05b Goals and Objectives PDF

Preview

Summary

Unit 5 lecture notes for quiz...

Description

Module 05b Perception II Goals and Objectives 1. With respect to word recognition, what are the frequency, recency, word superiority, and pronounceability effects? How is each explained within the context of the feature net model? Pay particular attention to which type of detector plays the key role in explaining each effect. I. Determining what pieces and fragments of the retinal image belong to the same object a. Perceptual organization II. Determining what those objects are b. Objection recognition 2. What are the different layers in the feature net model? How does activation at one layer depend upon activation at the lower layer? How in general do we recognize words within the model? A word is recognized (or read) when its total activation exceeds some threshold level. The word that we recognize is the one that corresponds to the first word detector that exceeds its threshold. The activation level of a detector (at any level in the diagram) can vary depending on how much evidence there is for its feature/letter/bigram/word in the stimulus—the more evidence the higher the activation level. The more highly When a detector is activated, once the stimulus that activated the detector is removed, the activation level of the detector does not return all at once to its resting state. Rather, it strength decays gradually, eventually returning to the resting state. This decay occurs over a time interval of several seconds. •

Current activation level is determined by – Word’s frequency of occurrence Recency with which the word has been seen

• •

• •

Frequency effect: common words are easier to recognize than rarer words. A word’s resting level of activation is assumed to be higher the more frequently that word is seen Priming: more recently read words are easier to recognize than words we have not seen for awhile. If we see the word again before it has returned all the way to its resting level, then less additional activation is required Word superiority effect: A letter is easier to recognize when presented in the context of a word than alone or in the context of a non-word Pronounceability effect: A letter is easier to recognize in a pronounceable non-word than in a non-pronounceable non-word (Bigram)

3. What were the two important additions made by McClelland and Rumelhart to the feature net model? Inhibition acts between what units? What determines how much one unit inhibits another? What are the new excitatory connections added by McClelland and Rumelhart? What level of detector is eliminated in the McClelland and Rumelhart model?

Inhibitory connections between nodes at the same level a. Words inhibit other words b. Letters inhibit other letters c. Amount of inhibition is proportional to the level of activation Excitatory connections from the word to the letter level and from the letter the feature level Note the absence of bigram detectors 4. What are the fundamental assumptions of Biederman’s RBC theory? What is a geon? What are the desirable properties of geons? What is a main goal of the theory? What does Biederman mean by the Principle of Componential Recovery? What role does the relation between geons play in the theory? How is an object recognized according to RBC theory? How do we mentally represent the visual appearance of an object according to RBC? What are some of the short-comings of the theory? 5. Be familiar with the Biederman and Gerhardstein experiment that we discussed in class on the effects of orientation and perspective in terms of what they did, what they found, and how they interpreted those findings. You should also know the Tarr and Pinker study we discussed in class as a reply to the Biederman and Gerhardstein study. What did they do, why, and what did they find? What do structural description theories predict they would find? What do image based theories predict they would find? Which study is the more ecologically valid? Can we say that one study is better than the other? Is being ecologically valid necessarily a good thing? What was the purpose of Joliceour’s experiment? What did he do? What did he find? What do structural description theories predict he would find? What do image based theories predict he would find? What ambiguity did Tarr and Pinker see in the interpretation of the Joliceour study? How did they correct that in their later study? What aspects (if any) of Tarr and Pinker’s second study support image-based models and what parts of their data (if any) support structural description (aka Biederman) theories, or are at least inconsistent with image-based views? 6. Image-based (aka view-based) models are the alternative to Biederman’s RBC approach. What are their main claims? How does object recognition proceed within those models? How are they different from Biederman’s approach? How do we mentally represent the visual appearance of an object according to these theories? 7. What is a possible reconciliation between RBC and the view-based approach?

Alternatives to feature theory It’s important to realize that template models are not models of reading words; they are models of letter identification. They do not replace the entire feature net model. Instead they replace only the first level (the level of features) and it’s connections with the second level. Biederman’s Recognition by Components (RBC) Theory • Fundamental assumptions: – There exist a small set of primitive, 3-D features, called geons (Geometrical Ions) • 24 – 36 geons – 3-D objects are defined by which geons comprise them and how they are put together/their relationship to one another

• • •

– A geon can occur in different sizes, orientations, and proportions from one object to another So, to recognize objects, we need to have stored in our long term memory for each object, a) a list of the geons comprising that object b) information about how those geons connect to one another for that object, that is, how the geons are arranged with respect to one another.

The principle of componential recovery says that a. That if I can recognize all the geons in an object and how they connect to one another, then I can recognize the object

Account for View Invariance-View invariance refers to the fact that people can quite easily recognize an object from many different perspectives or views Geons: Desirable Properties• View invariance: geon can be identified when viewed from a number of different perspectives • Discriminability: each geon can be distinguished from all others when viewed from a number of different perspectives • Resistant to Noise: geons can be identified even when partially obscured – Particularly when relation among geons preserved

One final point. Biederman argues that if one can identify each geon in an object as well as its structural relation to the other geons, then one has sufficient information to identify the object. This claim has become known as the principle of componential recovery.

Evidence for RBC and Geons1. We can generally parse even unfamiliar objects into their parts and we generally agrees as to what those parts are 2. Contour deletion experiments we deleted the contours where one geon joins another geon. That makes the object relatively difficult to identify. Shortcomings of RBC• RBC allows for the recognition of broad classes of objects, but not for discriminating two objects within the same class Biederman & Gerhardstein: View InvarianceThe important results concern the effects of rotation in phase 2 of the experiment. Rotation had no statistically significant effect. Although there is perhaps a hint of an increase as the rotation increases from 0 degrees to 67.5, there is none at all between 67.5 and 135 degrees. (This last

result is especially important, because the competing view-based models—see several slides further on—would predict that the greater the degree of rotation the longer reaction times would be.) In brief, Biederman and Gerhardstein’s result is that the amount of priming (improvement in processing upon seeing a repeated stimulus) is not affected by rotation.

Naming times would be flat as a function of the amount of rotation, regardless of whether the objects were different examples of the same object or identical how do we represent objects in our long term memory? a. As a single, viewpoint independent description of the components comprising the object and a specification of how those components are connected to one another b.

supports structural description

Tarr and Pinker criticized Biederman and Gerhardstein’s experiment for having used common objects. In the second phase of the experiment, they presented the same objects but now at various rotations and measured how long it took people to name them. Reaction times increased as the amount of rotation Tarr and Pinker advocated a class of models referred to as image-based or view-based models that stand in contrast to structural description theories. According to view-based models, we learn object descriptions by storing images of actual objects seen in the real-world.; they are view point dependent. Through experience, I build up many different object descriptions of chairs, and since I have seen chairs from a number of different perspectives, Should I encounter a chair at a novel orientation, though, I will need to rotate a mental image of it before it can match one of my object descriptions for chair and hence before I can recognize it as a chair. This rotation takes time, and the more rotation I have to do the longer the time. We should note that the images these theories assume that we store of objects are incredibly detailed, and include information about color and texture as well as just the component parts. They are more like photographs than abstract descriptions in terms of geons and their spatial arrangements. We should also note that the images are treated as wholes, as unitary objects. They are not broken down into a list of geons and their arrangement.

Reaction times were longer the farther away the new orientation was from the nearest practice orientation

Jolicouer took the logic of Tarr and Pinker one step farther. Presumably, after seeing one of these objects at a new orientation, people store an image of it at that orientation. Hence, the next time they see it, they can use that image to recognize the object. There is no need for them to do any mental rotation. Consequently, they should no longer show an increase in reaction time relative to the origin orientation when they again see the object at the new orientation. In a somewhat similar experiment to Tarr and Pinker’s, also using novel objects, Jolicouer first trained participants on names for novel objects In round 2, participants again named the objects presented at various rotations. All these rotations were rotations used in the first round. Now, there was no effect of orientation. Jolicouer argued that was because participants learned a new object description (stored a new image) for each object when they saw it in a new orientation during the round 1 of testing. This explanation is consistent with view-based models, Tarr and Pinker, in a later experiment, did suggest an alternative explanation for Jolicouer’s results. Perhaps, after seeing the same object in several orientations, participants developed a view-independent or view-invariant description for the object (like those hypothesized by RBC theory), stored that description in long-term memory, and used that to identify the object on future presentations. What they are suggesting is that perhaps people really do develop viewpoint independent representations (as Biederman would argue, and contrary to Tarr and Pinker’s image-based model). But in order to develop that representation, they first have to see the object from multiple perspectives—seeing it at just one rotation is not sufficient.

person needs to see an object from multiple viewpoints in order to develop an viewpoint independent representation of that object...