Minutes of the Project coordinators meeting on October 28th, 2019 (DRAFT 2)

1. Part A. 

[Agenda item #1] Presentation "Deep learning for diffraction imaging" by Julian
Zimmermann (Project 2.1)

The presentation discussed strategies for using neural networks for classifying
high-volume diffraction images, using the example of the data for liquid-helium
droplets collected at FERMI (lambda=32-65nm).  The dataset contains a very
large number of shapes and sizes, making traditional classification techniques
impractical.  Supervised-learning techniques for a convolutional neural network
were described first. The discussion of unsupervised-learning approach and the
comparison of the two approaches followed.

The presentation was followed by a lively discussion.

Dr Kornilov asked what is the final goal of the investigation? Is it a general
diffraction-image analysis system, or a system specific to helium droplets?

Mr. Zimmermann replied that system could in principle work with any images, but
may require additional work to achieve this capability. In this respect,
training the neural networks with simulated data may be more useful that using
manually-classified data in the training step.

Dr. Rouzée asked what is the next step in this investigation? Can it be
combined with the known physics (e.g. Mie functions)?

Mr. Zimmermann replied that at the present time, the goal of the system is
image classification, rather than extracting physical parameters. However, some
physical parameters have very clear signatures in the images, which are
automatically detected by the neural network. For example, both sizes and
shapes can be directly inferred from the network output.

Dr Nibbering asked whether the classification produced by the network is
unique? Is there any guarantee of correctness? Is it possible for the training
to terminate at a local optimum?

Mr. Zimmermann replied that generally, no guarantee of uniqueness or
correctness exists, since the process is black-box by design. It is also
generally not possible to to predict what is the useful information content of
the output, and no general mathematical guarantees are available for the
results.

Prof. Elsaesser asked what is the minimum ellipticity what can be detected by
this technique? The entropy approach to the loss function seems similar to some
of the techniques used in diffraction imaging, which are known to work well as
long as the image resolution is not too high.

Mr. Zimmermann replies that the neural networks are indeed sensitive to the
image resolution, and need to be specifically trained for each resolution.

Dr. Kornilov asked how could the problem-specific data be added to the network
for the subsequent analysis? It is often known what the physics is!

Mr. Zimmermann replied that no special measures need to be taken; the neural
network learns about the relevant problem-specific physics automatically.

Dr. Schnuerer asked what does the network do with data which does not belong
to any data set (e.g. due to collection glitches)?

Mr. Zimmermann replied that networks trained using the supervised-learning
approach are very good at rejecting such data. It may become problematic for
the non-supervised networks.

Dr. Schnuerer asked what would happen if the network were to asked to classify
a series of experimental images of the same object?

Mr. Zimmermann replied that this wasn't tried.

1. Part B. 

[Agenda item #2] Update on the audit "Beruf und Familie" by Dr. Alexander
Grimm.

Dr. Grimm first presented a brief retrospective of the MBI certification under
the "Beruf und Familie" program. The initial certification was achieved in
September 2015, followed by re-certification in September 2018. On 25.09.2019
the 1st report for the re-audit was submitted. The two main issues addressed in
the report are mobile work for family reasons and and the role of the appraisal
interviews (Mitarbeitergesprächen). Two internal workshops were conducted, with
the following outcomes:

For the "mobile work" subject, the current practice of managing the mobile work
will be described and communicated by the managing director to the staff. It is
further necessary to discuss the experiences with mobile work between staff
and the executive members on a regular basis.

For the "Mitarbeitergesprach" subject, the guidelines including the list of
typical issues to be discussed will be provided to all staff members. It will
be made clear that every member of the staff has a the possibility to request
an appraisal interview.

The issue of the family-friendly leadership still has not been discussed. For
this issue, an internal workshop is planned within the first quarter of 2020.

The slides of the presentation by Dr. Grimm are available at (MBI intranet
only):
http://intern.mbi-berlin.de/en/organization/projectcoordinators/2019-10-28/Presentation_Audit_28_10_2019.pdf

The report was followed by a discussion.

Prof. Elsaesser remarked that the appraisal interview process looks very
bureaucratic. It is hard to formalize this process, and it may hamper the
informal communication. We should try not to introduce extra formal steps in
the interaction between staff and the executive members.

Prof Eisebitt replied that this is exactly the reason why the proposed
appraisal interview process has a lot of flexibility. The questions and topic
are only advisory, and exist only to suggest the possibilities to the staff
members.

[Agenda item #3] Other items

[Item #3.1] Purchasing and the new purchasing platform - Prof. Eisebitt

Prof. Eisebitt briefly described the ongoing, long project for the
implementation of the FVB-wide electronic purchasing system. MBI has its own
electronic purchasing system, which predated the FVB effort, and remains the
only institute will not part of the FVB system. The two operational aspects
affected by the FVB-wide purchasing system are a) The ordering process; and b)
Recording of the purchasing process.

At the end of October, there has been a presentation of the FVB system to
selected members of the staff. The system appears very complicated, and
requires complex interaction (a lots of "clicks"). The system was tried by
three members of the staff (Dr. Kornilov, Dr. Schnuerer, and ?????). The test
involved listening to the presentation, followed by creating a test order using
the system.

It is now necessary to decide whether MBI will switch to the new system, or
remain with the current one. For technical reasons, the switch must occur at
the end of the calendar year.

Prof. Eisebitt asked the present members of the test group to comment on their
experiences.

Dr. Kornilov reported that the at the end of the 3-hour workshop (9:30 to
12:30) he was able to successfully create a test order. Dr. Kornilov attributed
the success to his prior experience as a computer programmer, which allowed him
to infer the underlying logic behind the interface. He expressed the doubt that
he would have succeeded otherwise. In his opinion, the system is not designed
for research purchasing. Rather, the intent is to support catalogue shopping.
The interface presents an excessive number of fields and options, with the
meaning and intent not clear to scientists. As the result, it will likely be
impossible for scientists to work with the system. Rather, it would be
necessary to rely on specially-trained employees to perform the ordering, based
on instructions for scientists. Furthermore, the system does not allow for
partially-completed orders.  In Dr. Kornilov's opinion, the FVB purchasing
system will not result in any time savings compared to the existing system.

Dr. Schnuerer reported that he has also managed to place a test order, although
it took him 2x longer than for Dr. Kornilov. One could be trained to use the
system, but it is necessary to know exactly which part of the system to enter
the information in.  In principle this is supposed to be an interactive
platform, but this appears to be complicated in practice, and it is not clear
how well the interactive component will work.  There is also a question of the
external access: the present FVB system allows external access from anywhere in
the world. This access is protected by a username/password combination, but
there are no possibilities of any additional safeguards. This situation is
supposed to be adequate, since issuing purchase orders requires manual
approval.

Prof. Eisebitt asked whether MBI staff can be productive with this system?

Dr. Woerner further asked what is the advantage of the FVB system for MBI?

Prof. Eisebitt replied that the main advantage is for the administrative staff,
for whom it simplifies the process. The system also automatically generates
purchasing audit logs, and supports automatic generation of the offers.

Prof. Vrakking replied that the FVB system has the advantage of collating all 
purchasing data together.

Prof. Eisebitt further remarked that the advantages for MBI appear to be
minimal, with the in-house electronic-purchasing system already implementing
much of the functionality in an easier to use form.

Dr. Nagy commented that for research purchasing, the quotations are rarely for
an exactly the same product or service from different offerors. As the result,
excessively formalizing the process for non-standard items will be problematic.

Prof. Elsaesser replied that this is why the FVB system contains an alternative
process, where the quotes are requested by the originator, and then loaded into 
the system.

Dr. Nagy asked whether we could hear about the experiences with the FVB system 
from some of the existing users?

Prof. Eisebitt (?) replied that it is not obvious how to do this: most
purchasing at other institutes is for the standard items, and is not directly
relevant to MBI.

Prof. Elsaesser commented that the overall experience at the central FVB
administration appears negative. The system appears to be too complicated for
much of the admin staff, which necessitated that purchasing is done through the
purchasing department, rather than directly through the purchasing system.

Dr. Nagy asked what is the the downside of not switching to the FVB system at
this time?

Prof. Eisebitt replied that no downside exists at this time. However, using the
system may become a legal requirement at some point in the future.

Dr. Kornilov noted that the FVB system may be preferable for the B2B contracts
("catalogue shopping").

Dr. Will asked whether the FVB system undergoes development, and whether it is
reasonable to expect that it will be improved for us in the future?

Prof. Eisebitt replied that this is a slightly-customized off-the-shelf system.
Further customization will be slow, and may even be impossible.

Prof. Eisebitt summarized the discussion, by stating that the comments overall
appear to be negative. As the result, the system will not be introduced at the
end of this year.

Dr. Kornilov further commented that a test account on the system is available.
Anybody who wants to try it should contact Dr. Kornilov directly.

Prof. Elsaesser further commented that the switch to the new system would need
to start no later that Nov. 15, due to the need for multiple training courses.

[Item #3.2] Future of the central workshop - Prof. Eisebitt

There has been a long discussion of how to improve the interaction with the
central workshop at MBI. In most physics institutes there exist central
workshops, which can be things better and faster than either the small lab
workshops or external companies. This does not seem to be the experience at
MBI. As the result, there is a proposal of strengthening the small laboratory
workshops, and contracting external companies for larger jobs.

Dr. Will asked what would be the effect on the turn-around time?

Prof. Eisebitt replied that fast turn-around may be hard to achieve with an
external company. The proposed change is a trial; no machines will be
eliminated at the central workshop at this time.

Dr. Nagy asked what exactly is planned for the central workshop?

Dr. Steinmeyer further asked how many people are currently at the central
workshop?

Prog. Vrakking replied that many staff members at the central workshop are
close to retirement. Next year, all but one are expected to retire. This is
about changing the model: administratively, all staff will be reassigned to
specific departments.

Dr. Nagy remarked that there is a risk that there will be nobody left to
maintain and operate the machines currently at the central workshop.

Prof. Vrakking replied that we need to take a look at which machines should be
kept. This has not been decided yet.

Dr. Schnuerer commented that one could be trained to use these machines, but
maintaining this competency requires constant usage. Otherwise, periodic
retraining would be required. Furthermore, having multiple people using the
machine, with no one ultimately responsible for it, may be problematic.

Dr. Nibbering commented that there are two discussions going on, namely the
questions of the administrative organization and the access to and maintenance
of the machines. These are separate questions.

Prof. Vrakking replied that the present discussion is the outcome of a long
process, going on for the last 5-6 years, involving massive complaints about
the machine shop. It is also a question of how much the workshop costs relative
to the external companies. The directors have asked Dr. Kornilov to perform the
cost analysis. According the the analysis, the in-house workshop costs about 4
times more than an external company for the same parts. The MBI budget is under
high pressure; it is necessary to look at the costs and possible savings. At
this time, running the central workshop costs EUR 0.5 million per year.

[Item #3.3] MBI evaluation report - Prof. Eisebitt

The preliminary evaluation report has been received by MBI. The report remains
confidential until approved by the Senate of the Leibniz Association (the next
sitting is planned for March 25th, 2020). All members of MBI can see the
hardcopy of the draft report, available in the directors' offices.

The meeting adjourned.
 
Minutes prepared by S. Patchkovskii on November 10th, 2019, with corrections
through November 11th, 2019