- TRACE- o Information feeds into this network from the bottom (feature level)o Information spreads through the networko Activation spreads from those nodeso Node- simple threshold unito Every word you know is competing to be recognized, depending on level of activation it will be stronger in competitiono If stays below threshold nothing happenso If threshold is reached its sent outo Connections at same layer and betweeno Primitive features combine to make letters, which combine to make wordso Auditory version looks the same but represents different thingso Instead of simple primitive visual features, its auditoryo Vertical connections: interlevel connectionso Horizontal connection: intralevel connections, inhibitory o Each letter has nested under it all the features that make up that lettero Activate all features and words of letters, 10/2/12- Settling process- network moves from this sytem to stable pattern of activation where 1 word is active, those letters and features are active, everything else is shutdowno Fully above threshold- Parameters- variables in the model that can be allowed to freely vary or fixed to a certain value- Always an issue on numbers of parameters- Degrees of freedom- if model has more freely varying data there can be a problemo Fitting data to theory is then an exercise of mathematics- Parameters allowed to vary are ones governing processes of things you are manipulating- Human world recognition is not perfect- Each word will be getting activation from letters that make up that word which- As soon as input is fed, it will be biased to words similar because features have slight advantage over all other features- There’s a time dimension- trace is not only processing this info, but in sequential moments in time- Look at location information of word, information from beginning is much more valuable/useful then info from end of the wordCohort theory- when a word is being processed, a cohort is selected (a group of words you know)o Word initial cohort theoryo All words that start with that sound activateo Rest of the info in that word is used to eliminate o Before second features come in, the first features are inhibiting the next featureso Earliness effect- trace is effected more by information early in the word, initial info starts the network running, as soon as it starts running inhibition is active as the rest of the info from the word comes in- Certain types of information is advantaged- Where the constraint comes from is going to have an enormous effect on what thesystem will do o Direct constraint- constraint at same levelo Indirect constraint- activating others- Letters that appear by themselves vs. in context of word- Will trace settle on letter by itself or in word faster?10/4/12- Trace was not designed to handle world superiority- Letters are recognized more quickly in word then by itself- How fast network can settle- Way to speed up word recognition is to construct a situation where there is more constraint- Less constraint if read backwards because you reversed the letters- More constraint in network when word is present, slight advantage in direct constraint- Featureswords using computational network architecture- Neural nets look like trace, but is radically different- Trace handles earliness, word recognition, word superiorityBeiderman’s RBC Theory- Recognition by component- Goal to come up with theory that can handle object perception ina rich dynamic world- Theories can’t handle this level of complexity - Set of components and we build our objects from them- Two stages of object recognitiono Identification of boundarieso Identification of basic shapes- Identification of boundarieso Edges and regionso Invariant properties- Geons- alphabet of components, simple 3 dimensional geometric shapeso Genons make up objects when combinedo Relationships among geons mattero- Complex objects- Object identification- Datao Object completenesso Object complexityo Degraded objects- Weaknesseso Face recognitiono Recognition of dynamic events10/9/12- Human visual perception starts with edge detection- Biederman attempts to solve the problem that we can recognize objects by only seeing one edge of the object- Visual input from environment changes with switch of viewing angle- Invariant propertieso Smooth continuation- can detect it is all one line or edge, no matter viewing angleo Cotermination- can determine 2 lines meet, junction between themo Parallelism- can tell 2 lines are parallelo Symmetry- always symmetric- Humans can detect these 4 properties- Only information used in invariant properties of edges- not all the information present at the retina - Regions are spaces separated by an edge- Edges-invariant properties-regions-geons- Edges that inclose regions are boundaries- Combine geons to construct objects- Have to access this information in memory- The way information about objects in the world are stored in memory is in ordered list of geons that make up object and structural relations- Written in a language called: lisp, list based language- Biederman built the original of this in lispo Idea was that you would match geons in parallel against all the lists stored in memoryo RBC has been reprogrammed in many ways now as a neural net, not dependent as list orientation anymore- Object completeness and complexity research is based on the assumption that objects are massively redundant, can identify an object based on relatively small portion - Pictures shown to subjects- number of geons present in object manipulatedo All thereo Remove 5% or 10% etco If object is presented without full geons, it is incomplete object- Removal of geons should not have much of an effect10/11/12- Human visual system will detect the shape- Object completion experiment- can remove a lot of the geons and does not effect it- Independent variable- number of geons removed from the object- Object complexity- the more complex objects are, the easier it is to identify- Independent variable- number of geons that make up the objecto Most of the competitors for RBC are variances of tracing theorieso Trace- mechanisms systems use to recognize objects trace outline of edges- Degraded objects- edges are detected, invariant properties of edges turn them into boundaries inclosing regions, regions create geons which make objectso Degrade objects by removing some of the information that is present, erase lineso If system is relying on these propertieso Remove same amount of