Janice Chen <kanile@mit.edu>

> 1. It seems that a problem with SPR and the use of Bayes Decision Theory
> may be that BDT is not behaviorally useful, that is, humans do not actually
> make decisions in the manner described by the theory. In fact, it is common
> knowledge that most people, even with training, score very poorly on tests of
> probability comprehension. Although these kinds of calculations might be
> achievable by the perceptual system, it is less believable that categorization
> operates in this manner. Perhaps a method could be devised to determine
> subjects' priors - maybe about a controlled laboratory topic - and then
> compare their categorical perceptions to results calculated by Bayes net
> principles.
> 
> 2. Briefly describe the steps of the "alignment approach" model of
> object recognition.
> 
> 3. Which of these is NOT one of the image transforms used in the alignment
> approach model of object recognition?
> a. centering
> b. inverse (correct)
> c. orientation
> d. scaling

Yuri Ostrovsky <yo@mit.edu>

1) Research Question

Do current statistical learning paradigms capture the low- and high-level aspects of human object recognition implicitly, or are they only capturing the low-level aspects, succeeding where there happens to be enough low-level information? Does it even make sense to ask this question?

2) Short Answer Question

In the Nearest Neighbor approach to statistical learning, what is the potential problem known as the "accidental outlyer?"

3) Multiple Choice Question

What is the definition of the L2 distance function d(x,y)?

a. d(x,y) = [sigma(i) (x_i-y_i)^2]^(1/2)

b. d(x,y) = max(i) |x_i - y_i|

c. d(x,y) = sigma(i) |x_i-y_i|

d. d(x,y) = max(i) x_i^2 - y_i^2