Giu/analistica
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

ex-7: review

Browse Source
This commit is contained in:
Michele Guerini Rocco 2020-06-01 23:28:44 +02:00
parent ed73ea5c8b
commit e7a5185976
1 changed files with 153 additions and 149 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

302
notes/sections/7.md
View File

@ -1,51 +1,50 @@
# Exercise 7
## Generating points according to Gaussian distributions {#sec:sampling}
Two sets of 2D points $(x, y)$ - signal and noise - is to be generated according
to two bivariate Gaussian distributions with parameters:
$$
\text{signal} \quad
\begin{cases}
\mu = (0, 0) \\
\sigma_x = \sigma_y = 0.3 \\
\rho = 0.5
\end{cases}
\et
\text{noise} \quad
\begin{cases}
\mu = (4, 4) \\
\sigma_x = \sigma_y = 1 \\
\rho = 0.4
\end{cases}
$$
## Generating random points on the plane {#sec:sampling}
Two sets of 2D points $(x, y)$ --- signal and noise --- is to be generated
according to two bivariate Gaussian distributions with parameters:
\begin{align*}
\text{signal}\:
\begin{cases}
\mu = (0, 0) \\
\sigma_x = \sigma_y = 0.3 \\
\rho = 0.5
\end{cases}
&&
\text{noise}\:
\begin{cases}
\mu = (4, 4) \\
\sigma_x = \sigma_y = 1 \\
\rho = 0.4
\end{cases}
\end{align*}
where $\mu$ stands for the mean, $\sigma_x$ and $\sigma_y$ for the standard
deviations in $x$ and $y$ directions respectively and $\rho$ is the bivariate
correlation, namely:
$$
\sigma_{xy} = \rho \sigma_x \sigma_y
\sigma_{xy} = \rho\, \sigma_x \sigma_y
$$
where $\sigma_{xy}$ is the covariance of $x$ and $y$.
In the code, default settings are $N_s = 800$ points for the signal and $N_n =
1000$ points for the noise but can be customized from the input command-line.
Both samples were handled as matrices of dimension $n$ x 2, where $n$ is the
number of points in the sample. The library `gsl_matrix` provided by GSL was
employed for this purpose and the function `gsl_ran_bivariate_gaussian()` was
used for generating the points.
In the programs, $N_s = 800$ points for the signal and $N_n = 1000$ points for
the noise were chosen as default but can be customized from the command-line.
Both samples were stored as $n \times 2$ matrices, where $n$ is the number of
points in the sample. The library `gsl_matrix` provided by GSL was employed for
this purpose and the function `gsl_ran_bivariate_gaussian()` was used for
generating the points.
An example of the two samples is shown in @fig:points.
![Example of points sampled according to the two Gaussian distributions
with the given parameters.](images/7-points.pdf){#fig:points}
Assuming not to know how the points were generated, a model of classification
is then to be implemented in order to assign each point to the right class
(signal or noise) to which it 'most probably' belongs to. The point is how
'most probably' can be interpreted and implemented.
Here, the Fisher linear discriminant and the Perceptron were implemented and
Assuming to not know how the points were generated, a model of classification,
which assign to each point the 'most probably' class it belongs to,
is implemented. Depending on the interpretation of 'most probable' many
different models can be developed.
Here, the Fisher linear discriminant and the perceptron were implemented and
described in the following two sections. The results are compared in
@sec:7_results.
@sec:class-results.
## Fisher linear discriminant
@ -54,24 +53,27 @@ described in the following two sections. The results are compared in
### The projection direction
The Fisher linear discriminant (FLD) is a linear classification model based on
dimensionality reduction. It allows to reduce this 2D classification problem
into a one-dimensional decision surface.
dimensionality reduction. It does so by projecting the data onto hyperplanes
that best divide the classes of points, consequently decreasing the dimension
to $n-1$. In the 2D case the projection is onto a line, therefore the problem
is reduced to simply selecting a threshold.
Consider the case of two classes (in this case signal and noise): the simplest
representation of a linear discriminant is obtained by taking a linear function
$\hat{x}$ of a sampled 2D point $x$ so that:
$\tilde{x}$ of a sampled 2D point $x$ so that:
$$
\hat{x} = w^T x
\tilde{x} = w^T x
$$
where $w$ is the so-called 'weight vector' and $w^T$ stands for its transpose.
An input point $x$ is commonly assigned to the first class if $\hat{x} \geqslant
An input point $x$ is commonly assigned to the first class if $\tilde{x} >
w_{th}$ and to the second one otherwise, where $w_{th}$ is a threshold value
somehow defined. In general, the projection onto one dimension leads to a
somehow defined. In general, the projection onto one dimension leads to a
considerable loss of information and classes that are well separated in the
original 2D space may become strongly overlapping in one dimension. However, by
adjusting the components of the weight vector, a projection that maximizes the
classes separation can be selected [@bishop06].
classes separation can be found [@bishop06].
To begin with, consider $N_1$ points of class $C_1$ and $N_2$ points of class
$C_2$, so that the means $\mu_1$ and $\mu_2$ of the two classes are given by:
$$
@ -83,46 +85,45 @@ $$
The simplest measure of the separation of the classes is the separation of the
projected class means. This suggests to choose $w$ so as to maximize:
$$
\hat{\mu}_2 − \hat{\mu}_1 = w^T (\mu_2 − \mu_1)
\tilde{\mu}_2 − \tilde{\mu}_1 = w^T (\mu_2 − \mu_1)
$$
This expression can be made arbitrarily large simply by increasing the magnitude
of $w$. To solve this problem, $w$ can be constrained to have unit length, so
that $| w^2 | = 1$. Using a Lagrange multiplier to perform the constrained
maximization, it can be found that $w \propto (\mu_2 − \mu_1)$, meaning that the
line onto the points must be projected is the one joining the class means.
This expression can be made arbitrarily large simply by increasing the
magnitude of $w$, fortunately the problem is easily solved by requiring $w$
to be normalised: $| w^2 | = 1$. Using a Lagrange multiplier to perform the
constrained maximization, it can be found that $w \propto (\mu_2 − \mu_1)$,
meaning that the line onto the points must be projected is the one joining the
class means.
There is still a problem with this approach, however, as illustrated in
@fig:overlap: the two classes are well separated in the original 2D space but
have considerable overlap when projected onto the line joining their means
which maximize their projections distance.
![The plot on the left shows samples from two classes along with the
histograms resulting fromthe projection onto the line joining the
class means: note that there is considerable overlap in the projected
histograms resulting from the projection onto the line joining the
class means: note the considerable overlap in the projected
space. The right plot shows the corresponding projection based on the
Fisher linear discriminant, showing the greatly improved classes
separation. Fifure from [@bishop06]](images/7-fisher.png){#fig:overlap}
separation. Figure taken from [@bishop06]](images/7-fisher.png){#fig:overlap}
The idea to solve it is to maximize a function that will give a large separation
between the projected classes means while also giving a small variance within
each class, thereby minimizing the class overlap.
The within-class variance of the transformed data of each class $k$ is given
by:
The overlap of the projections can be reduced by maximising a function that
gives, besides a large separation, small variance within each class. The
within-class variance of the transformed data of each class $k$ is given by:
$$
\hat{s}_k^2 = \sum_{n \in c_k} (\hat{x}_n - \hat{\mu}_k)^2
\tilde{s}_k^2 = \sum_{n \in c_k} (\tilde{x}_n - \tilde{\mu}_k)^2
$$
The total within-class variance for the whole data set is simply defined as
$\hat{s}^2 = \hat{s}_1^2 + \hat{s}_2^2$. The Fisher criterion is defined to
be the ratio of the between-class distance to the within-class variance and is
given by:
$\tilde{s}^2 = \tilde{s}_1^2 + \tilde{s}_2^2$. The Fisher criterion is therfore
defined to be the ratio of the between-class distance to the within-class
variance and is given by:
$$
F(w) = \frac{(\hat{\mu}_2 - \hat{\mu}_1)^2}{\hat{s}^2}
F(w) = \frac{(\tilde{\mu}_2 - \tilde{\mu}_1)^2}{\tilde{s}^2}
$$
The dependence on $w$ can be made explicit:
\begin{align*}
(\hat{\mu}_2 - \hat{\mu}_1)^2 &= (w^T \mu_2 - w^T \mu_1)^2 \\
(\tilde{\mu}_2 - \tilde{\mu}_1)^2 &= (w^T \mu_2 - w^T \mu_1)^2 \\
&= [w^T (\mu_2 - \mu_1)]^2 \\
&= [w^T (\mu_2 - \mu_1)][w^T (\mu_2 - \mu_1)] \\
&= [w^T (\mu_2 - \mu_1)][(\mu_2 - \mu_1)^T w]
@ -131,9 +132,9 @@ The dependence on $w$ can be made explicit:
where $M$ is the between-distance matrix. Similarly, as regards the denominator:
\begin{align*}
\hat{s}^2 &= \hat{s}_1^2 + \hat{s}_2^2 = \\
&= \sum_{n \in c_1} (\hat{x}_n - \hat{\mu}_1)^2
+ \sum_{n \in c_2} (\hat{x}_n - \hat{\mu}_2)^2
\tilde{s}^2 &= \tilde{s}_1^2 + \tilde{s}_2^2 = \\
&= \sum_{n \in c_1} (\tilde{x}_n - \tilde{\mu}_1)^2
+ \sum_{n \in c_2} (\tilde{x}_n - \tilde{\mu}_2)^2
= w^T \Sigma_w w
\end{align*}
@ -162,7 +163,7 @@ Differentiating with respect to $w$, it can be found that $F(w)$ is maximized
when:
$$
w = \Sigma_w^{-1} (\mu_2 - \mu_1)
$$
$$ {#eq:fisher-weight}
This is not truly a discriminant but rather a specific choice of the direction
for projection of the data down to one dimension: the projected data can then be
@ -178,63 +179,73 @@ with the `gsl_blas_dgemv()` function provided by GSL.
### The threshold
The threshold $t_{\text{cut}}$ was fixed by the condition of conditional
The threshold $t_{\text{cut}}$ was fixed by the conditional
probability $P(c_k | t_{\text{cut}})$ being the same for both classes $c_k$:
$$
t_{\text{cut}} = x \, | \hspace{20pt}
t_{\text{cut}} = x \text{ such that}\quad
\frac{P(c_1 | x)}{P(c_2 | x)} =
\frac{P(x | c_1) \, P(c_1)}{P(x | c_2) \, P(c_2)} = 1
$$
where $P(x | c_k)$ is the probability for point $x$ along the Fisher projection
line of being sampled according to the class $k$. If each class is a bivariate
Gaussian, as in the present case, then $P(x | c_k)$ is simply given by its
projected normal distribution with mean $\hat{m} = w^T m$ and variance $\hat{s}
= w^T S w$, being $S$ the covariance matrix of the class.
With a bit of math, the following solution can be found:
line of being sampled from $k$. If $\tilde{x} > t_\text{cut}$ then more likely
$x \in c_1$, otherwise $x \in c_2$.
If each class is a bivariate Gaussian distribution, as in the present case,
then $P(x | c_k)$ is simply given by its projected normal distribution with
mean $\tilde{m} = w^T m$ and variance $\tilde{s} = w^T S w$, being $S$ the
covariance matrix of the class.
After some algebra, the threshold is found to be:
$$
t_{\text{cut}} = \frac{b}{a}
+ \sqrt{\left( \frac{b}{a} \right)^2 - \frac{c}{a}}
$$
where:
- $a = \hat{s}_1^2 - \hat{s}_2^2$
- $b = \hat{\mu}_2 \, \hat{s}_1^2 - \hat{\mu}_1 \, \hat{s}_2^2$
- $c = \hat{\mu}_2^2 \, \hat{s}_1^2 - \hat{\mu}_1^2 \, \hat{s}_2^2
- 2 \, \hat{s}_1^2 \, \hat{s}_2^2 \, \ln(\alpha)$
- $a = \tilde{s}_1^2 - \tilde{s}_2^2$
- $b = \tilde{\mu}_2 \, \tilde{s}_1^2 - \tilde{\mu}_1 \, \tilde{s}_2^2$
- $c = \tilde{\mu}_2^2 \, \tilde{s}_1^2 - \tilde{\mu}_1^2 \, \tilde{s}_2^2
- 2 \, \tilde{s}_1^2 \, \tilde{s}_2^2 \, \ln(\alpha)$
- $\alpha = P(c_1) / P(c_2)$
The ratio of the prior probabilities $\alpha$ is simply given by:
In a simulation, the ratio of the prior probabilities $\alpha$ can
simply be set to:
$$
\alpha = \frac{N_s}{N_n}
$$
The projection of the points was accomplished by the use of the function
`gsl_blas_ddot()`, which computes the element wise product between two vectors.
`gsl_blas_ddot()`, which computes a fast dot product of two vectors.
Results obtained for the same samples in @fig:points are shown in
@fig:fisher_proj. The weight vector and the treshold were found to be:
Results obtained for the same samples in @fig:points are shown below in
@fig:fisher-proj. The weight vector and the threshold were found to be:
$$
w = (0.707, 0.707) \et
t_{\text{cut}} = 1.323
$$
<div id="fig:fisher_proj">
![View of the samples in the plane.](images/7-fisher-plane.pdf)
![View of the samples projections onto the projection
line.](images/7-fisher-proj.pdf)
::: { id=fig:fisher-proj }
![Scatter plot of the samples.](images/7-fisher-plane.pdf)
![Histogram of the Fisher-projected samples.](images/7-fisher-proj.pdf)
Aerial and lateral views of the samples. Projection line in blu and cut in red.
</div>
:::
Since the vector $w$ turned out to be parallel with the line joining the means
of the two classes (reminded to be $(0, 0)$ and $(4, 4)$), one can be mislead
and assume that the inverse of the total covariance matrix $\Sigma_w$ is
isotropic, namely proportional to the unit matrix.
That's not true. In this special sample, the vector joining the means turns out
to be an eigenvector of the covariance matrix $\Sigma_w^{-1}$. In fact: since
$\sigma_x = \sigma_y$ for both signal and noise:
### A mathematical curiosity
This section is really a sidenote which grew too large to fit in a margin,
so it can be safely skipped.
It can be seen that the weight vector turned out to parallel to the line
joining the means of the two classes (as a remainder: $(0, 0)$ and $(4, 4)$),
as if the within-class covariances were ignored. Strange!
Looking at @eq:fisher-weight, one can be mislead into thinking that the inverse
of the total covariance matrix, $\Sigma_w$ is (proportional to) the identity,
but that's not true. By a remarkable accident, the vector joining the means is
an eigenvector of the covariance matrix $\Sigma_w^{-1}$. In
fact: since $\sigma_x = \sigma_y$ for both signal and noise:
$$
\Sigma_1 = \begin{pmatrix}
\sigma_x^2 & \sigma_{xy} \\
@ -247,15 +258,14 @@ $$
\end{pmatrix}_2
$$
$\Sigma_w$ takes the form:
$\Sigma_w$ takes the symmetrical form
$$
\Sigma_w = \begin{pmatrix}
A & B \\
B & A
\end{pmatrix}
\end{pmatrix},
$$
Which can be easily inverted by Gaussian elimination:
which can be easily inverted by Gaussian elimination:
\begin{align*}
\begin{pmatrix}
A & B & \vline & 1 & 0 \\
@ -271,7 +281,7 @@ Which can be easily inverted by Gaussian elimination:
\end{pmatrix}
\end{align*}
Hence:
Hence, the inverse has still the same form:
$$
\Sigma_w^{-1} = \begin{pmatrix}
\tilde{A} & \tilde{B} \\
@ -279,21 +289,15 @@ $$
\end{pmatrix}
$$
Thus, $\Sigma_w$ and $\Sigma_w^{-1}$ share the same eigenvectors $v_1$ and
$v_2$:
For this reason, $\Sigma_w$ and $\Sigma_w^{-1}$ share the same eigenvectors
$v_1$ and $v_2$:
$$
v_1 = \begin{pmatrix}
1 \\
-1
\end{pmatrix} \et
v_2 = \begin{pmatrix}
1 \\
1
\end{pmatrix}
v_1 = \begin{pmatrix} 1 \\ -1 \end{pmatrix}
\et
v_2 = \begin{pmatrix} 1 \\ 1 \end{pmatrix}
$$
and the vector joining the means is clearly a multiple of $v_2$, causing $w$ to
be a multiple of it.
The vector joining the means is clearly a multiple of $v_2$, and so is $w$.
## Perceptron
@ -311,16 +315,18 @@ The aim of the perceptron algorithm is to determine the weight vector $w$ and
bias $b$ such that the so-called 'threshold function' $f(x)$ returns a binary
value: it is expected to return 1 for signal points and 0 for noise points:
$$
f(x) = \theta(w^T \cdot x + b)
f(x) = \theta(w^T x + b)
$$ {#eq:perc}
where $\theta$ is the Heaviside theta function.
where $\theta$ is the Heaviside theta function. Note that the bias $b$ is
$-t_\text{cut}$, as defined in the previous section.
The training was performed using the generated sample as training set. From an
initial guess for $w$ and $b$ (which were set to be all null in the code), the
perceptron starts to improve their estimations. The training set is passed point
by point into a iterative procedure a customizable number $N$ of times: for
every point, the output of $f(x)$ is computed. Afterwards, the variable
$\Delta$, which is defined as:
perceptron starts to improve their estimations. The training set is passed
point by point into a iterative procedure $N$ times: for every point, the
output of $f(x)$ is computed. Afterwards, the variable $\Delta$, which is
defined as:
$$
\Delta = r [e - f(x)]
$$
@ -355,7 +361,7 @@ To see how it works, consider the four possible situations:
the current $b$ and $w$ overestimate the correct output: they must be
decreased.
Whilst the $b$ updating is obvious, as regarsd $w$ the following consideration
Whilst the $b$ updating is obvious, as regards $w$ the following consideration
may help clarify. Consider the case with $e = 0 \quad \wedge \quad f(x) = 1
\quad \Longrightarrow \quad \Delta = -1$:
$$
@ -366,53 +372,51 @@ $$
Similarly for the case with $e = 1$ and $f(x) = 0$.
![Weiht vector and threshold value obtained with the perceptron method as a
![Weight vector and threshold value obtained with the perceptron method as a
function of the number of iterations. Both level off at the third
iteration.](images/7-iterations.pdf){#fig:iterations}
iteration.](images/7-iterations.pdf){#fig:percep-iterations}
As far as convergence is concerned, the perceptron will never get to the state
with all the input points classified correctly if the training set is not
linearly separable, meaning that the signal cannot be separated from the noise
by a line in the plane. In this case, no approximate solutions will be gradually
approached. On the other hand, if the training set is linearly separable, it can
be shown that this method converges to the coveted function [@novikoff63].
As in the previous section, once found, the weight vector is to be normalized.
by a line in the plane. In this case, no approximate solutions will be
gradually approached. On the other hand, if the training set is linearly
separable, it can be shown (see [@novikoff63]) that this method converges to
the coveted function. As in the previous section, once found, the weight
vector is to be normalized.
With $N = 5$ iterations, the values of $w$ and $t_{\text{cut}}$ level off up to the third
digit. The following results were obtained:
Different values of the learning rate were tested, all giving the same result,
converging for a number $N = 3$ of iterations. In @fig:iterations, results are
converging for a number $N = 3$ of iterations. In @fig:percep-iterations, results are
shown for $r = 0.8$: as can be seen, for $N = 3$, the values of $w$ and
$t^{\text{cut}}$ level off.
The following results were obtained:
$$
w = (0.654, 0.756) \et t_{\text{cut}} = 1.213
w = (-0.654, -0.756) \et t_{\text{cut}} = 1.213
$$
In this case, the projection line is not parallel with the line joining the
means of the two samples. Plots in @fig:percep_proj.
In this case, the projection line is not exactly parallel with the line joining
the means of the two samples. Plots in @fig:percep-proj.
<div id="fig:percep_proj">
![View from above of the samples.](images/7-percep-plane.pdf)
![Gaussian of the samples on the projection
line.](images/7-percep-proj.pdf)
::: { id=fig:percep-proj }
![Scatter plot of the samples.](images/7-percep-plane.pdf)
![Histogram of the projected samples.](images/7-percep-proj.pdf)
Aerial and lateral views of the projection direction, in blue, and the cut, in
red.
</div>
Aerial and lateral views of the samples. Projection line in blu and cut in red.
:::
## Efficiency test {#sec:7_results}
## Efficiency test {#sec:class-results}
Using the same parameters of the training set, a number $N_t$ of test
samples was generated and the points were divided into noise and signal
applying both methods. To avoid storing large datasets in memory, at each
iteration, false positives and negatives were recorded using a running
statistics method implemented in the `gsl_rstat` library. For each sample, the
numbers $N_{fn}$ and $N_{fp}$ of false negative and false positive were obtained
this way: for every noise point $x_n$, the threshold function $f(x_n)$ was
computed, then:
Using the same parameters of the training set, a number $N_t$ of test samples
was generated and the points were classified applying both methods. To avoid
storing large datasets in memory, at each iteration, false positives and
negatives were recorded using a running statistics method implemented in the
`gsl_rstat` library. For each sample, the numbers $N_{fn}$ and $N_{fp}$ of
false negative and false positive were obtained this way: for every noise point
$x_n$, the threshold function $f(x_n)$ was computed, then:
- if $f(x) = 0 \thus$ $N_{fn} \to N_{fn}$
- if $f(x) \neq 0 \thus$ $N_{fn} \to N_{fn} + 1$
@ -438,13 +442,13 @@ solution, the most powerful one, according to the Neyman-Pearson lemma, whereas
the perceptron is only expected to converge to the solution and is therefore
more subject to random fluctuations.
-------------------------------------------------------------------------------------------
$\alpha$ $\sigma_{\alpha}$ $\beta$ $\sigma_{\beta}$
----------- ------------------- ------------------- ------------------- -------------------
Fisher 0.9999 0.33 0.9999 0.33
------------------------------------------------------
$α$ $σ_α$ $β$ $σ_β$
----------- ---------- ---------- ---------- ---------
Fisher 0.9999 0.33 0.9999 0.33
Perceptron 0.9999 0.28 0.9995 0.64
-------------------------------------------------------------------------------------------
Perceptron 0.9999 0.28 0.9995 0.64
------------------------------------------------------
Table: Results for Fisher and perceptron method. $\sigma_{\alpha}$ and
$\sigma_{\beta}$ stand for the standard deviation of the false