Data Science 311
Lab 6 (10 points)
Due at 10am on Nov, 28, 2022
Read all of the instructions. Late work will not be accepted.
Overview
In this lab you will work from the notebook presented in class today imagenet cluster.ipynb.
You will use clustering to investigate characteristics of the ImageNet ILSVRC dataset that
are not immediately apparent from the original labels or a cursory visual inspection. I have
curated a toy subset of the ILSVRC validation data containing 8 of the original classes with 50
images each. imagenet cluster.ipynb contains all the code you need to extract the image
feature vectors from a pretrained Resnet50 model.
Collaboration
For this lab, you can
ainstorm with any classmates about ideas for datasets and data curation
methodology. However, your dataset and co
esponding explanation of the intended analysis
should be unique. Your submission must acknowledge ideas or suggestions you received from
other classmates in an acknowledgement section at the end of your notebook.
Details
Tasks
1. Using the columns of ”label” and ”cluster” make a matrix with the rows the original
label [0-9] and the columns the assigned cluster [0-nclusters]. The entries of the matrix
at entry i, j will be the number of class i assigned to cluster j. I suspect you will find that
most images from the same original class are assigned to the same cluster. However there
may be some ”oddball” images that get assigned to some other cluster. You may even
find that images from one orignal label straddle two distinct clusters in the distribution.
Perform the following analyses:
(a) Describe the distribution of classes over clusters. Some histograms may help in
addition to the class/cluster count matrix.
(b) Look at some oddball images from each class. What are your observations about
how these images might differ from the rest of the class images.
(c) Are there any classes that get grouped together into the same cluster? What are
they? Why might these images from the separate classes be grouped together?
(d) Provide a short paragraph describing your findings and any other interesting ob-
servations from this clustering exercise. Make sure to display images that illustrate
your conclusions in the notebook
1
2. Next you will perform ”within class” K-means clustering for your choice of 4 classes from
our toy dataset. Use K = 2. For each within class clustering discuss if you find some
meaningful patterns in the clusters that indicate distinct visual concepts. For instance,
from a visual inspection in lecture it looked like the ”dome” class might be logically split
etween ”inside dome” and ”outside dome”. Discuss your findings. Make sure to display
images that illustrate your conclusions in the notebook.
Submitting Your Work
You will submit a single file, Firstname Lastname lab6.zip. containing your lab6.ipynb file.
(where spelling, spacing and capitalization matter) and upload the zip via Canvas.
Grading
ImageNet cluster analysis will be graded on the following:
• 23% of total grade will be on co
ectness.
• 13% of total grade will be on clear exposition of findings.
• 12% of total grade will be on task 1.
• 12% of total grade will be on task 2.
2