2014HS: 33065 Applications of Artificial Intelligence

Here are the answers I would have (I might also make mistakes)

1)

For me it clearly depends on the problem you want to solve. This could be found by experiment.

There is certainly a good size for the human speech but I don't recall.

2)

GSF : I think this show how the sentence is close to a given grammar (higher = better).

WIP : The model tries to make out as much words as possible from a given written sentence (I would say medium = better, we don't want to invent words from a sentence). 

3)

HMM is based on Markov Model that is a way to describe probabilities to pass from a state to another. With HMM we hide the probabilities states and rely on extern informations to try to compute the hidden transition probabilities (hidden means we do not know them for us).

As Markov Model are the way to compute state from one to another we can easily combine different Markov Model to have a bigger Markov Model and thus a bigger HMM. This is not the case with ANN because the number of "cells" is fixed when the training begins. HMM are a special case of RNN.

4)

Here I might be mistaken but we can see PCA this way. We simply try to found an ellipse that enclose the data given. Given that ellipse we could fit an orthogonal basis in the center of it. Then we could find a Gaussian distribution that fits over the "largest" base and use it to compare new vectors with the data.

5)

Imagine a situation where you have 3 different classes. During the training you admit that if a vector is from class A or B it is the same. If you first train for class A and then B you might loose a lot of information about the class A. This is why you should have a NN with memory. You could also imagine a case where the data is changing over the time and from a class A a new class B which is linked to A appears. (a new way to pronounce words for example)