D.A. Forsyth --- 3310 Siebel Center

Office Hours Time: WF 14h00-15h00, Location: 3310 Siebel

TA's:

You may do this homework in groups of no more than three, and by submitting you are certifying the homework is your work.

Submission: Course submission policy is TBA

Problem 1
- Mean field inference for binary images The MNIST dataset consists of 60, 000 images of handwritten digits, curated by Yann LeCun, Corinna Cortes, and Chris Burges. You can find it here , together with a collection of statistics on recognition, etc. We will use the first 500 of the training set images.
  - Obtain the MNIST training set, and binarize the first 500 images by mapping any value below .5 to -1 and any value above to 1. For each image, create a noisy version by randomly flipping 2% of the bits.
  - Now denoise each image using a Boltzmann machine model and mean field inference. Use theta_{ij}=0.2 for the H_i, H_j terms and theta_{ij}=2 for the H_i, X_j terms. To hand in: Report the fraction of all pixels that are correct in the 500 images.
  - To hand in: Prepare a figure showing the original image, the noisy image, and the reconstruction for
    - your most accurate reconstruction
    - your least accurate reconstruction

Instructions to follow