Here are some assignment questions with answers by students, as selected by the course staff.
How might people be able think logically, despite the difficulties raised by Wason? What does your answer indicate about the place of logic in cognitive science?
A: Wason's Selective Experiments made many psychologists skeptical about mental logic and its psychological plausibility. He demonstrated that most people apply modus ponens in the letter-and-number experiments. However, many people neglect modus tollens and some even choose cards that are irrelevant. The experiment pointed out that people approach certain kinds of reasoning tasks with representations and computations that are different from those used in formal logic. However, subsequent experiments shown that people have little difficulty in solving the same problem if they are given familar concrete examples, like the bar-and-age problem. People may still be able to think logically despite difficulties raised by Wason since mental logic can give an appropriate account of some narrow kinds of human reasoning, such as applying modus ponens. However, more vivid representations such as pragmatic reasoning schemas and mental models are needed to account for more complex kinds of human reasoning. Using pragmatic reasoning schemas, people do much better with the concrete bar-and-age problem because the permission schema (If one is to do X, then one must satisfy precondition Y) is naturally applied in this case but not in the card-and-number problem. In terms of mental models, Johnson-Laird and Byrne claimed that people construct a mental representation when interpreting a conditional, they explained many people's performance in the selection task by supposing that people consider only those conditions that are explicitly represented in their models of rule. The theory of mental models apply reasoning with quantifiers that first proceed using inference rules, then use propositional rules of inference to make inferences, and finally reapply quantifiers using additional rules of inference. The point raised here indicates that formal logic is an important part of cognitive science and human reasoning, but logical approach is not the only possible way to understanding human thinking. Therefore, cognitive science should also pursue other approaches besides formal logic. Nonetheless, logic in philosophy and AI provide a mathematical analysis of what constitutes optimal reasoning in developing formal logical models of how people and intelligent system 'should' think.
What makes cognition nonmonotonic? Use rules for examples.
A: Cognition is nonmonotonic because it involves not only the accumulation of new information, but also the modification or abandonment of old representations. For example, if I go to a certain barbershop and receive an excellent haircut, I may form a rule which says, IF I want a good haircut, THEN I should go to Bill's Barbershop. If I went two or three times and received good haircuts again, this rule would be strengthened. If, however, I found out that Bill himself gives excellent haircuts but his apprentice gives bad haircuts, my rule would have to change: IF I want a good haircut, THEN I should go to Bill's Barbershop and specifically ask for Bill to cut my hair. If Bill began to have an unsteady hand due to the effects of aging, then I would have to abandon the rule altogether and go elsewhere for my haircuts.
A: Cognition is a nonlinear process whereby changes and deletions are often made to prior beliefs, thus making it nonmonotonic. Such changes in direction of human thought and reasoning are evident when examining several examples provided by rules. Firstly, rules offer a default and are rough generalizations for expectations that are activated by a search mechanism. In this way they are inherently different from their monotonic relatives based on logical deduction. Unlike logic, rules need not be interpreted as universally true. This is a mistake often made by children as they acquire language. Upon learning a new rule children tend to overgeneralize it to situations in which it does not hold ("I goed" or "the geeses", for example). They eventually surpass this stage in normal development however and go on to delete or change these rules, acquiring a more correct vocabulary. The nonmonotonic nature of cognition is also seen when we witness an alteration in a stereotype (or even the obliteration of one when it rarely occurs). Such rules as IF x is an accountant THEN x is dull may need to be deleted or changed if we meet an exciting, interesting CA. Rules can also be modified through specialization, the alteration of a rule for a specific situation (i.e. IF you have an exam THEN you should study can become IF you have an exam and it's for a "bird" course THEN don't study (this is of course a purely hypothetical example)). In addition, monotonicity is violated when the strengths of rules are increased or decreased based on experience or when they are used bidirectionally to solve a problem (such as how to get from A to B). In conclusion, it is clear that the modifications and deletions that can be made to rules as well as their by directional use makes cognition distinctly nonmonotonic.
How is language useful in understanding cognition?
A: Language can be a very useful tool in understanding cognition, just as cognition can be used to understand language. Several aspects of lanuage indicate this. "People's ability to comprehend language, their ability to produce utterances, and children's universal ability to learn langauge" [Thagard 17] suggest that it is "possible that language is a unique cognitive capacity" [Thagard 17] that is independent of the environment. Further evidence of language's usefulness in understanding cognition can be seen in many cognitive representations. Chomsky postulated a level of "logical form" of language at which meaning is most explicitly represented, and stated that "children learn a language merely by recognizing which of a finite set of possibilities that rule employs" [Thagard 51]. By studying the manner in which meanings are represented in language and the process by which children recognize the rules that govern language, further understanding into the contribution of logic and rules to human cognition may be gained. The process of learning a language can also provide insight into the formation of concepts, since "it involves developing a whole conceptual system" [Thagard 69]. Metaphors, viewed as "a pervasive and valuable form of language" [Thagard 87] require the preception of an underlying analogy by both the speaker and listener. Study of the formation and use of metaphors might provide insight into the use of analogy in cognition, and since many metaphors are visual in origin, of imagery as well. Since language is such a fundamental element of our ability to acquire and communicate information, the study of how it is learned and used could provide valuable information about the validity of these cognitive representations.
A: Language involves rules, concepts, analogy, and other elements of cognition. According to Chomsky, the rules of grammar and syntax are innate: from a very young age, we use language without expressly stating rules. For example, young children pluralize words by adding -S without realizing that they are applying the rule. Rules (grammar) work in conjunction with concepts (words) to express ideas. New words and concepts are learned the same way: by observation or in terms of other words or concepts. Language, as the basis of metaphor, is also useful in studying the nature of analogy. Analogy helps problem-solving: destroying a tumour by weak beams from many directions was inspired by the army that converged on a castle along multiple roads. We can learn about the mind by studying what makes some sentences meaningful, grammatically correct, and effective. The converse is true as well, as a purpose of cognitive science is understanding language.
Are all concepts stereotypes? Why or why not?
A: These is no definite answer to whether all concepts are stereotypes. There are arguments supporting both views and the answer depends on how concepts is being considered in different circumstances. Minsky argued that thinking should be understood as frame application rather than logical deduction. He proposed that concept-like frames are the central form of knowledge representations: schemas and scripts. A frame is a data-structure for representing a stereotyped situation. Frames are never stored in long-term memory with unassigned terminal values. Instead, they are stored with weakly bound default assignments at every terminal that is filled by specific data. Frames are often useful, but can also becomes counterproductive stereotypes. Default assignments of frames have subtle, idiosyncratic influences on the path that people tend to follow in making analogies, generalization, and judgments, especially when the exterior influences on the choices are weak. Such stereotypes can serve as a storehouse of valuable heuristic plan skeletons if properly chosen, but they can also form paralyzing collection of irrational biases if badly selected. On the other hand, Abelson showed that a great deal of our social knowledge consists of scripts which describe typical sequential occurrences. Rumelhart described knowledge in terms of conceptlike structures call schemas that represent not the essence of a concept, such as 'dog', but what is typical of dogs. Likewise, Putnam argued that the meaning of concepts should be thought of in terms of stereotypes, but not in terms of defining conditions. Moreover, there are sometimes 'opposed' locations in a conceptual network. A certain concept (e.g. giant) can have opposite informations attached to it in different cases.
Mother Goose says "A man of words and not of deeds / Is like a garden full of weeds." Give a multiconstraint description of the analogy underlying this couplet. (Hint: Think of the concept of a garden and do not try to represent the moral.)
A:
Source Target garden man weeds words useful vegetation deeds quality credibility produces(garden,useful vegetation,weeds) produces(man,deeds,words) overshadow(weeds,useful vegetation) overshadow(words,deeds) reduce(weeds,quality) reduce(words,credibility) is-measured-by(garden,quality) is-measured-by(man,credibility) because-of(reduces,overshadow) because-of(reduces,overshadow)
A:
Source Target Attributes Attributes ========== ========== garden man flowers deeds weeds words Relations Relations ========= ========= hurt(weeds,garden) *hurt(words,man) improve(flowers,garden) *improve(deeds,man) Systems Systems ======= ======= grows-because(garden,improve) *grows-because(man,*improve) dies-because(garden,hurt) *dies-because(man,*hurt)
Note that this has all the requisite components, but is somewhat figurative, whch weakens the analogy. The relations and systems with a * imply the concepts brought across by the analogy. What is required is either a further analogy or an understanding that a man who "grows" is progressing, not just getting bigger, and one who is "dying" is not necessarily ceasing to live.
How do the limitations of the visual buffer affect visual mental imagery?
A: According to Kosslyn the visual buffer has several limitations these include:
A: As neurological evidence has shown, visual imagery requires the use of the visual buffer, and therefore the limitations of the visual buffer translate more or less directly to limitations in visual imagery. One immediate consequence is that since we only have one visual buffer, we can not really only examine two images in the same space. For example it is difficult to interpret both representations of the Necker cube at the same time in the same space. Of course, one can superimpose images, but this is not the same thing. Other limitations of the visual buffer are its size and precision. Large images interpreted as close up may overflow the visual buffer, and therefore not be seen in their entirety. Similarly small images interpreted as far away will not be seen in detail. Consequently there is a need for mental panning and zooming in order to examine images which have either overflowed the visual buffer, or can not be examined in sufficient detail. Finally, there is an adaptive feature of the visual buffer which leads to limitations in visual imagery. In order to prevent a mess of blurred information, the visual buffer retains images for a only very short amount of time. As a consequence, retention of visual images is difficult and requires mental effort. So the inspection and transformation of visual images requires a good deal of attention, and this is perhaps why it is difficult to mentally rotate unfamiliar complex pictures. In general, it is the fact that visual imagery is so depedent on the use of the visual buffer that leads to its limitations.
Is the backpropagation learning procedure for neural networks psychologically plausible? Why or why not?
A: Thagard (1996) lists two difficulties with using backpropagation to model human learning: it requires a supervisor to indicate whether an error has been made, and it requires many examples train a simple network. These two difficulties can be overcome by considering simple extensions to the backpropagation procedure. For artificial neural networks, the supervisor is always an external supervisor, usually a human operator who provides the feedback response. For a human, an unquestionably more complex system, the supervisor could be less explicit. For instance, when learning a language, the feedback may be as little as interpreting the change in facial expression with one to whom you are talking. It may be as complex as one mental process supervising another. Thus, if this extension is allowed, the psychological plausibility increases. Concerning the required number of examples, the human neural system may simply be more efficient at choosing both initial weightings and changing the weighting values. This efficiency could account for the relatively few examples needed for human learning. Combining these two extensions of backpropagation with results from simulations such as Seidenberg and McClelland (1989) and St. John (1992) proves that backpropagation is psychologically plausible.
A: The backpropagation learning procedure does have some psychological plausibility. Like much human learning, backpropagation occurs gradually and is sensitive to subtle statistical associations which are not picked up by other representations such as rules. Backpropagation has been used to model and predict a whole range of human psychological processes. However, it suffers two major shortcomings as an account of human learning. First, it requires a "teacher" to inform nodes of what their activiation levels should be after processing has taken place, while much human learning seems to take place in the absense of explicit instruction. Second, while some kinds of human learning are slow and gradual like backpropagation, others happen quite quickly, with a minimum of practice. Backpropagation cannot account for these types of learning.