Cognitive Psychology Lecture 6: Object Recognition PDF

Lecture 6: Object (& Word) Recognition Cognitive Psychology 2135 These are words  Complex cells: Hyper-complex: Simple cells: large receptive very large How do we narrow, fields that receptive fields assemble these elongated...

Lecture 6: Object (& Word) Recognition Cognitive Psychology 2135 These are words  Complex cells: Hyper-complex: Simple cells: large receptive very large How do we narrow, fields that receptive fields assemble these elongated combine that combine visual "features" to receptive fields: information information from recognize whole "line detectors" from simple complex cells. objects? cells. Respond Like edges with best to moving end points. edges in a "angle specific detectors" orientation & direction of motion  High level Vision How do humans (other animals) recognize objects in the world? Bottom-up processes - Data driven that depend on visual input Top-down processes - Conceptually driven processes that depend on knowledge Top-down Processing  Feature detection Objects - composed of separable, distinct features Recognition - decomposing objects into features and re-assembling them into representations of objects - comparing to representations in memory  Features What is a feature? - visual primitives (edges, colours, etc.) - innate feature detectors and receptive fields  Features Whole object is decomposed into its features - Feature detectors (innate and learned) are activated - Features are then integrated (assembled) - Feature integration takes time - Studied with a feature detection task Find the vertical line Find the vertical line Target detection not dependent on number of distractors Orientation is a single feature Find the red line Find the red line Target detection is not sensitive to the number of distractors “Red” is a single feature Find the vertical red line Identification of feature conjunctions takes longer  Recognition By Components Also known as Geon Theory (Irving Biederman) Objects based on 36 basic shapes: Geons  Recognition By Components Objects are composed of geons Geons are identified by low-level vision They are composed of non-accidental properties  Tests of Geon Theory Outlines are recognized as quickly as pictures - Prediction: If geons underlie object recognition, then simple outlines should be just as easy to identify as actual objects - Biederman & Ju (1988) - brief presentations (50 ms): line drawings & colour photos - time to name object did not differ  Geon Identification Information about geons is carried in object conjunctions - Non-accidental property information is identifiable Prediction: Missing information from conjunction will impair identification Middle column: "recoverable": gaps are within geons Righthand column "nonrecoverable": gaps are at conjunctions of geons  Problems... Structural theories (like geon theory) do not specify how objects with the same geons are identified - e.g., dog vs. wolf Many objects have no identifiable geons (loaf of bread) - Top-down knowledge is needed  Word Superiority (role of top-down information in recognizing letters alone vs. words) Do you see a “E” or a “K” D K V E Word Superiority Do you see a “E” or a “K” ZDVK GPVR WQGR WORK Feature Net  Feature Net Each detector in the network has an activation level - With input, this activation level increases Some detectors will be easier to activate than others Detectors fire when their response threshold is reached - Similar to a neuron's threshold for firing an action potential Individual detectors are complex assemblies of neural tissue, not individual neurons or groups of neurons  Feature Net Frequency: detectors that have fired frequently will have higher "resting" activation levels - An exercise effect (learning) Recency: detectors that have fired recently will have higher activation levels (not fully deactivated) - A warm-up effect ("priming")  Recognition Errors recognizing "CQRN" as "CORN" because corn is a familiar word  Distributed Knowledge Knowledge is not locally represented - Rather, feature nets contain distributed knowledge Knowledge in a network is reflected by relationships across detectors McClelland & Rumelhart Model Information flows bottom-up, top- down, and within the same level Higher-level word detectors can influence lower-level detectors Includes excitatory connections and inhibitory connections  Present TRIP very very briefly you "see" only RIP and a bit of the T activates letter detectors for R, I, & P (& a bit for T) those activate word detector for "TRIP" TRIP word detector inhibits other word detectors also activates letter detector for T you perceive word as "TRIP"  Face recognition Holistic (wholes matter more than parts) - faces are not (entirely) recognized by components - Memory for faces versus objects (houses) - If recognition is holistic then the wrong orientation would hurt performance on faces but not houses Now a memory test: Now a memory test: Now a memory test: Now a memory test:  Disruptions of Perception Study of patients with acquired (or developmental) brain damage Emphasis on the preserved cognitive abilities vs. deficits Agnosia - Apperceptive - Associative - Prosopagnosia Patients with Apperceptive Agnosia can perceive an object's features but have difficulty recognizing the object in its entirety Patient D.F.  Associative Agnosia fail to identify objects presented visually & to name visually- presented objects semantics (meaning) intact given testing in another modality (reading, hearing) The problem is the activation of semantics from visual stimulus  Prosopagnosia A visual agnosia specific to faces These patients may have otherwise intact object recognition - They can see details of faces, but cannot recognize a face as a coherent unit - Often fail to recognize their friends, family members, well known faces and even their own face - http://www.youtube.com/watch?v=VKa- PuJCrO4&feature=related  Quiz 1: Monday, January 27 15 m/c questions in a random order Open 5:00pm – 5:00pm (24 hours) Once you begin, you have 20 minutes We'll send out an email with further information