/STAD80-Review_Of_Double_Descent

A Review of Preetum Nakkiran's "More Data Can Hurt for Linear Regression: Sample-wise Double Descent"

Primary LanguageTeXMIT LicenseMIT

STAD80: A Review of Preetum Nakkiran's "More Data Can Hurt for Linear Regression: Sample-wise Double Descent"

Authors

  • Shawn Santhoshgeorge
  • Anaqi As Shafiq Bin Amir Razif

Summary

In overparameterized linear regression, the phenomenon of double descent can be observed were the Test MSE initially decreases with a fixed parameter equation and an increasing sample equation. It keeps on decreasing before increasing asymptotically towards infinity when equation before it begins decreasing again. We call the regime where equation to be the overparameterized regime and the underparameterized regime is when equation. There are interesting concepts that can be found when researching and analyzing these regimes especially around the area of equation. This repository looks into the claims made in "More Data Can Hurt for Linear Regression: Sample-wise Double Descent" by Preetum Nakkiran. It includes the a deeper analysis on Section 3.1 of paper and works to explain the conclusions of the paper.

Results

This graph was created based on a simple simulation based on the configuration stated in the paper and was a test to see if we could replicate the claims.

References