Pattern recognition (…and object perception)
Traditional Gestalt Psychology Topic Outline: Traditional Gestalt Psychology Laws of Organization Problems with this stuff Solution to Pragnanz II. Identifying “Figures” Properties of figures/ground Intrinsic/extrinsic contours Illusory contours III. Models of object recognition a. Pandemonium Theory b. Identification by Components c. Marr’s Computational approach d. Gibson’s “direct perception” e. Connectionist/PDP models
Traditional Gestalt Psychology -originated as a counter to structuralist viewpoint (Koffka, Wertheimer, Kohler) Structuralist: “the whole is equal to the sum of the parts” Koffka: “the whole is other than the sum of the parts” In other words, cannot fully explain object perception based solely on a description of the elements of that perception: emergent properties One of the first examples: Phi phenomenon
Emergent properties
Emergent properties
Emergent properties
Emergent properties
Emergent properties
Emergent properties
Emergent properties
Emergent properties
Emergent properties
Gestalt laws of Organization
Gestalt laws of Organization Law of Proximity
Gestalt laws of Organization Law of Similarity
Gestalt laws of Organization Good Continuation
Gestalt laws of Organization Closure
Gestalt laws of Organization Common fate
Gestalt laws of Organization Common fate
Gestalt laws of Organization Common fate
Gestalt laws of Organization Common fate
Gestalt laws of Organization Common fate
Gestalt laws of Organization Common fate
Gestalt laws of Organization Common fate
Gestalt laws of Organization All the other laws are subsumed under the Law of Pragnanz: -patterns will be perceived in such a way that the resulting structure is as simple as possible Problems: 1. Descriptive, with no indication of underlying mechanisms of how it is accomplished 2. Lack of predictive power
Gestalt laws of Organization All the other laws are subsumed under the Law of Pragnanz: -patterns will be perceived in such a way that the resulting structure is as simple as possible Problems: 1. Descriptive, with no indication of underlying mechanisms of how it is accomplished 2. Lack of predictive power 3. No definition of “simplest”
More recent research tries to clear up that last one: “simplest” = least amount of data/information required 2-d = 26 angles 3-d = 24 angles
More recent research tries to clear up that last one: “simplest” = least amount of data/information required 2-d = 18 angles 3-d = 24 angles
II. Identifying “Figures”
II. Identifying “Figures” Relative to “grounds”, figures are: -in front -a clearly defined shape -higher perceived contrast -processed in greater detail -the process involves distinguishing intrinsic from extrinsic contours
-can give rise to complex perceptions II. Identifying “Figures” -subjective/illusory contours: -can give rise to complex perceptions
-can give rise to complex perceptions II. Identifying “Figures” -subjective/illusory contours: -can give rise to complex perceptions
-can give rise to complex perceptions II. Identifying “Figures” -subjective/illusory contours: -can give rise to complex perceptions
III. Models of object recognition -bottom-up versus top-down distinction -earliest model: template matching -discarded due to lack of flexibility, too much storage needed
III. Models of object recognition 1. Selfridge’s Pandemonium model
III. Models of object recognition 1. Selfridge’s Pandemonium model
III. Models of object recognition 1. Selfridge’s Pandemonium model Problems: Defining a feature Overly data-driven (no room for context effects) Decision demon will need stored templates to compare to and recognize the object Was originally designed for language only
III. Models of object recognition 1. Selfridge’s Pandemonium model 2. Biederman’s identification by components model -another feature model, but the features are now “Geons” -small set of shapes that, in combination, can represent objects
III. Models of object recognition 1. Selfridge’s Pandemonium model 2. Biederman’s identification by components model -again, not much room for context to play a role -again, the job of actually identifying the object is left vague
III. Models of object recognition 1. Selfridge’s Pandemonium model 2. Biederman’s identification by components model 3. Marr’s computational approach -stages of processing from simple to complex -primal sketch: edges and boundaries identified based on intensity differences (called zero crossings) -2.5d sketch: that info is processed using grouping rules into larger groups, surfaces and distances interpolated into image (wire frame) -3d sketch: textures added and volume applied
III. Models of object recognition 1. Selfridge’s Pandemonium model 2. Biederman’s identification by components model 3. Marr’s computational approach 4. Gibson’s direct perception -question is, how do we perceive a constantly changing stimulus array with any consistency? -light array contains sufficient information that we don’t need internal models or deconstructions -changing array still has invariant properties -affordances can be directly perceived -deletion/accretion, optic flow, motion parallax -not enough detail to be taken seriously
III. Models of object recognition 1. Selfridge’s Pandemonium model 2. Biederman’s identification by components model 3. Marr’s computational approach 4. Gibson’s direct perception 5. Connectionist/PDP models -neural networks modeled via interconnected nodes -combinations of nodes converge to give rise to complex codings being programmed at higher layers
Different from Selfridge in several ways: -connections can occur within a single level -connections can be excitatory or inhibitory -system can learn through back-propogation and application of delta rule -apply input, assign initial weights based on nodes firing at the same time (Hebb rule) -output compared with desired input to compute the degree of error -this difference is used to compute a delta rule for changing the weights -reapply the stimulus and recalculate delta, update weights again -keep cycling through until detection is accurate
Context effects
No matter how hungry you are, try not to….
Is there a sink in the picture?
Is there a sink in the picture?
Duncker Wheel
Duncker Wheel
McGurk Effect
Word superiority effect GZQRM XXXXX What were the letters?
Word superiority effect WORKS XXXXXX