Psy 260 Announcements All late CogLab Assignment #1’s due today CogLab #2 (Attention) is due Thurs. 9/21 at the beginning of class Coglab booklets and disks--along with a printer that usually works--are available for use in the Psychology Resource Room (enter through Psych B 120) Quiz alert! Neural network models Nodes - processing units used to abstractly represent elements such as features, letters, and words Links, or connections between nodes Activation - excitation or inhibition that spreads from one node to another Word superiority effect, revisited Word superiority effect, revisited Cond. 1: WORD XXXXX Cond. 2: ORWD Cond. 3: D XXXXX XXXXX Test: Which one did you see? K K D D K D Word superiority effect, revisited Word level Letter level Feature level See Reed, p. 36 Input Word superiority effect: An interactive activation model WORK K | / \ See Reed, p. 36 Input: K or WORK or ORWD Interactive Activation Model of the word superiority effect (McClelland & Rumelhart, 1981) Interactive Activation Model of the word superiority effect (McClelland & Rumelhart, 1981) (Email example of mangled text!!) James Cattell, 1886: Word superiority effect (Reicher, 1969; Cattell, 1886) Subjects recognized flashed words more accurately than flashed letters. He proposed a word shape model. Evidence for word shape model: Word superiority effect Lowercase text is read faster than uppercase. Proofreading errors tend to be consistent with word shape. Evidence for word shape model: Word superiority effect Lowercase text is read faster than uppercase. Proofreading errors tend to be consistent with word shape. It’S dIfFiCuLt To ReAd WoRdS iN aLtErNaTiNg CaSe. Perception and Pattern Recognition III: Faces How do people recognize faces? Consider these types of theories: Template theories Feature theories Structure theories Prototype theories Feature theories Patterns are represented in memory by their parts. In perception, the parts are first recognized and then assembled into a meaningful pattern. Piecemeal (as opposed to holistic) What are the distinctive features for faces ? Eyes, nose, mouth - NOT! What are the distinctive features for faces ? Eyes, nose, mouth - NOT! Revisit Eleanor Gibson’s criteria: Each feature should be present in some patterns and absent in others A feature should be invariant (unchanged) for all instances of a particular pattern Each pattern has a unique combination of features The number of features should be fairly small A set of features is evaluated by how well it can predict perceptual confusions. Who are these people? Same or different? Who are these people? Same or different? Inspiration: Caricatures “More like the face than the face itself” What are the distinctive features of a face - say, Richard Nixon’s??? Ski jump nose Jowly face Curly-textured hair Receding bays in hairline Boxy chin (David Perkins, 1975) Contraindicated features: Worse than missing features (Perkins, 1975) A D B C E F QuickTime™ and a TIFF (LZW) decompressor are needed to see this picture. Revisit: Problems w/ feature theories How to determine the right set of features? What about the relationships between features? What if all the features are present in the pattern, but scrambled? Features theories predict: No problem! (and that’s the problem.) Face recognition is holistic (Tanaka & Farah, 1993) Structure theories Build on feature theories Patterns are represented in memory by features AND by the relations between them. Holistic The context of the pattern plays an important role in pattern recognition. A structure theory: RBC (Biederman) Recognition by Components Geons: simple volumes (~35 of them) Construct objects by combining geons RBC Theory Analyze an object into geons Determine relations among the geons The relation among geons is critical! RBC Theory It’s hard to recognize an object without the information about relations among geons. Hard! RBC Theory It’s hard to recognize an object without the information about relations among geons. Easier! RBC Theory Basic properties of Geons View invariance Discriminability Resistance to visual noise RBC Theory - Problems Explains how people distinguish categories of objects (types) - like cups vs. briefcases. But how do people distinguish individual objects (tokens) that come from the same category (like faces)?? Neurons are to tuned respond to much smaller elements than those represented by geons! Recap so far: Theory: What it explains: Template Feature Structural Prototype Bar codes (by machines) Letter learning & confusions Biederman’s data (geons) Face recognition (Piecemeal or holistic?) (A “special” case of pattern recognition?) We see faces everywhere. Image from Mars’ surface by Viking Orbiter 1 (Mcneill, 1998, p. 5) Are faces “special”? How many faces can you recognize? Are faces “special”? How many faces can you recognize? Gibson: Patterns are easier to encode as faces than as writing Are faces “special”? How many faces can you recognize? Gibson: Patterns are easier to encode as faces than as writing Faces vs. writing Are faces “special”? How many faces can you recognize? Gibson: Patterns are easier to encode as faces than as writing Prosopagnosia We don’t need much information to recognize a familiar face. Guess who? We don’t need much information to recognize a familiar face. Guess who? Why is face recognition so interesting? It’s important! Faces are highly similar to one another. Yet we’re really good at it: we can tell an astounding number of faces apart. Not all facial information is created equal. Could machines ever do as well as people? Or even better? Are faces somehow “special”? Why is face recognition so interesting? It’s important! Faces are highly similar to one another. Yet we’re really good at it: we can tell an astounding number of faces apart. Not all facial information is created equal. Could machines ever do as well as people? Or even better? Are faces somehow “special”? Faces are hard to recognize in photographic negative (Galper & Hochberg, 1971) Faces are hard to recognize upside down (Yin, 1969) Faces are hard to recognize upside down (Yin, 1969) “Early processing in the recognition of faces” http://www.diss.fu-berlin.de/2003/35/Kap4.pdf Faces are hard to recognize upside down (Yin, 1969) “Early processing in the recognition of faces” http://www.diss.fu-berlin.de/2003/35/Kap4.pdf Margaret Thatcher effect (Thomson, 1980) Margaret Thatcher effect (Thomson, 1980) Why? The configural processing hypothesis: When faces are inverted, the relationships among features are disturbed. So we don’t notice the odd configuration in the Thatcher illusion. (Bartlett & Searcy, 1993) Faces are hard to recognize upside down (Yin, 1969) “Early processing in the recognition of faces” http://www.diss.fu-berlin.de/2003/35/Kap4.pdf What kind of theory accounts for face recognition? Theory: Objection: Template Different lighting, orientation, motion, hair, glasses, age What is a facial “feature”? Invariant vs. transient features Feature Structural Prototype Familiar vs. unfamiliar faces “Attribute Checking Theory” A feature theory For familiar faces, internal features seem to be more important than outside features. For new faces, we pay more attention to outside features (hair, face shape, etc.) (Bradshaw & Wallace) Familiar vs. unfamiliar faces “Early processing in the recognition of faces” http://www.diss.fu-berlin.de/2003/35/Kap3.pdf Children recognize faces differently than adults do. Children under 10 use transient features to distinguish unfamiliar faces. Strangers wearing the same hat seem similar, and are confusable. (Susan Carey) What makes faces confusable? (Harmon, 1973) Application: Face recognition by eyewitnesses Problem: Identikit: piecemeal, featural Photo methods: Introduce interference, bias Lineup: when the perpetrator is not present, 20-40% of witnesses select someone anyway. With photos and lineups, witnesses compare the suspects and choose the most similar one False convictions often have eyewitness testimony as the strongest evidence in the The right way to do a lineup: “Showup” - view suspects or pictures one at a time, ideally only once If multiple viewings, then view each one the same number of times, always in random order (avoid between-suspect comparisons) The one showing the faces must be blind to whom law enforcement believes suspect is (Otherwise, impossible to avoid bias) Then false IDs drop to 10%. Mistaken identity! What about a structural theory of face recognition? Pro: The relationships between features are very important. Pro: We often fail to recognize a familiar face when we see it out of context. Con: A structural theory doesn’t explain how we can distinguish so many highly similar, individual tokens. (Moving right along: A prototype theory What is a caricature? An exaggerated representation of a face More like a face than the face itself! The Caricature Generator (Brennan, 1982) The average (prototype) face QuickTime™ and a TIFF (LZW) decompressor are needed to see this picture. Veridical (traced) drawing QuickTime™ and a TIFF (LZW) decompressor are needed to see this picture. Veridical (traced) drawing QuickTime™ and a TIFF (LZW) decompressor are needed to see this picture. Ronald Reagan A prototype theory of face recognition When drawings were recognized, caricatures were faster than veridical drawings, which were faster than “anti-caricatures.” QuickTime™ and a TIFF (LZW) decompressor are needed to see this picture. QuickTime™ and a TIFF (LZW) decompressor are needed to see this picture. Average face 0 distortion Caricature (Rhodes, Brennan, & Carey, 1987) QuickTime™ and a TIFF (LZW) decompressor are needed to see this picture. 50% Caricature Caricatures & Anti-Caricatures For a face, maybe we encode the difference from a prototype. Face Space What kind of theory accounts for face recognition? Theory: Objection: Template Different lighting, orientation, motion, hair, glasses, age What is a facial “feature”? Invariant vs. transient features Faces are highly similar tokens with the same structure! This works! (but maybe not for unfamiliar faces and not for kids) Feature Structural Prototype Is face recognition “special”? No! There are other classes of patterns for which people can distinguish huge numbers of individuals (tokens). Ornithologists recognize individual birds New England Kennel Club judges recognize individual dogs There is even prosopagnosia for things other than faces! Some sources George Lovell’s slides from Roth & Bruce http://www.face-rec.org/interesting-papers/Other/FaceRecognition.pdf “Early processing in the recognition of faces” http://www.diss.fu-berlin.de/2003/35/Kap3.pdf Harmon, L. D. (1973). The recognition of faces. Scientific American, 229(5), 71-82.