Saturday, March 8, 2008


The midterm is worth 25 points (25% of your final grade).

There will be three problems (8 points each) and three simple exercises (3 points each). Thus if you solve any two problems + all exercises, you will still get the maximum grade. If you do everything, you will get 8 bonus points.

The material covered in the following lectures will NOT be on the midterm:
  1. Generalization bounds (the lecture before the midterm)
  2. Reductions
  3. Sanjoy's talk on projections, Gaussian scale mixtures, and the EM
One of the problems will be on the semantics of loss functions. You should be able to answer questions of the form: What loss function does the following quantity minimize? (See this post, for example.)

The other two problems will be on any two of the following algorithms:
  1. Winnow
  2. Halving
  3. (Randomized) Weighted Majority
  4. Perceptron
The exercises will cover the rest of the material.


COMS 4771 said...

A student asks: "Will the problem questions be of the nature of deriving proofs or application based"

Reply: You won't have to produce proofs, but I may ask you to produce a sequence of examples that forces a given algorithm to make a certain number of mistakes.

Web mining said...

How long will the exam take?

COMS 4771 said...


Zhenia said...

is it possible to post some practice problems for the exam?

COMS 4771 said...

That's what homeworks were for!

COMS 4771 said...

A good way to find practice problems is to go to other Machine Learning classes and check out their homework assignments and exams (e.g.,