Chapter 3: Recognizing objects
Chapter 3: Recognizing objects Visual Perception -
obtain information about the world via sensory modalities vision is the dominant sense form perceptiono the process by which you manage to see what the basic shape ad size of an object are object recognition o process through which you identify what the object is
Why is object recognition crucial? -
without recognition you cannot bring you knowledge to the world for learning o usually have to combine new info with previously learned
Beyond the information given -
-
-
gestalt psychologists o perception of the visual world is organized in ways that the stimulus input is not o organization must be contributed by the perceiver and is different from the sum of its parts Jerome Bruner o “Beyond the information given” coined that phrase o our perception of a stimulus differs from the stimulus itself o Necker cube – ambiguous cause it can be perceived in more than one way Figure/ground organization- determination of what is the figure and what is the ground
Organization and Features -
our interpretation (organization of the input) happens before we start cataloguing the inputs basic features (certain features have an eye in the beholder point of view) creates a paradox: o on one hand- perception must start with the stimulus and governed by what is in that stimulus- the features must be in place before an interpretation is offered
o other hand- features one find in an input depend on how the figure is interpreted therefore it must be interpretation not features that must be first o SOLUTION: the brain relies on parallel processing where the brain areas analyzing a patterns basic features do their work at the same time as the brain areas analyzing the patterns large scale configuration Logic of perception -
what matters for familiarity is the figure as it is perceived the brain comes to the most basic answer and don’t rely on coincidences
Object recognition Recognition: Some Early Considerations + Features -
-
many objects are recognized by their parts recognition begins with identification of features such as vertical lines, curves, diagonals we recognize objects b detecting the presence of the relevant features visual search task – participants have to indicate whether a certain target is or is not present in a display detection of features is a separate step in object recognition, followed by other steps in which the features are assembled into more complex wholes Integrative Agnosia- damage to the parietal cortex, appear normal in tasks where they have to detect whether particular features are present or absent but are impaired when asked to judge how the features are bound together to form complex objects Disruption in the parietal lobe had no impact on a single feature detection (find the red shape) but a conjunction of features was very hard to do (find the shape that is red and round)
Word Recognition Factors influencing recognition - tachistoscope- device designed to present stimuli for precisely controlled amounts of time - each stimulus is followed by a mask- random jumble of letters which will disrupt any continued processing that participants might try to do - familiarity impacts the ability to recognize - recency of view impacts ability to recognize - repetition priming- exposing yourself to a word and then later that day exposing yourself to it again The word-superiority Effect
-
words viewed frequently are easier to perceive words are easier to perceive as compared to isolated letters this is called the word-superiority effect
Degrees of Well-Formedness -
letter strings like GLAKE are easier to recognize than “JPSRW” familiarlity of a letter formation doesn’t actually matter what matters is that the string is well formed according to the rules of the language pronounce ability is a major factor that helps recognition
Making Errors -
we make systematic errors strong tendency to misread less-common letter sequences as if they were more common patterns and irregular patterns are misread as regular patterns over-regularization errors- people perceiving input to be more regular than it actually is
Feature Nets and Word Recognition The design of a feature net - when different letters detectors of the feature are activated perhaps it activates different letter detectors - a network of detectors organized in layers with each subsequent layer concerned with more complex larger-scale objects - 1. feature net is the bottom layer that is concerned with features - activation level- reflect how activated the detector is at just that moment - response threshold- activation levels very high in a detector - fire- the response threshold will cause the detector to fire which sends a signal to other detectors to which it is connected - the more a detector is activated the easier it will activated (frequency, recent firing) - even a weak signal can bring the detectors that are used most often, to life The Feature Net and Well-Formedness -
3. bigram detectors- detectors of letter pairs well formed words involve familiar letter combinations
Recovery from Confusion -
if only given a word to view for a couple seconds your brain automatically assumes that the bigram you are looking for is the most common one
activated because other bigrams are never used and are therefore ruled out because they need a lot of activation power Ambiguous Inputs -
in the “the cat” example where the h and the a are the same, it still causes them to fire because they fire recently but only weak because only some of the a’s features are present
Recognition Errors -
the more regular we see words the more we make mistakes o if we see CQRN we will automatically read CORN because it is more common
Distributed Knowledge -
the detectors knowledge is not Locally represented- it is not stored in a particular location or built into a specific process distributed knowledge – it is represented in a fashion that is distributed across the network and detectable only if we consider how the entire network functions
Efficiency vs accuracy -
to maximize time efficiency we skip some letters in order to read faster
Descendants of the Feature Net The McClelland and Rumelhart Model -
the same as the other model expect it is able to accomplish the same stuff but without bigram detectors excitatory connections- activation of one detector serves to activate other detectors but in this model they can also inhibit connections higher level detectors can influence lower level detectors at any level (different from the other model)
Recognition by components Hummel and Biederman - offered a network theory dubbed recognition by components model (RBC) o intermediate level of detectors sensitive to Geons (geometric ions) o geons are an alphabet for which objects are constructed
o only 26 geons are needed to make all the shapes of objects in the world o viewpoint-independent- geons can be identified from any angle of view Recognition via multiple views Michael Tarr - different approach to object recognition - people have stored in memory a number of different view of each object and to recall the object someone just has to match what they are seeing to the image in their memory - viewpoint-dependent – needs to be seen from a certain viewpoint to be recognizable - data supports this theory because we can recall certain things at certain angles, faster than others o there is a hierarchy of detectors with each successive layer within the network concerned with more complex aspects Different objects different recognition systems? -
similar mechanisms used in things other than print face recognitions is served by specialized structures and mechanisms and proceeds according to principles different from the ones we have described so far
Faces -
-
Prosopagnosia- inability to recognize faces even though their other visual abilities seem to be intact Performance suffers from upside down photos specifically for faces and dogs This disorder is not just for faces but for cars and birds etc. We have a specialized recognition system using its own brain tissue and especially sensitive to orientation o This system works whenever a task has two characteristics: Involves recognizing specific individuals within a category Category has to be extremely familiar 2 parts to this model o system for recognizing parts and the assembly of those parts o system for configurations for analyzing patterns damage to the configuration leads to prosopagnosia damage to the fist system disrupts ability to recognize words, objects and other targets
Top Down influences on Object Recognition
Benefits of larger contexts -
more easy to understand a word if it is in a sentence than by itself or in a jumble of letters
interactive models data driven, bottem up processing – how the incoming information ttrigers a response by feature detectors which in turn triggers a sreponse by letter detectors and so on - object recognition also involved concept driven, top down processingprocessing that is diren by a broad pattern of knowlge and expectations -interactive models- involve both top down and bottem up processes
obtain information about the world via sensory modalities vision is the dominant sense form perceptiono the process by which you manage to see what the basic shape ad size of an object are object recognition o process through which you identify what the object is
Why is object recognition crucial? -
without recognition you cannot bring you knowledge to the world for learning o usually have to combine new info with previously learned
Beyond the information given -
-
-
gestalt psychologists o perception of the visual world is organized in ways that the stimulus input is not o organization must be contributed by the perceiver and is different from the sum of its parts Jerome Bruner o “Beyond the information given” coined that phrase o our perception of a stimulus differs from the stimulus itself o Necker cube – ambiguous cause it can be perceived in more than one way Figure/ground organization- determination of what is the figure and what is the ground
Organization and Features -
our interpretation (organization of the input) happens before we start cataloguing the inputs basic features (certain features have an eye in the beholder point of view) creates a paradox: o on one hand- perception must start with the stimulus and governed by what is in that stimulus- the features must be in place before an interpretation is offered
o other hand- features one find in an input depend on how the figure is interpreted therefore it must be interpretation not features that must be first o SOLUTION: the brain relies on parallel processing where the brain areas analyzing a patterns basic features do their work at the same time as the brain areas analyzing the patterns large scale configuration Logic of perception -
what matters for familiarity is the figure as it is perceived the brain comes to the most basic answer and don’t rely on coincidences
Object recognition Recognition: Some Early Considerations + Features -
-
many objects are recognized by their parts recognition begins with identification of features such as vertical lines, curves, diagonals we recognize objects b detecting the presence of the relevant features visual search task – participants have to indicate whether a certain target is or is not present in a display detection of features is a separate step in object recognition, followed by other steps in which the features are assembled into more complex wholes Integrative Agnosia- damage to the parietal cortex, appear normal in tasks where they have to detect whether particular features are present or absent but are impaired when asked to judge how the features are bound together to form complex objects Disruption in the parietal lobe had no impact on a single feature detection (find the red shape) but a conjunction of features was very hard to do (find the shape that is red and round)
Word Recognition Factors influencing recognition - tachistoscope- device designed to present stimuli for precisely controlled amounts of time - each stimulus is followed by a mask- random jumble of letters which will disrupt any continued processing that participants might try to do - familiarity impacts the ability to recognize - recency of view impacts ability to recognize - repetition priming- exposing yourself to a word and then later that day exposing yourself to it again The word-superiority Effect
-
words viewed frequently are easier to perceive words are easier to perceive as compared to isolated letters this is called the word-superiority effect
Degrees of Well-Formedness -
letter strings like GLAKE are easier to recognize than “JPSRW” familiarlity of a letter formation doesn’t actually matter what matters is that the string is well formed according to the rules of the language pronounce ability is a major factor that helps recognition
Making Errors -
we make systematic errors strong tendency to misread less-common letter sequences as if they were more common patterns and irregular patterns are misread as regular patterns over-regularization errors- people perceiving input to be more regular than it actually is
Feature Nets and Word Recognition The design of a feature net - when different letters detectors of the feature are activated perhaps it activates different letter detectors - a network of detectors organized in layers with each subsequent layer concerned with more complex larger-scale objects - 1. feature net is the bottom layer that is concerned with features - activation level- reflect how activated the detector is at just that moment - response threshold- activation levels very high in a detector - fire- the response threshold will cause the detector to fire which sends a signal to other detectors to which it is connected - the more a detector is activated the easier it will activated (frequency, recent firing) - even a weak signal can bring the detectors that are used most often, to life The Feature Net and Well-Formedness -
3. bigram detectors- detectors of letter pairs well formed words involve familiar letter combinations
Recovery from Confusion -
if only given a word to view for a couple seconds your brain automatically assumes that the bigram you are looking for is the most common one
activated because other bigrams are never used and are therefore ruled out because they need a lot of activation power Ambiguous Inputs -
in the “the cat” example where the h and the a are the same, it still causes them to fire because they fire recently but only weak because only some of the a’s features are present
Recognition Errors -
the more regular we see words the more we make mistakes o if we see CQRN we will automatically read CORN because it is more common
Distributed Knowledge -
the detectors knowledge is not Locally represented- it is not stored in a particular location or built into a specific process distributed knowledge – it is represented in a fashion that is distributed across the network and detectable only if we consider how the entire network functions
Efficiency vs accuracy -
to maximize time efficiency we skip some letters in order to read faster
Descendants of the Feature Net The McClelland and Rumelhart Model -
the same as the other model expect it is able to accomplish the same stuff but without bigram detectors excitatory connections- activation of one detector serves to activate other detectors but in this model they can also inhibit connections higher level detectors can influence lower level detectors at any level (different from the other model)
Recognition by components Hummel and Biederman - offered a network theory dubbed recognition by components model (RBC) o intermediate level of detectors sensitive to Geons (geometric ions) o geons are an alphabet for which objects are constructed
o only 26 geons are needed to make all the shapes of objects in the world o viewpoint-independent- geons can be identified from any angle of view Recognition via multiple views Michael Tarr - different approach to object recognition - people have stored in memory a number of different view of each object and to recall the object someone just has to match what they are seeing to the image in their memory - viewpoint-dependent – needs to be seen from a certain viewpoint to be recognizable - data supports this theory because we can recall certain things at certain angles, faster than others o there is a hierarchy of detectors with each successive layer within the network concerned with more complex aspects Different objects different recognition systems? -
similar mechanisms used in things other than print face recognitions is served by specialized structures and mechanisms and proceeds according to principles different from the ones we have described so far
Faces -
-
Prosopagnosia- inability to recognize faces even though their other visual abilities seem to be intact Performance suffers from upside down photos specifically for faces and dogs This disorder is not just for faces but for cars and birds etc. We have a specialized recognition system using its own brain tissue and especially sensitive to orientation o This system works whenever a task has two characteristics: Involves recognizing specific individuals within a category Category has to be extremely familiar 2 parts to this model o system for recognizing parts and the assembly of those parts o system for configurations for analyzing patterns damage to the configuration leads to prosopagnosia damage to the fist system disrupts ability to recognize words, objects and other targets
Top Down influences on Object Recognition
Benefits of larger contexts -
more easy to understand a word if it is in a sentence than by itself or in a jumble of letters
interactive models data driven, bottem up processing – how the incoming information ttrigers a response by feature detectors which in turn triggers a sreponse by letter detectors and so on - object recognition also involved concept driven, top down processingprocessing that is diren by a broad pattern of knowlge and expectations -interactive models- involve both top down and bottem up processes
