0.14 Maximum likelihood and complexity regularization  (Page 2/2)

 Page 2 / 2

Error bound

Suppose that we have a pool of candidate functions $\mathcal{F}$ , and we want to select a function $f$ from $\mathcal{F}$ using the training data. Our usual approach is to show that the distribution of $\stackrel{^}{{R}_{n}}\left(f\right)$ concentrates about its mean as $n$ grows. First, we assign a complexity $c\left(f\right)>0$ to each $f\in \mathcal{F}$ so that $\sum {2}^{-c\left(f\right)}\le 1$ . Then, apply the union bound to get a uniform concentration inequality holding for all models in $\mathcal{F}$ . Finally, we use this concentration inequality to bound the expected risk of our selected model.

We will essentially accomplish the same result here, but avoid the need for explicit concentration inequalities and instead make use of the information-theoretic bounds.

We would like to select an $f\in \mathcal{F}$ so that the excess risk is small.

$\begin{array}{ccc}\hfill 0& \le & R\left(f\right)-R\left({f}^{*}\right)\hfill \\ & =& \frac{1}{n}E\left[log{p}_{{f}^{*}}\left(Y\right)-log{p}_{f}\left(Y\right)\right]\hfill \\ & =& \frac{1}{n}E\left[log,\frac{{p}_{{f}^{*}}\left(Y\right)}{{p}_{f}\left(Y\right)}\right]\hfill \\ & \equiv & \frac{1}{n}K\left({p}_{f},{p}_{{f}^{*}}\right)\hfill \end{array}$

where

$K\left({p}_{f},{p}_{{f}^{*}}\right)=\sum _{i=1}^{n}\underset{K\left({p}_{f\left({x}_{i}\right)},{p}_{{f}^{*}\left({x}_{i}\right)}\right)}{\underbrace{\left(\int ,log,\frac{{p}_{{f}^{*}\left({x}_{i}\right)}\left({y}_{i}\right)}{{p}_{f\left({x}_{i}\right)}\left({y}_{i}\right)},·,{p}_{{f}^{*}\left({x}_{i}\right)},\left({y}_{i}\right),\phantom{\rule{0.166667em}{0ex}},\mathrm{d},{y}_{i}\right)}}$

is again the KL divergence.

Unfortunately, as mentioned before, $K\left({p}_{f},{p}_{{f}^{*}}\right)$ is not a true distance. So instead we will focus on the expected squared Hellinger distance as our measure of performance. We will get a bound on

$\frac{1}{n}E\left[{H}^{2},\left({p}_{f}\left(Y\right),{p}_{{f}^{*}}\left(Y\right)\right)\right]=\frac{1}{n}\sum _{i=1}^{n}\left(\int ,{\left(\sqrt{{p}_{f\left({x}_{i}\right)}\left({y}_{i}\right)},-,\sqrt{{p}_{{f}^{*}\left({x}_{i}\right)}\left({y}_{i}\right)}\right)}^{2},\phantom{\rule{0.166667em}{0ex}},\mathrm{d},{y}_{i}\right).$

Theorem

Li-barron 2000, kolaczyk-nowak 2002

Let ${\left\{{x}_{i},{Y}_{i}\right\}}_{i=1}^{n}$ be a random sample of training data with $\left\{{Y}_{i}\right\}$ independent,

${Y}_{i}\sim {p}_{{f}^{*}\left({x}_{i}\right)}\left({y}_{i}\right)\phantom{\rule{1.em}{0ex}},i=1,...,n$

for some unknown function ${f}^{*}$ .

Suppose we have a collection of candidate functions $\mathcal{F}$ , and complexities $c\left(f\right)>0,f\in \mathcal{F}$ , satisfying

$\sum _{f\in \mathcal{F}}{2}^{-c\left(f\right)}\le 1.$

Define the complexity-regularized estimator

$\stackrel{^}{{f}_{n}}\equiv arg\phantom{\rule{0.166667em}{0ex}}\underset{f\in \mathcal{F}}{min}\left\{-,\frac{1}{n},\sum _{i=1}^{n},log,\phantom{\rule{0.166667em}{0ex}},{p}_{f},\left({Y}_{i}\right),+,\frac{2c\left(f\right)log2}{n}\right\}\phantom{\rule{0.166667em}{0ex}}.$

Then,

$\begin{array}{ccc}\hfill \frac{1}{n}E\left[{H}^{2},\left({p}_{f}\left(Y\right),{p}_{{f}^{*}}\left(Y\right)\right)\right]& \le & -\frac{2}{n}E\left[log,\left(A,\left(,{p}_{f},\left(Y\right),,,{p}_{{f}^{*}},\left(Y\right),\right)\right)\right]\hfill \\ & \le & \underset{f\in \mathcal{F}}{min}\left\{\frac{1}{n},K,\left({p}_{f},{p}_{{f}^{*}}\right),+,\frac{2c\left(f\right)log2}{n}\right\}\phantom{\rule{0.166667em}{0ex}}.\hfill \end{array}$

Before proving the theorem, let's look at a special case.

Gaussian noise

Suppose ${Y}_{i}=f\left({x}_{i}\right)+{W}_{i}\phantom{\rule{1.em}{0ex}},{W}_{i}\stackrel{i.i.d.}{\sim }\mathcal{N}\left(0,{\sigma }^{2}\right)$ .

${p}_{f\left({x}_{i}\right)}\left({y}_{i}\right)=\frac{1}{\sqrt{2\pi {\sigma }^{2}}}{e}^{-\frac{{\left({y}_{i}-f\left({x}_{i}\right)\right)}^{2}}{2{\sigma }^{2}}}.$

Using results from example 1 , we have

$\begin{array}{ccc}\hfill -2logA\left({p}_{\stackrel{^}{{f}_{n}}},\left(Y\right),,,{p}_{{f}^{*}},\left(Y\right)\right)& =& \sum _{i=1}^{n}\phantom{\rule{0.166667em}{0ex}}-2logA\left({p}_{\stackrel{^}{{f}_{n}}\left({x}_{i}\right)},\left({Y}_{i}\right),,,{p}_{{f}^{*}\left({x}_{i}\right)},\left({Y}_{i}\right)\right)\hfill \\ & =& \sum _{i=1}^{n}\phantom{\rule{0.166667em}{0ex}}-2log\int \sqrt{{p}_{\stackrel{^}{{f}_{n}}\left({x}_{i}\right)}\left({y}_{i}\right)·{p}_{{f}^{*}\left({x}_{i}\right)}\left({y}_{i}\right)}\phantom{\rule{0.166667em}{0ex}}\mathrm{d}{y}_{i}\hfill \\ & =& \frac{1}{4{\sigma }^{2}}\sum _{i=1}^{n}{\left(\stackrel{^}{{f}_{n}},\left({x}_{i}\right),-,{f}^{*},\left({x}_{i}\right)\right)}^{2}.\hfill \end{array}$

Then,

$-\frac{2}{n}E\left[log,\phantom{\rule{0.166667em}{0ex}},A,\left(,{p}_{\stackrel{^}{{f}_{n}}},,,{p}_{{f}^{*}},\right)\right]=\frac{1}{4{\sigma }^{2}n}\sum _{i=1}^{n}E\left[{\left(\stackrel{^}{{f}_{n}},\left({x}_{i}\right),-,{f}^{*},\left({x}_{i}\right)\right)}^{2}\right].$

We also have,

$\begin{array}{ccc}\hfill \frac{1}{n}K\left({p}_{f},{p}_{{f}^{*}}\right)& =& \frac{1}{n}\sum _{i=1}^{n}\frac{{\left(f,\left({x}_{i}\right),-,{f}^{*},\left({x}_{i}\right)\right)}^{2}}{2{\sigma }^{2}}\hfill \\ \hfill -log\phantom{\rule{0.166667em}{0ex}}{p}_{f}\left(Y\right)& =& \sum _{i=1}^{n}\frac{{\left({Y}_{i}-f\left({X}_{i}\right)\right)}^{2}}{2{\sigma }^{2}}\phantom{\rule{0.166667em}{0ex}}.\hfill \end{array}$

Combine everything together to get

$\stackrel{^}{{f}_{n}}=arg\underset{f\in \mathcal{F}}{min}\left\{\frac{1}{n},\sum _{i=1}^{n},\frac{{\left({Y}_{i}-f\left({X}_{i}\right)\right)}^{2}}{2{\sigma }^{2}},+,\frac{2c\left(f\right)log2}{n}\right\}\phantom{\rule{0.166667em}{0ex}}.$

The theorem tells us that

$\frac{1}{4n}\sum _{i=1}^{n}E\left[\frac{{\left(\stackrel{^}{{f}_{n}},\left({x}_{i}\right),-,{f}^{*},\left({x}_{i}\right)\right)}^{2}}{{\sigma }^{2}}\right]\le \underset{f\in \mathcal{F}}{min}\left\{\frac{1}{n},\sum _{i=1}^{n},\frac{{\left(f,\left({x}_{i}\right),-,{f}^{*},\left({x}_{i}\right)\right)}^{2}}{2{\sigma }^{2}},+,\frac{2c\left(f\right)log2}{n}\right\}$

or

$\frac{1}{n}\sum _{i=1}^{n}E\left[{\left(\stackrel{^}{{f}_{n}},\left({x}_{i}\right),-,{f}^{*},\left({x}_{i}\right)\right)}^{2}\right]\le \underset{f\in \mathcal{F}}{min}\left\{\frac{2}{n},\sum _{i=1}^{n},{\left(f,\left({x}_{i}\right),-,{f}^{*},\left({x}_{i}\right)\right)}^{2},+,\frac{8{\sigma }^{2}c\left(f\right)log2}{n}\right\}.$

Now let's come back to the proof.

Proof
$\begin{array}{ccc}\hfill {H}^{2}\left({p}_{\stackrel{^}{{f}_{n}}},,,{p}_{{f}^{*}}\right)& =& \int {\left(\sqrt{{p}_{\stackrel{^}{{f}_{n}}}\left(y\right)},-,\sqrt{{p}_{{f}^{*}}\left(y\right)}\right)}^{2}\phantom{\rule{0.166667em}{0ex}}\mathrm{d}y\hfill \\ & \le & -2log\underset{affinity}{\underbrace{\left(\int ,\sqrt{{p}_{\stackrel{^}{{f}_{n}}}\left(y\right)·{p}_{{f}^{*}}\left(y\right)},\phantom{\rule{0.166667em}{0ex}},\mathrm{d},y\right)}}\hfill \end{array}$
$⇒$
$E\left[{H}^{2},\left({p}_{\stackrel{^}{{f}_{n}}},,,{p}_{{f}^{*}}\right)\right]\le 2\phantom{\rule{0.166667em}{0ex}}E\left[log,\left(\frac{1}{\int \sqrt{{p}_{\stackrel{^}{{f}_{n}}}\left(y\right)·{p}_{{f}^{*}}\left(y\right)}\phantom{\rule{0.166667em}{0ex}}\mathrm{d}y}\right)\right]\phantom{\rule{0.166667em}{0ex}}.$

Now, define the theoretical analog of $\stackrel{^}{{f}_{n}}$ :

${f}_{n}=arg\underset{f\in \mathcal{F}}{min}\left\{\frac{1}{n},K,\left({p}_{f},,,{p}_{{f}^{*}}\right),+,\frac{2c\left(f\right)log2}{n}\right\}.$

Since

$\begin{array}{ccc}\hfill \stackrel{^}{{f}_{n}}& =& arg\underset{f\in \mathcal{F}}{min}\left\{-,\frac{1}{n},log,{p}_{f},\left(Y\right),+,\frac{2c\left(f\right)log2}{n}\right\}\hfill \\ & =& arg\underset{f\in \mathcal{F}}{max}\left\{\frac{1}{n},\left(log,{p}_{f},\left(Y\right),-,2,c,\left(f\right),log,2\right)\right\}\hfill \\ & =& arg\underset{f\in \mathcal{F}}{max}\left\{\frac{1}{2},\left(log,{p}_{f},\left(Y\right),-,2,c,\left(f\right),log,2\right)\right\}\hfill \\ & =& arg\underset{f\in \mathcal{F}}{max}\left\{log,\left(\sqrt{{p}_{f}\left(Y\right)},·,{e}^{-c\left(f\right)log2}\right)\right\}\hfill \\ & =& arg\underset{f\in \mathcal{F}}{max}\left\{\sqrt{{p}_{f}\left(Y\right)},·,{e}^{-c\left(f\right)log2}\right\}\hfill \end{array}$

we can see that

$\frac{\sqrt{{p}_{\stackrel{^}{{f}_{n}}}\left(Y\right)}{e}^{-c\left(\stackrel{^}{{f}_{n}}\right)log2}}{\sqrt{{p}_{{f}_{n}}\left(Y\right)}{e}^{-c\left({f}_{n}\right)log2}}\ge 1\phantom{\rule{0.166667em}{0ex}}.$

Then can write

$\begin{array}{ccc}\hfill E\left[{H}^{2},\left({p}_{\stackrel{^}{{f}_{n}}},,,{p}_{{f}^{*}}\right)\right]& \le & 2\phantom{\rule{0.166667em}{0ex}}E\left[log,\left(\frac{1}{\int \sqrt{{p}_{\stackrel{^}{{f}_{n}}}\left(y\right)·{p}_{{f}^{*}}\left(y\right)}\phantom{\rule{0.166667em}{0ex}}\mathrm{d}y}\right)\right]\hfill \\ & \le & 2\phantom{\rule{0.166667em}{0ex}}E\left[log,\left(\frac{\sqrt{{p}_{\stackrel{^}{{f}_{n}}}\left(Y\right)}{e}^{-c\left(\stackrel{^}{{f}_{n}}\right)log2}}{\sqrt{{p}_{{f}_{n}}\left(Y\right)}{e}^{-c\left({f}_{n}\right)log2}},·,\frac{1}{\int \sqrt{{p}_{\stackrel{^}{{f}_{n}}}·{p}_{{f}^{*}}}\phantom{\rule{0.166667em}{0ex}}\mathrm{d}y}\right)\right]\phantom{\rule{0.166667em}{0ex}}.\hfill \end{array}$

Now, simply multiply the argument inside the $log$ by $\sqrt{\frac{{p}_{{f}^{*}}\left(Y\right)}{{p}_{{f}^{*}}\left(Y\right)}}$ to get

$\begin{array}{ccc}\hfill E\left[{H}^{2},\left({p}_{\stackrel{^}{{f}_{n}}},,,{p}_{{f}^{*}}\right)\right]& \le & 2\phantom{\rule{0.166667em}{0ex}}E\left[log,\left(\frac{\sqrt{{p}_{{f}^{*}}\left(Y\right)}}{\sqrt{{p}_{{f}_{n}}\left(Y\right)}},\frac{\sqrt{{p}_{\stackrel{^}{{f}_{n}}}\left(Y\right)}}{\sqrt{{p}_{{f}^{*}}\left(Y\right)}},\frac{{e}^{-c\left(\stackrel{^}{{f}_{n}}\right)log2}}{{e}^{-c\left({f}_{n}\right)log2}},·,\frac{1}{\int \sqrt{{p}_{\stackrel{^}{{f}_{n}}}\left(y\right)·{p}_{{f}^{*}}\left(y\right)}\phantom{\rule{0.166667em}{0ex}}\mathrm{d}y}\right)\right]\hfill \\ & =& E\left[log,\left(\frac{{p}_{{f}^{*}}\left(Y\right)}{{p}_{{f}_{n}}\left(Y\right)}\right)\right]+2c\left({f}_{n}\right)log2\hfill \\ & & \phantom{\rule{4pt}{0ex}}\phantom{\rule{4pt}{0ex}}\phantom{\rule{4pt}{0ex}}\phantom{\rule{4pt}{0ex}}+2E\left[log,\left(\frac{\sqrt{{p}_{\stackrel{^}{{f}_{n}}}\left(Y\right)}}{\sqrt{{p}_{{f}^{*}}\left(Y\right)}},·,\frac{{e}^{-c\left(\stackrel{^}{{f}_{n}}\right)log2}}{\int \sqrt{{p}_{\stackrel{^}{{f}_{n}}}\left(y\right)·{p}_{{f}^{*}}\left(y\right)}\phantom{\rule{0.166667em}{0ex}}\mathrm{d}y}\right)\right]\hfill \\ & =& K\left({p}_{{f}_{n}},,,{p}_{{f}^{*}}\right)+2c\left({f}_{n}\right)log2\hfill \\ & & \phantom{\rule{4pt}{0ex}}\phantom{\rule{4pt}{0ex}}\phantom{\rule{4pt}{0ex}}+2E\left[log,\left(\frac{\sqrt{{p}_{\stackrel{^}{{f}_{n}}}\left(Y\right)}}{\sqrt{{p}_{{f}^{*}}\left(Y\right)}},·,\frac{{e}^{-c\left(\stackrel{^}{{f}_{n}}\right)log2}}{\int \sqrt{{p}_{\stackrel{^}{{f}_{n}}}\left(y\right)·{p}_{{f}^{*}}\left(y\right)}\phantom{\rule{0.166667em}{0ex}}\mathrm{d}y}\right)\right]\phantom{\rule{0.166667em}{0ex}}.\hfill \end{array}$

The terms $K\left({p}_{{f}_{n}},,,{p}_{{f}^{*}}\right)+2c\left({f}_{n}\right)log2$ are precisely what we wanted for the upper bound of the theorem. So, to finish theproof we only need to show that the last term is non-positive. Applying Jensen's inequality, we get

$2E\left[log,\left(\frac{\sqrt{{p}_{\stackrel{^}{{f}_{n}}}\left(Y\right)}}{\sqrt{{p}_{{f}^{*}}\left(Y\right)}},·,\frac{{e}^{-c\left(\stackrel{^}{{f}_{n}}\right)log2}}{\int \sqrt{{p}_{\stackrel{^}{{f}_{n}}}\left(y\right)·{p}_{{f}^{*}}\left(y\right)}\phantom{\rule{0.166667em}{0ex}}\mathrm{d}y}\right)\right]\le 2log\left(E,\left[{e}^{-c\left(\stackrel{^}{{f}_{n}}\right)log2},·,\frac{\sqrt{\frac{{p}_{\stackrel{^}{{f}_{n}}}\left(Y\right)}{{p}_{{f}^{*}}\left(Y\right)}}}{\int \sqrt{{p}_{\stackrel{^}{{f}_{n}}}\left(y\right)·{p}_{{f}^{*}}\left(y\right)}\phantom{\rule{0.166667em}{0ex}}\mathrm{d}y}\right]\right)\phantom{\rule{0.166667em}{0ex}}.$

Both $Y$ and $\stackrel{^}{{f}_{n}}$ are random, which makes the expectation difficult to compute. However, we can simplify the problem using the union bound,which eliminates the dependence on $\stackrel{^}{{f}_{n}}$ :

$\begin{array}{ccc}\hfill 2E\left[log,\left(\frac{\sqrt{{p}_{\stackrel{^}{{f}_{n}}}\left(Y\right)}}{\sqrt{{p}_{{f}^{*}}\left(Y\right)}},·,\frac{{e}^{-c\left(\stackrel{^}{{f}_{n}}\right)log2}}{\int \sqrt{{p}_{\stackrel{^}{{f}_{n}}}\left(y\right)·{p}_{{f}^{*}}\left(y\right)}\phantom{\rule{0.166667em}{0ex}}\mathrm{d}y}\right)\right]& \le & 2log\left(E,\left[\sum _{f\in \mathcal{F}},{e}^{-c\left(f\right)log2},·,\frac{\sqrt{\frac{{p}_{f}\left(Y\right)}{{p}_{{f}^{*}}\left(Y\right)}}}{\int \sqrt{{p}_{f}\left(y\right)·{p}_{{f}^{*}}\left(y\right)}\phantom{\rule{0.166667em}{0ex}}\mathrm{d}y}\right]\right)\hfill \\ & =& 2log\left(\sum _{f\in \mathcal{F}},{2}^{-c\left(f\right)},\frac{E\left[\sqrt{\frac{{p}_{f}\left(Y\right)}{{p}_{{f}^{*}}\left(Y\right)}}\right]}{\int \sqrt{{p}_{f}\left(y\right)·{p}_{{f}^{*}}\left(y\right)}\phantom{\rule{0.166667em}{0ex}}\mathrm{d}y}\right)\hfill \\ & =& 2log\left(\sum _{f\in \mathcal{F}},{2}^{-c\left(f\right)}\right)\hfill \\ & \le & 0.\hfill \end{array}$

where the last two lines come from

$E\left[\sqrt{\frac{{p}_{f}\left(Y\right)}{{p}_{{f}^{*}}\left(Y\right)}}\right]=\int \sqrt{\frac{{p}_{f}\left(y\right)}{{p}_{{f}^{*}}\left(y\right)}}·{p}_{{f}^{*}}\left(y\right)\phantom{\rule{0.166667em}{0ex}}\mathrm{d}y=\int \sqrt{{p}_{f}\left(y\right)·{p}_{{f}^{*}}\left(y\right)}\phantom{\rule{0.166667em}{0ex}}\mathrm{d}y$

and

$\sum _{f\in \mathcal{F}}{2}^{-c\left(f\right)}\le 1.$

where we get a research paper on Nano chemistry....?
what are the products of Nano chemistry?
There are lots of products of nano chemistry... Like nano coatings.....carbon fiber.. And lots of others..
learn
Even nanotechnology is pretty much all about chemistry... Its the chemistry on quantum or atomic level
learn
da
no nanotechnology is also a part of physics and maths it requires angle formulas and some pressure regarding concepts
Bhagvanji
hey
Giriraj
Preparation and Applications of Nanomaterial for Drug Delivery
revolt
da
Application of nanotechnology in medicine
what is variations in raman spectra for nanomaterials
I only see partial conversation and what's the question here!
what about nanotechnology for water purification
please someone correct me if I'm wrong but I think one can use nanoparticles, specially silver nanoparticles for water treatment.
Damian
yes that's correct
Professor
I think
Professor
Nasa has use it in the 60's, copper as water purification in the moon travel.
Alexandre
nanocopper obvius
Alexandre
what is the stm
is there industrial application of fullrenes. What is the method to prepare fullrene on large scale.?
Rafiq
industrial application...? mmm I think on the medical side as drug carrier, but you should go deeper on your research, I may be wrong
Damian
How we are making nano material?
what is a peer
What is meant by 'nano scale'?
What is STMs full form?
LITNING
scanning tunneling microscope
Sahil
how nano science is used for hydrophobicity
Santosh
Do u think that Graphene and Fullrene fiber can be used to make Air Plane body structure the lightest and strongest. Rafiq
Rafiq
what is differents between GO and RGO?
Mahi
what is simplest way to understand the applications of nano robots used to detect the cancer affected cell of human body.? How this robot is carried to required site of body cell.? what will be the carrier material and how can be detected that correct delivery of drug is done Rafiq
Rafiq
if virus is killing to make ARTIFICIAL DNA OF GRAPHENE FOR KILLED THE VIRUS .THIS IS OUR ASSUMPTION
Anam
analytical skills graphene is prepared to kill any type viruses .
Anam
Any one who tell me about Preparation and application of Nanomaterial for drug Delivery
Hafiz
what is Nano technology ?
write examples of Nano molecule?
Bob
The nanotechnology is as new science, to scale nanometric
brayan
nanotechnology is the study, desing, synthesis, manipulation and application of materials and functional systems through control of matter at nanoscale
Damian
Is there any normative that regulates the use of silver nanoparticles?
what king of growth are you checking .?
Renato
What fields keep nano created devices from performing or assimulating ? Magnetic fields ? Are do they assimilate ?
why we need to study biomolecules, molecular biology in nanotechnology?
?
Kyle
yes I'm doing my masters in nanotechnology, we are being studying all these domains as well..
why?
what school?
Kyle
biomolecules are e building blocks of every organics and inorganic materials.
Joe
how did you get the value of 2000N.What calculations are needed to arrive at it
Privacy Information Security Software Version 1.1a
Good
Got questions? Join the online conversation and get instant answers!