686
L. Lyons
15.9.2 Deep Learning
This involves the use of sophisticated techniques[45] for achieving nearly optimal
extraction of information from data, but which are still relatively unfamiliar to many
scientists. It is important to develop a set of protocols to ensure that they perform in
a reliable manner, and are not introducing subtle biases of which users are unaware.
15.9.3 Unfolding Data or Smearing Theory?
Observed experimental distributions are almost always smeared versions of ‘the true
distributions of Nature’. It is simpler to compare theory and data by smearing the
theory, rather than trying to unfold the experimental effects from the data, as the
latter is a less stable procedure and also introduces correlations among the bins
of the unfolded distribution. Some fields tend to favour deconvolution; this is partly
because it is rarer for them to have a dominant theoretical model with which the data
is to be compared. Unfolding does have the advantage that it provides an estimate
of the ‘true’ distribution, with which any future theory can be compared. Also it can
be looked at by a physicist, but we are not accustomed to readily interpreting data
where the contents of the histogram bins are highly correlated.
There are some situations where unfolding is desirable. For example, it allows the
comparison of distributions from different experiments, with different resolutions.
Another is using experimental data for tuning Monte Carlo generators; smearing the
data at each step of the optimisation increases the computation time too much.
Even for checking in future whether new theories are compatible with data does
not necessarily require unfolding. Provided that the smearing matrix of the detector
is provided, the future data can be smeared, and then compared with the actual (not
unfolded) data. However, including the effects of systematics can be a complication.
Sessions at the 2011 PHYSTAT workshop[9] and at CERN’s PHYSTATν
meeting[12] were devoted to unfolding. Blobel[97] has reviewed the topic, while
ref. [98] contains a statistician’s view of the statistical issues involved in unfolding.
15.9.4 Visualisation
The combination of the human eye and brain is very powerful at detecting patterns
in data (even if sometimes they are not there!) This can be useful in deciding how
to analyse the data; as a check on whether the result of an analysis is plausible;
whether a machine learning method for separating signal from background is
performing sensibly; etc. Such human inspection of data is feasible if there are only
a small number (below 4) of relevant variables. Techniques for inspecting multidimensional data would be valuable.
L. Lyons
15.9.2 Deep Learning
This involves the use of sophisticated techniques[45] for achieving nearly optimal
extraction of information from data, but which are still relatively unfamiliar to many
scientists. It is important to develop a set of protocols to ensure that they perform in
a reliable manner, and are not introducing subtle biases of which users are unaware.
15.9.3 Unfolding Data or Smearing Theory?
Observed experimental distributions are almost always smeared versions of ‘the true
distributions of Nature’. It is simpler to compare theory and data by smearing the
theory, rather than trying to unfold the experimental effects from the data, as the
latter is a less stable procedure and also introduces correlations among the bins
of the unfolded distribution. Some fields tend to favour deconvolution; this is partly
because it is rarer for them to have a dominant theoretical model with which the data
is to be compared. Unfolding does have the advantage that it provides an estimate
of the ‘true’ distribution, with which any future theory can be compared. Also it can
be looked at by a physicist, but we are not accustomed to readily interpreting data
where the contents of the histogram bins are highly correlated.
There are some situations where unfolding is desirable. For example, it allows the
comparison of distributions from different experiments, with different resolutions.
Another is using experimental data for tuning Monte Carlo generators; smearing the
data at each step of the optimisation increases the computation time too much.
Even for checking in future whether new theories are compatible with data does
not necessarily require unfolding. Provided that the smearing matrix of the detector
is provided, the future data can be smeared, and then compared with the actual (not
unfolded) data. However, including the effects of systematics can be a complication.
Sessions at the 2011 PHYSTAT workshop[9] and at CERN’s PHYSTATν
meeting[12] were devoted to unfolding. Blobel[97] has reviewed the topic, while
ref. [98] contains a statistician’s view of the statistical issues involved in unfolding.
15.9.4 Visualisation
The combination of the human eye and brain is very powerful at detecting patterns
in data (even if sometimes they are not there!) This can be useful in deciding how
to analyse the data; as a check on whether the result of an analysis is plausible;
whether a machine learning method for separating signal from background is
performing sensibly; etc. Such human inspection of data is feasible if there are only
a small number (below 4) of relevant variables. Techniques for inspecting multidimensional data would be valuable.