LFD Book Forum - Figure 4.3(b)

LFD Book Forum (http://book.caltech.edu/bookforum/index.php)

- Chapter 4 - Overfitting (http://book.caltech.edu/bookforum/forumdisplay.php?f=111)

- - Figure 4.3(b) (http://book.caltech.edu/bookforum/showthread.php?t=4528)

The title of Figure 4.3(b) is "Deterministic noise". However, the "overfitting level first decrease, hit sweet spot, then increase" trend with increase of Qf is the result of not only the deterministic noise, but of the stochastic noise as well. The reason that the fixed hypothesis set overfits the data when the target function is extremely simple is the existence of stochastic noise. If we eliminate that, a 10th order polynomial hypothesis set will fit a 2nd order polynomial target function exactly as it is.
In this sense, the title "Deterministic noise" of figure 4.3(b) is a bit misleading.

Actually, I became aware of this while doing exercise 4.3. At first I thought the answer should be "deterministic noise will go up all the way with increase of target complexity", then I looked at Fig. 4.3(b) and thought "hey, it says deterministic noise will first go down and then go up!". But come to think of it, the figure is really Eout(H10)-Eout(H2), which include effect of both stochastic and deterministic noise, hence the post.

Re: Figure 4.3(b)

An explanation by @yaser was posted on the 2014 edX forum:

Quote:

Q. Why is the bottom part of this figure behaving differently?

A. This is an artifact and it has to do with our choice of the models (2nd and 10th order polynomials). For Qf≤10 there is no deterministic noise for H10 (we can perfectly fit them).

Q. Why did we add stochastic noise to the target when generating the above figure; aren’t we just analyzing deterministic noise?

A. We wanted to compare the two figures fairly. When plotting the impact of the stochastic noise, we already had some built-in deterministic noise in our target function as well.

Re: Figure 4.3(b)

Another question about this figure: is there any intuition as to why regions of a given color (fixed overfit measure) are roughly linear in the stochastic N-sigma^2 graph, but non-linear (exponential?) in the deterministic N-Qf graph (for Qf > 10)?

Thank you!

Re: Figure 4.3(b)

Quote:

Originally Posted by ypeels (Post 11872)

The analysis of the stochastic noise figure may be doable given the clean analytic components of the simulation (Legendre, pseudo-inverse, Gaussian noise). In the deterministic noise figure, the noise value is quantified by the complexity of the target

. While deterministic noise (the part of

that cannot be captured by $\mathcal H$ ) is indeed related to

, it is not necessarily linearly related to it so that direct parallel with what happens with stochastic noise does not hold.

BTW, the LaTeX stuff is done by delimiters [ math ] and [ /math ] (without the spaces) instead of $ and $. A bit cumbersome when you use a lot of math.