classifying Objects

• Pattern, object and face recognition how does
  this remarkable feat happen?
• processes whereby we match a stimulus with stored
  representations for the purpose of identification
• I. Bottom-up Views of Pattern Recognition
• views of pattern recognition differ primarily in
  their characterization of bottom-up processes
• basic question--when engaged in object
  recognition, what are we matching the information
  to in memory?
• A. Structural-Description-Based (SDB) Approaches
• compare patterns to a structural description in
  memory includes a list of the visual features,
  as well as the relationships between these
  features (often called feature analysis)
• representation in memory is not visually or
  spatially analogous to the pattern being
  recognized
• compare features of incoming stimulus to an
  abstract description of those features in memory
• advantage particular orientation or view of the
  pattern is not important no matter what the
  perspective, patterns are broken down into
  component parts and compared to a structural
  description in memory (also called view-point
  independent)

4
Pandemonium Model

• feature analysis is carried out by hierarchically
  organized demons
• -entities each carry out different jobs that
  culminate in pattern recognition
• -each set of demons passes their info on to the
  next set for further analysis
• image demons responsible for the initial
  encoding of the pattern
• feature demons analyze the now encoded stimulus
  in terms of its component elements (e.g., /, -,
  and \) different feature demons look for each of
  these elements
• cognitive demons monitor work of feature demons,
  looking for evidence that supports a specific
  pattern
• -each cognitive demon represents a different
  pattern (e.g., A, V, W, B)
• -cognitive demons shout with each piece of
  evidence (i.e., info from feature demons like /,
  -, or \) that supports their designated pattern
• -multiple cognitive demons will be shouting if
  the to-be-identified pattern shares features with
  several cognitive demons
• -e.g., the A cognitive demon, plus the V
  cognitive demon will be shouting
• decision demons assess shouting of cognitive
  demons
• evaluate which one (e.g., A or V) is shouting
  the loudest (i.e, has detected the most
  evidence) use that information to decide on
  pattern

5
Pandemonium model

• model fits well with what we know about lower
  level perceptual processes
• there are cells in the visual cortex that
  correspond to lines of certain lengths,
  orientations, etc.
• so it makes sense to characterize pattern
  recognition as a gradual process of evidence
  accumulation based on a feature-by-feature
  analysis of incoming information
• problems
• -limited to explanations of how we recognize
  simple patterns like letters
• -representational economy--unlikely that we have
  a near-infinite number of feature analyzers to
  analyze a near-infinite number of features

6
Visual Pathways
7
Extracting Sensory MessagesVisual Processing in
the Brain (Pp. 130-131)

• Neural messages travel to brain via optic nerve
• Splits at optic chiasm so that information from
  left visual field goes to right hemisphere and
  vice versa
• Within each hemisphere, information goes to the
  lateral geniculate nucleus (90 in thalamus) and
  superior colliculus (10 in midbrain)
• Former deals with colour, texture, depth
• Former deals with movement
• Demonstrates parallel processing
• From the lateral geniculate nucleus,
  informationgoes to primary visual cortex (in
  occipital lobe)
• Feature detectors Cells in cortex that react to
  simple visual stimuli such as edges, lines,
  angles (may be orientation specific)

8
Feature Detectors in Visual Cortex
9
Marrs View

• goal of the visual system is to transform a
  2-dimensional retinal image into a 3-dimensional
  percept that is quickly and easily identified
• processing stages
• first stage register the information contained
  in the retinal image
• note the specifics of the retinal image depend
  on the particular view of the object)
• information about light intensity, boundaries,
  discontinuities, edges, and groupings
• used to form a primal sketch--rough rendition of
  the most primitive elements of the object
• information from the primal sketch is used to
  construct a 2½ -D sketch
• includes information about orientation, relative
  depth of the visible surfaces also about
  discontinuities in depth and orientation

10
Marrs View

• information derived from the primal sketch and
  2½-D sketch forms the basis for the construction
  of a 3-D model of the object
• the parts of 3-D model are termed volumetric
  primitives, but are basically generalized 3-D
  cylinders (a pipe-cleaner version of the object
  were looking at) -I.e., there is less detail
• unlike the primal sketch and the 2½ -D sketch
  (which differ depending on the perspective of the
  observer), the 3-D model is not view-specific
  (viewpoint independent)

11
Recognition-by-Components (RBC) Theory

• object recognition is a matter of parsing objects
  into features
• analyzed features are not simple line segments,
  angles, and curves the features are basic 3-D
  shapes termed geons
• a total of 36 geons that serve as visual
  primitives--simple shapes that can combine to
  form most other more complex shapes
• similar to Marrs theory in that it proposes a
  series of hierarchically arranged stages whereby
  information about component features is used to
  identify the object
• edge extraction process--looks for differences in
  features like texture, luminance, color, and
  results in a simple line drawing of the object
• -next, two processes occur simultaneously
• -the detection of nonaccidental features--actual
  features of the stimulus, rather than some
  accident of a peculiar observer perspective
• -parsing the object at areas where there appear
  to be boundaries between the parts of the object

12
Recognition-by-Components (RBC) Theory

• next, with the information gained to this point,
  the components of the figure, the geons, are
  determined
• this set of components is matched with object
  representations in memory
• when a match is found, the object is identified
• critical predictions
• objects should become less identifiable the
  harder it gets to recover their components

13
Biederman and Cooper (1991)

• participants were presented with sketches of
  objects in which 50 of the contours had been
  deleted
• for some figures, the contour deletion disrupted
  the figure at the points of segmentation that
  would be used to carve the object into its
  component geons
• Predictions
• object recognition should suffer
• for other figures, the contour deletion did not
  prevent recovery of the geons (although the same
  amount of the contour was deleted)
• object recognition should not be affected
• findings were consistent with the RBC theory
• rotation of objects should not hinder recognition
• changes in orientation do not influence the basic
  components of the object and their relationships
  to one another

14
View-Based (VB) Approaches

• objects are recognized wholistically through a
  process of comparison with a stored analog
• when a match is found, the pattern is recognized
• termed viewpoint-dependent because identification
  of an object depends critically on the particular
  perspective the viewer has
• to identify the object, an image matching this
  particular view must be found or, the incoming
  stimulus image must be manipulated in some way
  (e.g., rotated) until a match is found with
  images represented in memory
• An Early Attempt Template Matching
• our store of general knowledge includes a set of
  templates (copies) of every pattern that we might
  encounter
• when we encounter a pattern that needs to be
  identified the mind quickly rifles through its
  set of templates when a match is found, the
  pattern is given the label stored with the
  template (i.e., the pattern is recognized)
• even the slightest change in the pattern will
  lead to a recognition failure
• problems
• too rigid
• lack of economy

15
Modern Versions of the View-Based Approach
Tarr and Pinker (1989)

• taught participants names for shapes shapes were
  always presented in the same orientation during
  training
• during a test phase, these shapes were presented
  for recognition
• participants responded quickly if the shapes were
  presented at the same orientation as in training,
  but were successively slower as the degree of
  rotation from that original position increased,
  indicating that perception is viewpoint-dependent
• after training on the new orientations,
  participants eventually became equally fast no
  matter what the orientation
• results parallel the everyday recognition of
  objects
• we start with one representation of objects
  (termed the canonical representation), and
  through experience with objects in many different
  orientations and viewed from many different
  perspectives, we develop multiple
  representations, or views of the objects
• these multiple views serve as the templates for
  later recognition
• orientation tends not to affect visual
  recognition under most circumstances since
  everything we must recognize has received
  extensive exposure from different view

16
Logothetis, Poggio, and Poggio (1995)

• taught monkeys to recognize novel 3D objects from
  a variety of different perspectives, and like
  humans, they eventually became equally proficient
  at recognizing these objects given any of the
  rotations
• neural activity associated with recognition
  different sets of cells responded most strongly
  to certain objects, indicating that certain
  networks were devoted to certain objects
• a given set of cells responded most strongly when
  that object appeared in the same orientation as
  it had during training the responses decreased
  systematically with increases in the rotation
  from that perspective
• monkeys seemed to have what might be termed
  physiological templates that were devoted to
  recognizing a specific object in a specific
  orientation

17
Object Recognition Views or Structural
Descriptions?

• basic question--is object recognition is truly
  independent of the particular view there is
  support for both views
• Tarr and Bulthoff (1995) suggest that object
  recognition should be conceived of as a continuum
• one end are heavily view-point dependent
  mechanisms that are used for making subtle
  discriminations among similar exemplars
  (distinguishing a finch from a sparrow)
• explains Tarr and colleagues results--stimuli
  used required fairly subtle discriminations
• object recognition appears to be viewpoint
  dependent
• other end are heavily view-point independent
  mechanisms that are used when gross categorical
  judgments are required (distinguishing a hammer
  from a sparrow)
• explains Biederman and colleagues
  results--stimuli used require fairly gross
  discriminations
• object recognition process appears to be
  view-point independent

18
II. Top-down Processing in Pattern Recognition

• A. The Word-Superiority Effect
• Reicher (1969)
• participants were briefly presented with letter
  strings that either did or did not form a word
  (e.g., OWRK or WORK, respectively)
• following a rapid display of such a letter
  string, participants were queried about the
  component letters
• two alternatives were presented and participants
  were to pick the one they had just seen
• identification accuracy was higher if the letter
  had been presented in the context of a word,
  relative to when it had been presented in the
  context of a non-word
• if letter identification had been based solely on
  bottom-up processing, then letter identification
  accuracy should have been equivalent in the two
  conditions
• the data that make up the letter D do not vary
  depending on the context in which D appears a
  D is a D

19
An Interactive Approach to Word Recognition

• word superiority effect is the result of an
  interplay between the bottom-up processes and the
  top-down processes
• McClelland and Rumelhart (1981) model assumes
  that words are represented in our mental
  dictionaries at three different levels features,
  letters, and as whole words

• each type of information about a word is being
  analyzed simultaneously, and information about
  the words identity accumulates. -example the
  letter A
• activation of the A detector will excite nodes
  representing words that have an A, but inhibit
  nodes representing words that dont have an
  A--bottom-up
• activation of the A detector will excite nodes
  representing features that are part of the letter
  A (e.g., slanted lines), but inhibit nodes that
  are not part of the letter A (e.g., a curved
  line)--top-down

20
An Interactive Approach to Word Recognition

• when we read a word (e.g., work), info about the
  component letters leads to the activation of
  representations at the word level that include
  these letters
• this heightened activation at the word level then
  feeds back to the letter level, enhancing
  activation of component letters of the words
  activated
• but only the letters in a word (e.g., work) are
  receiving the bottom-up (feature-based)
  activation therefore, evidence for these
  particular letters will accumulate the fastest,
  facilitating their identification
• nonwords dont have representations at the word
  level the letters in the nonword (e.g., owrk)
  dont receive this additional top-down
  activation therefore speed of identifying a
  letter in a non-word will be slower

21
A Scene-Superiority Effect

• identifying an object within a scene is
  facilitated when the object is consistent with
  the scene (e.g., refrigerator in kitchen),
  relative to when it is inconsistent (e.g.,
  refrigerator in farm)
• Biederman, Mezzanotte, and Rabinowitz (1982)
• participants saw the name of a common object,
  followed by a real-world scene, and then a mask
  that included a location cue
• participants were to determine whether the common
  object had appeared at that location in the scene
• detection performance was better when the object
  was consistent with the scene relative to when
  the object was inconsistent

22
Hollingworth and Henderson (1998)

• a subtle response bias may have been at play in
  the Biederman procedure
• participants were told what to look for in the
  scene, prior to the scenes presentation this
  expectation may have influenced the participants
  response
• seeing sheep might lead to an expectation of a
  farm scene
• if that scene is indeed presented, then
  participants will have a bias to respond yes
• when an inconsistent scene is presented (e.g., an
  office) the participants will demand a lot of
  visual evidence (i.e., bottom-up information)
  before theyll acknowledge that a sheep was in
  the office--in other words, they will have a bias
  to respond no
• employed a procedure much like the one used by
  Reicher (1969) to investigate word-superiority
• after presentation of a scene (e.g., farm),
  participants were given a forced choice
  recognition test in which two alternatives were
  presented
• either both scene consistent (sheep, pig) or both
  scene inconsistent (coffee maker, mixer)
• participants had to pick the one that was in the
  scene

23
Hollingworth and Henderson (1998)

• predictions
• if scene context facilitates recognition, then in
  the scene-consistent trials, participants should
  successfully choose the correct option a majority
  of the time
• if scene superiority effect had been due to a
  guessing bias, then this procedure should
  eliminate the scene superiority effect
• when faced with two alternatives that both fit
  with the context, neither item will have an
  advantage--theyre both likely choices
• no scene-superiority effect was found
• participants were just as good at choosing which
  of two scene inconsistent objects had occurred as
  they were at guessing which of two scene
  consistent objects had occurred
• authors conclude that scene context does not
  facilitate the recognition of scene components
• the processes underlying object perception seem
  to be isolated from information about objects
  that might occur in any given scene
• the lack of effect of contextual knowledge on
  object recognition is probably a good thing
• if identification of objects involved consulting
  general knowledge about objects both relevant and
  irrelevant to the scene, it would get bogged
  down top-down processing would be an unnecessary
  drag on the system

24
Hollingworth and Henderson (1998), cont.

• so why does it top-down information facilitate
  the recognition of letters?
• our experience with any given word is not as
  extensive as our experience with a given natural
  scene
• the letters that can appear in a word are much
  more constrained by context (i.e., context
  provides more information about the component
  letters of a word) than are the objects that can
  appear in a scene
• Two-System Views of Object Recognition
• objects are represented in two separate
  systems--a structural description system and an
  episodic system
• the structural description system includes
  information about the global shape of an object,
  as well as the relationships among the objects
  parts
• the episodic system encodes semantic (i.e.,
  meaning) and visual information about the
  objects, such as their identity, their function,
  and specifics of their visual presentation

25
Cooper and Shepard (1992)

• encoding phase participants are asked to make a
  judgment about possible and impossible objects
• possible objects are ones whose surfaces and
  edges are configured such that they could exist
  in a 3-D world
• impossible objects are those that could not exist
  in three dimensions

• two encoding tasks
• global structure task--decide whether the object
  faced right or left
• induce participants to encode the global
  properties of the objects, and the
  interrelationships between the parts
• meaningful properties task--name something that
  each object resembled
• required a meaningful elaboration of the object
• later phase participants are given two different
  object recognition tests
• one test, they are asked whether they saw each
  figure earlier
• other test, they are asked to rapidly classify a
  series of objects as possible or impossible
• some of these objects had been presented earlier

26

• dependent variable is whether participants have
  an easier time making the possible-impossible
  judgment for figures that they had seen before
  (i.e., priming)
• benefit from having seen an object before is
  called priming
• results
• if participants had to recognize the object as
  one they had seen before, the most effective
  encoding was to think about it in meaningful
  terms (i.e., meaningful properties task)
• foreshadows levels-of-processing effect remember
  material better when its processed in terms of
  its meaning rather than its physical structure
• if participants had to rapidly classify objects
  as possible or impossible, the most effective
  encoding was to notice the global structure of
  the object (i.e., the global structure task)

27
Structural Description System

• includes a stored representation of the overall
  structure of objects and is used as the basis for
  their rapid recognition
• subserved performance on the speeded
  possible-impossible classification test
• the only encoding method that primed the
  representation in this system was the global
  structure task (which induced participants to
  notice the overall structure of objects)
• However, rapid recognition of impossible objects
  was not aided by global structure encoding task.
• Suggestion The structural description system
  cant recognize impossible objects?
• based on bottom-up processing this system
  contains the actual data that we analyze

28
Episodic System

• includes a stored representation of the identity
  of the object and its distinctive physical
  characteristics this system is the basis for our
  memory of and knowledge about objects
• subserved performance on the recognition test
  this system stores information about what objects
  are, recognition was most affected by the
  encoding task that required participants to say
  what each object resembled
• based on top-down processing this system
  includes our conceptual knowledge about the
  objects that we need to identify
• global judgments primed representations in the
  structural description system but only for
  possible objects the rapid recognition of
  impossible objects showed no benefit from the
  global structure judgment
• the structural description system is incapable of
  representing objects that could not exist in
  three dimensions
• indicates the importance of top-down processing
  in object recognition--even the structural
  description system, which provides the data for
  bottom-up processing, is influenced by our
  previous experience

29
III. Recognizing Faces

• The Face-Inversion Effect
• the deleterious effect of inversion is
  disproportionately great for faces compared to
  other objects
• to recognize objects, we need first-order
  relational information--information about the
  parts of an object, and how those parts relate to
  one another
• first-order relational information is not enough
  to recognize faces noticing that two eyes are
  above the nose, which is above the mouth may be
  enough to recognize that something is a face, but
  doesnt allow for recognition of who the face is
• to recognize faces, we need second-order
  relational information
• second-order relational information involves
  comparing the first order analysis to facial
  features of a typical or average face
• this typical face is built up through experience,
  and serves as an implicit standard against which
  we compare faces that we see
• when a face is inverted, this disrupts the
  encoding of second-order relational information
  therefore inversion disproportionately harms face
  recognition

30
Diamond and Carey (1986)

• in addition to replicating the basic inversion
  effect with human faces, they also investigated
  recognition of dog faces
• compared dog experts with dog non-experts
• dog experts are so experienced with dogs that
  they encode dog faces in terms of second-order
  relational properties
• inversion should have adverse effects on the dog
  non-experts recognition of human faces, but not
  dog faces
• inversion should have adverse effects on dog
  experts recognition of both human faces and dog
  faces
• predicted results were obtained
• Wholistic Processing of Faces
• faces are encoded, stored, and retrieved from
  memory as whole configurations, rather than as a
  set of features or parts

31
Tanaka and Farah (1993)

• to the degree that a given object is stored as a
  set of features, then those features ought to be
  useful cues in retrieving the remaining
  information about the object
• if an object is stored as a whole configuration
  then presenting part of that whole will not be
  particularly helpful in recognition
• presented participants with sketches of faces and
  sketches of houses, both decomposable in terms of
  distinct features each face and house was given
  a label, such as Larrys house, or Larrys
  face
• on a later recognition test, participants were
  asked about the faces and houses
• isolated-part condition--given a choice of two
  object parts, and had to pick which one of them
  had been part of an earlier-presented object
  (e.g., Which of these is Larrys nose? or Which
  of these is Larrys door?)
• whole-object condition--given a choice of two
  whole objects, and had to pick out the one they
  had seen earlier (e.g., Which of these is Larrys
  face? or Which of these is Larrys house?)

32
Tanaka and Farah (1993)

• results
• the type of question asked didnt matter for
  recognition of houses participants were just as
  good at recognizing parts of houses as they were
  at recognizing whole houses
• for faces, the type of question did matter
  participants were not as good at recognizing face
  parts as they were at recognizing whole faces
• face recognition is similar to view based
  approaches to pattern/object recognition

33
Dissociations and Associations in Face Recognition

• some suggest at least two subsystems for
  recognizing letters, objects, and faces rather
  than a special mechanism for face recognition
• alexia deficits in the ability to recognize
  printed words
• object agnosia deficits in the ability to
  recognize everyday objects
• prosopagnosia an inability to recognize familiar
  faces
• There are associations and dissociations among
  these three disorders
• deficits in recognizing faces are often found in
  association with at least some deficits in
  recognizing objects
• deficits in recognizing letters/words are often
  found in association with at least some deficits
  in recognizing objects
• these two findings indicate that object
  recognition does share some mechanisms with both
  letter/word and face recognition

34
Dissociations and Associations in Face Recognition

• deficits in object recognition with spared face
  and letter/word processing or the opposite
  pattern (deficits in face and letter/word
  processing with spared object recognition) are
  rarely found
• the fact that face recognition and letter/word
  recognition are almost never spared together or
  impaired together indicates that face recognition
  and letter/word recognition rely on quite
  distinct mechanisms
• the fact that problems in object recognition
  quite often line up with problems in either face
  or letter recognition indicates that object
  recognition relies on some of the mechanisms
  required for each
• visual recognition involves two primary
  mechanisms
• one mechanism used for representation/combination
  of parts
• important for letter/word recognition not
  important for face recognition
• second mechanism used for representation/combinati
  on of complex wholes
• important for face recognition not important for
  letter/word recognition
• a combination of both is important for object
  recognition