Take Home Quiz, ECE 547, Fall 1998

Don't bust yourself on this - it's a reality check for me, to see what people understand and don't understand.
  1. Say something nice about backpropagation, and say something nasty about backpropagation.

  2. You are fitting a density to the one-dimensional dataset shown below. The parametric form of your density is a mixture of two gaussians, and you are using EM for the optimization. Right now, your two gaussians are in the configuration shown below.

    Vaguely estimate the values of the z's you would next calculate (these are the variables which assign responsibility for points in the dataset to gaussians in the mixture, ie z5,1 is how must responsibility gaussian 1 has for point 5.)

    points
    1234567
    gaussians1       
     2       

    Fill in the table of z values; points are numbered left to right; just get them vaguely right, ie the correct ordinal relationships and correctly close to 0 or 1 and suchlike.

  3. You are hired by your country's navy to build a NN system which, when attached to a gun, makes it refuse to fire upon friendly or surrendering soldiers. Describe what kind of data you would need, what kind of network architecture you would use, how you would train it, how you would test it, and (this is the ethical component of the course required by ABET) whether you think the project is a good idea.

Barak Pearlmutter <bap@cs.unm.edu>