PSTAT197A/CMPSC190DD Fall 2024
Sign up in a Group (3-5 members). [here]
task: create a method vignette on a data science topic or theme
goal: create a reference that you or someone else might use as a starting point next term
deliverable: public repository in the Capstone-24-25 workspace
clustering methods
neural net architecture(s) for … [images, text, time series, spatial data]
analysis of network data
numerical optimization
bootstrapping
geospatial data structures
anomaly detection
functional regression
Your repository should contain:
Thursday 11/21: No formal lecture - time to prepare final project
let us know your topic by the end of day Monday 11/25
No class on Tuesday 11/26
present a draft in class Tuesday 12/3 and Thursday 12/5
finalize repository by Friday 12/13
You’ll need to yourself learn about the topic and implementation by finding reference materials and code examples.
It is okay to borrow closely from other vignettes in creating your own, but you should:
cite them
use different data
do something new
It is not okay to make a collage of reference materials by copying verbatim, or simply rewrite an existing vignette.
An image filter is a function that takes in a local spatial neighborhood of pixel values and detects the presence of some pattern in that data.
Let \(X\) be the original image, of size \(d\); then pixel \(i\) of the output image is specified by:
\[ Y_i = F \cdot (X_{i-1}, X_i) \] This process of applying the filter to the image to create a new image is called convolution.
Number of filters: \(m_l\)
Size of one filter: \(k_l \times k_l \times m_{l-1} + 1 \text{ (for the bias value for this one filter).}\)
Stride \(s_l\): The stride determines the spacing at which the filter is applied to the image.
Input tensor size: \(n_{l-1} \times n_{l-1} \times m_{l-1}\)
Padding \(p_l\): Refers to the number of extra pixels (typically with value \(0\)) added around the edges of the input.
Max-pooling is a simple yet powerful operation in CNNs:
Here is the form of a typical convolutional network:

