D.A. Forsyth --- 3310 Siebel Center

Looking up at the surface along a cliff face from 20m

TA's:

MCS-DS version of the course
- Krishna Kothapalli Lead
- Daniel Calzada
- Taiyu Dong
- Shreya Rajpal
- Ehsan Saleh
Online version of the course
- Binglin Chen
In person version of the course
- Christopher Benson
- Ji Li
- Maghav Kumar

Remission of Sins: Due 10 May 23h59 (Thur; midnight)

You may do this homework in groups of up to 3 contributors. Groups of 1 or of 2 are just fine, too. A group can consist of any mixture of any type of student (MSCS-DS/Online/Face). We do not offer coordination services for complex group interactions, and you may want to take this into account when forming your group.

Submission: Submission details as for earlier homeworks.

No programming environment warnings for this homework.

This homework is optional, and is intended for remission of sins (i.e. if some other homework was submitted late, went bad, got eaten by the dog, etc.). We will compute a grade for this homework, and substitute that for your worst grade, if you submit this and ask us to. However, we will not substitute for Homework 9. By the way, if you have great grades, you could do this homework anyhow. It won't affect your grade (we're not going over 100%, that way lies madness) but you should learn a lot, which is the purpose of the exercise.

Denoising autoencoders: We will evaluate denoising autoencoders applied to the MNIST dataset.

Obtain (or write! but this isn't required) a tensorflow code for a denoising autoencoder. Train this autoencoder on the MNIST dataset. Use only the MNIST training set. You should use at least three layers in the encoder and in the decoder.
We now need to determine how well this autoencoder works. For each image in the MNIST test dataset, compute the residual error of the autoencoder. This is the difference between the true image and the reconstruction of that image by the autoencoder. It is an image itself. Prepare a figure showing the mean residual error, and the first five principal components. Each is an image. You should preserve signs (i.e. the mean residual error may have negative as well as positive entries). The way to show these images most informatively is to use a mid gray value for zero, then darker values for more negative image values and lighter values for more positive values. The scale you choose matters. You should show
- mean and five principal components on the same gray scale for all six images, chosen so the largest absolute value over all six images is full dark or full light respectively and
- mean and five principal components on a scale where the gray scale is chosen for each image separately.

Variational autoencoders: We will evaluate variational autoencoders applied to the MNIST dataset.

Obtain (or write! but this isn't required) a tensorflow code for a variational autoencoder. Train this autoencoder on the MNIST dataset. Use only the MNIST training set.
We now need to determine how well the codes produced by this autoencoder can be interpolated.
- For 10 pairs of MNIST test images of the same digit, selected at random, compute the code for each image of the pair. Now compute 7 evenly spaced linear interpolates between these codes, and decode the result into images. Prepare a figure showing this interpolate. Lay out the figure so each interpolate is a row. On the left of the row is the first test image; then the interpolate closest to it; etc; to the last test image. You should have a 10 rows and 9 columns of images.
- For 10 pairs of MNIST test images of different digits, selected at random, compute the code for each image of the pair. Now compute 7 evenly spaced linear interpolates between these codes, and decode the result into images. Prepare a figure showing this interpolate. Lay out the figure so each interpolate is a row. On the left of the row is the first test image; then the interpolate closest to it; etc; to the last test image. You should have a 10 rows and 9 columns of images.