At 9:30 AM CEST we have launched our “Virtual Summer School: Deep Learning for Language Analysis” #vssdl4la, continuing the series of summer schools which deals with the application of deep learning methods to language data (see here for the announcements of 2018 and 2019 – in German).
- Day 1: Monday Aug .31 – A warm welcome with a dog as supporting actor.
- Day 2: Tuesday Sept. 1 – To text, or to audio, that is the question.
- Day 3: Wednesday Sept. 2 – Guessing locations, social edition.
- Day 4: Thursday Sept. 3 – Going Newrow.
- Day 5: Friday Sept. 4 – Three prices, a panel and a passage.
As the name of the school of this year already indicates, the Covid-19 pandemic prevented us from holding the event at the University of Cologne. On the one hand, this is of course a great pity, because in the last two years it was a great pleasure for us to welcome participants with very different backgrounds – regarding their place of residence as well as their academic orientation – at Cologne to offer a program that allows networking also beyond the teaching sessions. This networking during the coffee breaks at the Forum of the Philosophikum and during evening meetings in a brewery or on the “Mäuerchen”, a meeting point close to the Univeristy at the Zülpicher Straße was a central element of the Summer Schools that we were very reluctant to do without. In updates of this post I would like to report what we came up with instead and whether these program elements were an adequate substitute for face-to-face-meetings.
Again, we organized the #vssdl4la with a team consisting of Felix Rau (now DCH), Claes Neuefeind (now CCeH) and myself (still IDH). Administrative support was again provided by Competence Area III (Quantitative Modelling of Complex Systems) in person of Christian Fischer, who was thrown in at the deep end shortly before the conference (because his predecessor Frederick von Heymann moved to the Humboldt Foundation in August) and this year by the Center of Data and Simulation Science (CDS, in the person of Alexander Heinlein). This means that all three of Colognes digital humanities institutions are involved, as well as two interfaculty centers, but this has not made our cooperation more complicated. On the contrary – it was a very fruitful collaboration.
I hope to be able to give an updated report here in the next few days on how the virtual school is running. So far (status: Monday noon) everything has worked, 68 participants from four continents have found the opening session as well as their parallel workshops. See you soon.
In the morning, we discussed the schedule one last time, and started the Summer School on time at 9:30 AM CEST with a meeting in Microsoft Teams. It took a little while until (almost) all participants found the meeting. After he had asked his dog out of the room – which started to whistle punctually at the start of the meeting 🙂 – Christian welcomed our participants.
Øyvind Eide as IDH Head of Department gave the official welcome address and thematized briefly the manifold connections between Deep Learning approaches and the Digital Humanities. Afterwards we (Alexander, Felix, Claes and me) had the opportunity to introduce ourselves and the institutions we belong to before Christian explained our program once again.
After that we were able to clarify a few questions and point out that there is the opportunity for further questions in the Teams chatrooms for the entire duration of the school. Not least this possibility had led us to the decision to use Teams instead of Zoom. Although we all had good experiences with Zoom within the last semester, we lacked the possibility to publish information beyond individual meetings. If you close a Zoom meeting, the whole chat history is deleted (unless you export it statically). Within Teams you have the possibility to create chat rooms in addition to the meetings. You can also create several meetings in parallel, so that the participants can switch between all of them. Although all of us were at least annoyed about problems concerning our Microsoft accounts, this chat possibilities were the deciding factor for using Teams. For us, the complexity of the software seems appropriate for summer schools and conferences.
After the introductory session, our lecturers Nils Reiter (DeepLearning4TextProcessing) and Joscha Rieber (DeepLearning4AudioProcessing) started their parallel workshops, which will continue until Thursday morning. So I will still have time tomorrow to go into them in more detail, because we are now starting a trial run for our social event on Wednesday.
As you can see in the schedule above, the Summer School consists of four different parts: Two parallel workshops, a course given by NVIDIA, a panel discussion and a social event. By far the most extensive are the two parallel workshops taught by the two lecturers Nils Reiter and Joscha Rieber that started yesterday and will end with their last session Thursday morning. In order to ensure a balanced distribution between these two workshops, we asked the applicants to decide already at the time of registration which of them they would like to attend – text processing with Nils or audio processing with Joscha. In the end we were able to accept 24 applicants for the audio workshop and another 32 for text.
Organizational and technical both workshops are designed in the same way. Since Deep Learning applications are very demanding in terms of processors (ideally GPUs) and RAM, we decided to rent machines from AWS (Amazon Web Services) for our setup and made them available to the participants. Each participant received a virtual machine with a preinstalled Jupyter Notebook, that is used as programming environment in our courses. Required libraries like tensorflow, keras, scikit-learn, numpy, pandas and librosa were also pre-installed. The company StormReply helped us with the setup of the machines and was a very reliable partner.
Both workshops were divided into (more theoretical) lessons in the plenum and (more practical) exercises in smaller groups. In addition to our two lecturers Joscha and Nils, our advanced students Johanna Binnewitt and Katja Bertholdt (for text) and Julia Kappes and Thomas Oehlke (for audio) were available as tutors for the exercises. You might know Julia and Johanna from my reports about the former Summer Schools (see above), they are already experienced tutors.
The main goals of the workshop “Text Analysis with Deep Learning” are to give an introduction to the basics of Deep Learning as well as to enlight on the specifics of language data. From this, best practices for the application of the one to the other will be derived. These will then be tested in practical exercises.
In the workshop “Audio Mining and Speech Recognition”, participants are introduced to use machine learning to solve problems related to audio data and in particular audio recordings of speech. The participants work with multilingual audio data and focus on language identification. They learn how to augment speech audio data and train a language-recognition model using deep learning.
Between this two offered workshops the decision is of course really hard. Which one would you have chosen?
An essential feature of our previous Summer Schools was that the participants not only met in the classrooms, but also during the coffee breaks, that they had lunch together and also met in a brewery in the evening. Due to the circumstances this year all this was unfortunately not possible.
Originally, we wanted to try out the tool Yotribe for our coffee breaks, but we did not manage to get access in time – the invitation mail unfortunately seems to have somehow got stuck in my spam filter. Well, maybe the participants were also quite happy to have screen breaks between the extensive learning sessions.
But we wanted to organize at least one social event for all registered students, tutors, lecturers, and organizers. Some of us already had experience with virtual conferences, where, for example, very well received pub quizzes were used as social events. We also thought that a kind of knowledge game might be a good idea, but we also want to involve the participants themselves.
Therefore we have asked them to send in photos of the place they live and the location data, where these photos were shot. Then we added the Locations to a KML file that we sent back to them and asked the participants to import it as a Google Earth project. During the event, we showed the numbered photos and the attenders had to guess, to which location the photo belongs to. We had a lot of help from Slavina Stoyanova and Lukas Sept (both IDH), who did a great job in preparing and moderating this GeoGuessingGame.
Although it was a completely new development, it worked out quite well. The interaction of could have been a bit better, we still have to think about it. But because we think the game might also be a good idea for other online events with participants from all over the world, we’re going to put the necessary materials to explain and play this game on GitHub (I will add the link later).
We are very curious to see how many locations the participants have correctly assigned – the evaluation is still ongoing. Of course there are also prizes to be won, which can certainly be used at the next online conference.
In the morning, the two parallel workshops on text and audio processing ended with a comparison of the performances achieved by the systems of different teams on specific tasks. The winning teams from both workshops will present their systems at the final session on Friday.
For the 24 hours following, the Summer School will take place outside of our Teams group, as the NVIDIA workshop is on the program, which will use Newrow as collaboration plattform. Since we already had a very good experience with NVIDIA last year, we are happy that we could win the company as a partner for our school again. The offered workshop Building Transformer-Based Natural Language Processing Applications is a brand new development and can actually only be booked from October (an exception was made for us). It focuses on transformer architectures that are used, for example, in the currently hotly debated GPT-3.
The best thing is that the participants have access to the workshop documents and the necessary NVIDIA hardware for a whole year, so that they can continue experimenting after the school, which unfortunately ends on Friday. 😢
For the last session of the Summer School, we had planned a panel to which we invited academics who deal with deep learning in their professional practice. In this way we wanted to bring an application-oriented perspective to the school so that our participants would be taught how to apply the knowledge they had acquired in their workshops.
Almost 60 participants had gathered in Teams (a very good numer for a Friday afternoon) and before the panel starts we announced the winners of the geoguessing quiz. No less than five of the participants had assigned all of the photos to the right locations. So we had to take the time of their transmissions into account before we could award our three prizes. Between first and second place was less than a minute! Afterwards, the workshops of Nils and Joscha were summarized by two of their students before the panel started.
The panelists were Birgit Kirsch (Fraunhofer IAIS, Natural Language Understanding), and David Biesner and Lars Hillebrand (both University of Bonn/Fraunhofer IAIS), who each presented the application of machine learning methods to the analysis of their data (from financial resp. legal domain). At the end there was an fruitful question and answer session before Øyvind Eide – who had already started the school – finished it.
I am sorry here that I cannot directly draw a conclusion after this long week. The feedback we have received so far has all been very positive, even though it was the first virtual school we have hosted. The use of Microsoft Teams proved to be a good choice. On the one hand, because you could manage a lot of different (sub)sessions, on the other hand, because you could store information visible for everyone without having to write mass mails all the time. This can certainly be transferred to an on-site conference, which we would like to host again next year. Until then, get well through the remaining pandemic!