This project is built to explore how ridge regression with optimal regularization behaves on datasets with non-gaussian noise applied to the training data labels. This is mostly to observe if the phenomena described as "double descent" can be mitigated under these alternative parameters.
The double descent phenomena is described as a critical interval where adding more data causes the test error to rise, then re-descend.
Deep Double Descent: Where Bigger Models and More Data Hurt
- Calculate an optimal lamba (for regression) as described in Optimal Regularization Can Mitigate Double Descent.
- Plotting the 2D models
- Store results in xlsx file.
- Batch generation
- Gaussian & Non gaussian noise methods