One thought I had while looking into regression recently is to consider the model created with a given regularization coefficient not as a line on a 2 dimensional graph but as a slice of a surface on a 3 dimensions graph where the third dimension is the regularization coefficient.

In my case the model was for logistic regression and it was the boundary lines of the classification, but the thought is largely the same. Viewing it as a 3d shape form by boundary lines and considering hill tops as areas where entire classification boundaries disappeared as the regularization coefficient grew large enough to eliminate them. Impractical to do on models of any size and only useful when looking at two features at a time, but a fun consideration.

More on topic with the article, how well does this work with considering multiple features and the different combinations of them. Instead of sigma(n => 50) of x^n, what happens if you have sigma(n => 50) of sigma(m => 50) of (x^n)*(y^n). Well probably less than 50 in the second example, maybe it is fair to have n and m go to 7 so there are 49 total terms compared to the original 50, instead of 2500 terms if they both go to 50.