| 1 |
An artificial neural network maps a point in the space of input observables to
|
| 2 |
some value of neural network output $x$. The neural network training error is
|
| 3 |
given by equation~\ref{eq:NNerrorFunc}. A given point in the vector space
|
| 4 |
spanned by the neural network input observables (denoted as ``feature space'')
|
| 5 |
contributes to the neural network training error $E$ by
|
| 6 |
\begin{equation}
|
| 7 |
E' = (1 - x)^2\cdot\rho^\tau + x^2\cdot\rho^{QCD}
|
| 8 |
\end{equation}
|
| 9 |
where $\rho^\tau (\rho^{QCD})$ denotes the training sample density of the
|
| 10 |
$\tau$ signal and QCD--jet background at that point in feature space.
|
| 11 |
|
| 12 |
The value $x$ assigned by the neural network to this region in feature space
|
| 13 |
should satisfy the requirement of minimal error:
|
| 14 |
\begin{align}
|
| 15 |
\frac{\partial E'}{\partial x} &= 0 \nonumber \\
|
| 16 |
0 &= -2(1-x)\cdot\rho^\tau+2x\cdot\rho^{QCD} \nonumber \\
|
| 17 |
x &= \frac{\rho^\tau} {\rho^\tau + \rho^{QCD}} \label{eq:probFracToX} \\
|
| 18 |
\rho^\tau &= x(\rho^\tau + \rho^{QCD}) \nonumber \\
|
| 19 |
\frac{\rho^{QCD}}{\rho^\tau} &= \frac{1}{x} - 1 \label{eq:rawTransformX}
|
| 20 |
\end{align}
|
| 21 |
|
| 22 |
N.B. that the ratio $\frac{\rho^{QCD}}{\rho^\tau}$ corresponds to the ratio of
|
| 23 |
the normalized probability density functions of signal and background input
|
| 24 |
observable distributions, i.e. $\int \rho^{\tau} d\vec x = 1$.
|
| 25 |
|
| 26 |
In the case of multiple neural networks, one can derive a formula that maps the
|
| 27 |
output $x_j$ of the neural network corresponding to decay mode $j$ according to
|
| 28 |
the ``prior probabilities'' $p_j^\tau (p_j^{QCD})$ for true $\tau$ lepton
|
| 29 |
hadronic decays (quark and gluon jets) to pass the preselection criteria and
|
| 30 |
be reconstructed with decay mode $j$.
|
| 31 |
|
| 32 |
By substituting $\rho^s \rightarrow \rho^s p_j^s$ for $s \in \{\tau, QCD\}$ in
|
| 33 |
equation~\ref{eq:probFracToX}, the output $x_j$ can be related to $p_j^\tau
|
| 34 |
(p_j^{QCD})$ by
|
| 35 |
\begin{equation}
|
| 36 |
x_j' = \frac{\rho^\tau \cdot p_j^\tau}
|
| 37 |
{\rho^\tau \cdot p_j^\tau + \rho^{QCD} \cdot p_j^{QCD} }
|
| 38 |
= \frac{p_j^\tau}
|
| 39 |
{p_j^\tau + \frac{\rho^{QCD}}{\rho^\tau} \cdot p_j^{QCD} }
|
| 40 |
\label{eq:probFracToXWithPriors}
|
| 41 |
\end{equation}
|
| 42 |
|
| 43 |
Substituting equation~\ref{eq:rawTransformX} into
|
| 44 |
equation~\ref{eq:probFracToXWithPriors} yields the transformation of the output
|
| 45 |
$x_j$ of the neural neural network corresponding to any selected decay mode $j$
|
| 46 |
to a single discriminator output $x_j'$ which for a given point on the optimal
|
| 47 |
performance curve should be independent of $j$.
|
| 48 |
|
| 49 |
\begin{equation}
|
| 50 |
x_j' = \frac{p_j^\tau}
|
| 51 |
{p_j^\tau + \left(\frac{1}{x_j}-1\right)\cdot p_j^{QCD} }
|
| 52 |
\end{equation}
|
| 53 |
|
| 54 |
|
