I. Kernel Eigenfaces
In this section, you are going to do face recognition using eigenface and fisherface.
Reference: https://www.csie.ntu.edu.tw/~mhyang/papers/fg02.pdf
● Data
o The Yale Face Database.zip contains 165 images of 15 subjects
(subject01, subject02, etc.). There are 11 images per subject, one for each of the following facial expressions or configurations: center-light, w/glasses, happy, left-light, w/no glasses, normal, right-light, sad, sleepy, surprised, and wink.
o Thesedataareseparatedintotrainingdataset(135images)andtesting dataset(30 images). You can resize the images for easier implementation.
● What you are going to do
o Part1:UsePCAandLDAtoshowthefirst25eigenfacesandfisherfaces,
and randomly pick 10 images to show their reconstruction. (please refer
to the lecture slides).
o Part2:UsePCAandLDAtodofacerecognition,andcomputethe
performance. You should use k nearest neighbor to classify which subject
the testing image belongs to.
o Part3:UsekernelPCAandkernelLDAtodofacerecognition,and
compute the performance. (You can choose whatever kernel you want, but you should try different kernels in your implementation.) Then compare the difference between simple LDA/PCA and kernel LDA/PCA, and the difference between different kernels.
II. t-SNE
Here are nice implementations of t-SNE in different programming languages:
https://lvdmaaten.github.io/tsne/
● Data & reference code o Download link:
https://lvdmaaten.github.io/tsne/code/tsne_python.zip,
o mnist2500_X.txt:contains2500featurevectorswithlength784,for
describing 2500 mnist images.
o mnist2500_labels.txt:providescorrespondinglabels o tsne.py: reference code
● What you are going to do
o Part1: Try to modify the code a little bit and make it back to symmetric
SNE. You need to first understand how to implement t-SNE and find out the specific code piece to modify. You have to explain the difference between symmetric SNE and t-SNE in the report (e.g. point out the crowded problem of symmetric SNE).
o Part2: Visualize the embedding of both t-SNE and symmetric SNE. Details of the visualization:
▪ Project all your data onto 2D space and mark the data points into different colors respectively. The color of the data points depends on the label.
▪ Use videos or GIF images to show the optimize procedure. o Part3: Visualize the distribution of pairwise similarities in both high-
dimensional space and low-dimensional space, based on both t-SNE and
symmetric SNE.
o Part4:Trytoplaywithdifferentperplexityvalues.Observethechangein
visualization and explain it in the report.
III. Report
- ● Submit a report in pdf format. The report should be written in English.
- ● Report format:
○ a. code with detailed explanations (40%)
- Paste the screenshot of your functions with comments and
explain your code. For example, explain the process to clustering
and show different kernels, etc.
- Note that if you don’t explain your code clearly, you cannot get any points in section b and c either.
● Kernel Eigenfaces ○ Part1 (10%)
○ Part2 (5%)
○ Part3 (10%) ● t-SNE
○ Part1 (10%) ○ Part2 (2%) ○ Part3 (2%) ○ Part4 (1%)
○ b. experiments settings and results (35%) & discussion (15%) ■ Show everything we asked you to show
● Kernel Eigenfaces ○ Part1 (5%) ○ Part2 (5%)
○ Part3 (5%) & (5%) ● t-SNE
○ Part1 (5%) & (5%) ○ Part2 (5%)
○ Part3 (5%)
○ Part4 (5%) & (5%)
○ c. observations and discussion (10%)
■ Anything you want to discuss, such as the meaning of eigenfaces
or trying different dimension reduction methods, comparing the advantages and disadvantages of them.
