Professional Documents
Culture Documents
Machine Learning
Summer School
(TMLSS)
Summary of the first edition
The first edition of TMLSS was held in Cluj-Napoca, Romania, and brought together talented
participants and amazing speakers from EE and the rest of the world, for six intense days of
lectures, practical sessions, discussions, and networking activities. Given the success of the
first edition, measured through the positive feedback provided by all the people involved
(participants, speakers, sponsors), we are committed to continuing the project and we hope
we will enjoy the same support from the local and international communities in the future.
Although this school is a small scale event, aimed at raising awareness for EE, the school is
not restricted to the region. We hope to enhance communication between communities,
within EE but globally as well, and to ease access to knowledge for all interested in Artificial
Intelligence. So we hope everyone felt welcome and we welcome everyone to apply and
participate to future editions of the summer school.
We are extremely grateful for all the support and help we received in making this school
happen. We thank Luigi Malago and Razvan Florian from Romanian Institute of Science and
Technology, Cluj-Napoca, for their active involvement in the organisation of the school.
We thank the speakers (Dumitru Erhan, Guido Montufar, Jan Chorowski, Kyunghyun Cho,
Lucian Busoniu, Luigi Malago, Maria-Florina Balcan, Marius Leordeanu, Nicolas Heess, Oriol
Vinyals, Raia Hadsell, Ulrich Paquet, David Szepesvari, Diana Borsa, Mihaela Rosca, Wojtek
Czarnecki) for donating their precious time and for sharing with generosity their knowledge
and research ideas.
We are extremely grateful to our sponsors and individual donors whose generosity was
essential in making this event a high quality learning experience but affordable and
accessible to everybody: DeepMind and Siemens at Diamond level; Amazon, Artificial
Intelligence Journal, Everseen, Intel, and The Curious AI Company at Platinum level;
Bitdefender, Bosch, certSIGN, Continental, facebook, General Intelligence, Rakuten, Google,
and Tora at Gold level; Accesa and Catalysts at Silver level.
Last, but not least, we would like to thank to several individuals whose contributions were
paramount in making the event a success: Shakir Mohamed and Ulrich Paquet for sharing
advice from their Indaba experience; Elena Pascanu for advice and support throughout the
organisation of the school, prof. Ion Bica for endorsing and promoting the event, Gabi
Marchidan for all the help in organisation and for an unlimited source of humour that got us
through the tough moments, Cosmin Huruiala and Vlad Piersec for useful advice about Cluj.
Razvan Pascanu
Viorica Patraucean
Doina Precup
Source: Source:
http://www.aiindex.org/2017-report.pdf https://twitter.com/nipsconference/stat
us/909064018559590402?lang=en
The industry is becoming also deeply involved in the field. Giants like Google, Facebook,
Amazon, Baidu, Intel and many more are regularly present at these conferences. For
example, for ICML 2017, about 20%-25% of accepted papers had industry involvement
(source). See below-right a breakdown of top institutions according to number of accepted
papers at the conference. Large corporations, beside carrying and publishing a lot of
research, are also investing in the field, offering scholarships or residency programs, e.g.
here, here, here or here.
years in AI
While there might be some uncertainty on how far along we really are with regard to AI
technology, one thing is clear. Machine learning, in one form or another, from simply
smarter devices to self-driving cars, from robotics to search engines, will impact and is
impacting everyone's life: how we interact with each other socially, how we entertain
ourselves, the job market, how we treat illnesses or how we do science.
Statistics on the research output in the field* (see figure below) show a disproportionate
distribution of where this research is carried out, the field being largely dominated by
research centres from US, Canada, some Western Europe countries (e.g. UK, Germany,
France) and China and Japan to some extent.
Source: http://www.jfgagne.ai/talent/
See here or here or here for in depth discussions on the importance of diversity. One
argument is that the tools we build reflect the culture and needs of those who build them.
Without a diverse set of researchers, we cannot build new technology that benefits
everyone. Our background influences what problems we tackle and the way we approach
them, but it also determines our potential inability to notice certain types of biases, for
example in the data we use to train our models. While following good practices and
guidelines might help in preventing intentional discrimination, a lot of negative
consequences can come from unintentional or unconscious biases -- and diversity in the
field might be crucial in fighting effectively against these.
How to fix the diversity problem in AI research (or elsewhere) is an open social problem with
no straightforward answer or quick fixes.
There are a number of ongoing projects that are trying to improve diversity and access to
knowledge in AI for underrepresented groups, e.g.: Deep learning indaba, LatinX in AI, Black
in AI, Deep AfricaAI, while some other important initiatives try to improve access to
education in general, e.g. Coursera, TwentyTu.
Current ML related events in the region include RAAI, TFML, PL in ML, DeepBayes, LeMAS.
Our goal is to join the efforts of strengthening the local research community by creating
high-quality international events that welcome participants from all over the world in EE.
* Statistics are collected based on publications in main ML conferences such as ICLR, NIPS and ICML in the last few years and
provide an english-centric view .See the source of the plot for how the data was collected and what it represents.
ML / AI research suffers from a significant gender imbalance; see male to female ratio for
authors of NIPS, ICML and ICLR submission in 2017 in figure below. The problem is larger
than ML, the lack of gender diversity being rampant in STEM sciences all over the world.
There are a multitude of causes, from momentum of the status quo to explicit
discrimination, some of it even disguised as cultural norm. And this imbalance has real
consequences, from the type of problems we address and type of solutions we find
acceptable, to biases in our datasets. For a detailed discussion of the topic we suggest
looking here or here or here; particularly see the talk given by Daphne Koller at ICLR 2017 or
this article.
Several efforts try to address this issue and encourage or provide support for women in
STEM, e.g. Women In ML, Women in CV, Black girls code, Stanford AI4ALL, etc. However,
correcting for the skewed distribution will be a long process that will need continuous effort
from the entire community. We want to openly acknowledge this issue and join the efforts
of encouraging and supporting women researchers in ML and AI.
Source:
https://medium.com/element-ai-research-lab/estimating-the-gender-ratio-of-ai-researchers-around-the-world-81d2b8d
be9c3
To get an idea on the affordability of ML education programs, let us focus on the cost of
attending conferences and summer schools in the field. The registration fees for attending
top ML related conferences, as a student, in 2018, not including accomodation, travel or
food, are on the order of: 300€ for ECCV or AISTATS, ~360-380€ for ICML, NIPS or ICLR,
~400€ for AAAI, ~490€ for CVPR. Table 1 shows estimated prices for several ML related
summer schools in 2017-2018. Note that, among these, the last two have considerably lower
participation fees: Indaba has free registration for students as it was designed with access to
education in mind, whereas DeepBayes is organised in Russia, where the costs of living are
lower compared to other countries.
While the high registration fees of most conferences and summer schools might be well
justified to ensure the self-sustainability of the events, they can become quickly prohibitive
for many, especially when taking into account the costs of travel, accommodation, and food.
Indeed, most of these events are organised in Western Europe, US, or Canada - countries
with high costs of living - so the subsistence costs might be even higher than the
registration fees. Well-funded research groups and doctoral schools might cover the costs
of attending such events for students, but it is not a standardised practice for all institutions,
and it surely is a very limited practice in EE. Hence, the students and researchers would
have to pay from their own pocket the fees. But when we look at the minimum wage per
country in Europe in Table 2, we can quickly see how such events are simply not affordable
for many. Some summer schools and conferences do offer scholarships to address such
issues. However, these are usually very limited in number and they might cover only a small
part of the overall cost. This makes it very difficult for researchers from certain regions (e.g.
EE) or less visible institutions to attend such events, which in turn makes it very difficult for
them to be part of the community, find collaborators, etc. This drives the lack of diversity by
affecting disproportionally different segments of the world. Here or here are some
discussions on these issues.
Source:
https://en.wikipedia.org/wiki/International_Mathematical_Olympiad
Leaving aside pure research, in terms of expertise in the field, the picture does not look
drastically different. The figure below shows statistics of LinkedIn profiles per country that
have ML or AI in their expertise. For details please see (http://www.jfgagne.ai/talent/). The
picture below shows correlation with the amount of published work.
Source:
http://www.jfgagne.ai/talent/
Possible causes
The factors underlying the current situation in Eastern Europe have complex historical,
political, and economical roots and their in depth analysis is beyond the scope of this report.
See this article for a discussion.
One possible cause is the large brain drain phenomenon that the region is experiencing,
which plays a significant negative role. The 2015 UN report on immigration placed Romania
and Poland on 2nd and 3rd positions after Syria in the increase rate of diaspora population [1].
Adding that in EE countries it is mostly the highly skilled people that immigrate (see source
here, here, and figure below), this could explain the poor quality of education at university
level.
[1] International Migration Report 2015, Highlights, Department of Economic and Social Affairs, United Nations
New York, 2016, ST/ESA/SER.A/375.
Source: https://www.imf.org/external/pubs/ft/sdn/2016/sdn1607.pdf
Source: https://www.imf.org/external/pubs/ft/sdn/2016/sdn1607.pdf
Here are some extracts from opinions expressed by ML researchers and faculties affiliated
in Romania and Poland in a survey that we carried related to these aspects:
Q: According to Times Higher Education 2018, Romania's best ranked university is in the range
601-700, Poland's best university is in 501-600. What do you think is the main issue explaining such
low rankings?
A: “One generic explanation would be that a CS department such as one in Warsaw can afford to finance
just 2-3 bigger projects (preferably cheap projects, such as projects in theoretical computer science). At
the moment we basically have big groups in algorithms and logic in computer science and a smaller
crypto group.” Henryk Michalewski, Warsaw University, Poland
A: “Lack of collaboration between universities, lack of collaboration between industry and universities,
outdated curricula, lack of internal university governance that promotes meritocracy.” Gabi Marchidan,
Iasi, Romania
A: “Lack of public financing for the education which leads to low paid teachers and no interest in
pursuing an academic carrier. Also, in Romania, there is an acute lack of communication between
scholars due to a misperceived need for individualism in research.” Adriana Stan, Technical University of
Cluj-Napoca, Romania
A: “1. insufficient funding, 2. insufficient support for obtaining grants, 3. researchers are not encouraged
(comp.sci) to publish and the teaching load is too large (60% - 65% personnel only), 4. no successful
cooperation with local companies.” - Anonymous, Romania
Q: Both Romania and Poland have low outcome in terms of research publications in top conferences
compared to Western countries. In your opinion, what is the main cause of this in your country?
A: “[...] In Machine Learning there is no serious research tradition - we are trying to establish it, but it
will be a long process.” Henryk Michalewski, Warsaw University, Poland
A: “[...] the universities are not that open to international students and prefer not to deal with the hassle
of paperwork. This is also true for undergraduate programmes. The openness to international students, I
think, would determine an increase in the level of involvement in research of both the students and the
professors, TAs, so on. At a larger scale, I think that access to the European research funding is quite
unbalanced. Meaning that the large grants are won mostly by the universities which already have lots of
expertise in the respective field. Hence, the access to such grants is restricted for new research labs, for
example.” Adriana Stan, Technical University of Cluj-Napoca, Romania
A: “Not having a meritocratic system and low investments in the educational system determines the
migration of skilled researchers to other areas where they can be well rewarded for their abilities. This
also means that, in time, the universities have had to suffer from a lack of well trained and capable
individuals willing to join them. We are now seeing the long term effect of the 2 above causes.”
Anonymous, Romania
A: “No long-term reward for publication, no consistency in the funding / rewarding / planning
mechanisms.” Anonymous, Romania
A: Research performance criteria in Romania are not aligned with Western countries, e.g. a NIPS paper
values less than a paper in some obscure ISI-indexed conference. There is high-quality teaching in
Romanian universities for Computer Science fundamentals, but the mentors to encourage students to
follow the research path are missing. The costs for attending top conferences in ML are very high
compared with income/budget in Romania. Elena Burceanu, ML researcher at Bitdefender, Romania
Q: If you could change something in the educational system in your country, what would that be?
A: “[...] students are starting to work in various companies at a very young age and we do not have any
systematic way to retain them. One method which seems to work is to send the best students to do
research internships (in contrast to engineering internships).” Henryk Michalewski, Warsaw University,
Poland
Q: If you could change something in your current position, what would that be (number of teaching
hours, salary, less/more/better students to supervise etc.)?
A: “A peculiarity of Polish system is that grants pay money to everyone engaged into a given project. With
this extra "grant" money a professor income is reasonable.” Henryk Michalewski, Warsaw University,
Poland
A: “[...] the students [...] are great and intelligent, however most of the professors prefer not to have them
think too much for themselves, so they end up quite disappointed by their university experience [...] and
start working full-time in IT companies as early as the second year. As a result, it is quite hard to find
creative, involved students to do a good Masters or PhD thesis.” Adriana Stan, Technical University of
Cluj-Napoca, Romania
A: “reduce the number of teaching hours and the administrative overload” Camelia Lemnaru, Technical
University of Cluj-Napoca, Romania
Due to the growing interest in the fields of ML/AI, there are a number of initiatives that aim,
directly or indirectly, to improve the state of research in Eastern Europe in fields related to
ML. Taking Romania as an example, these initiatives include:
● Research groups within universities focusing on ML related topics, created by
researchers publishing in top conferences. E.g.: Politehnica Bucuresti - Robotics and
Computer Vision group, Technical University Cluj - Reinforcement Learning, Image
Processing, Computer Vision.
● Large international companies opened offices in Romania (e.g. Intel-Movidius,
Amazon, Siemens, etc.), having R&D as an important sector of activity. These co-exist
with important Romanian companies such as Bitdefender, and start-ups. Their
contribution within the local community is very important due to their efforts to
support the local universities through collaborations, sponsorships, internships.
● SSIMA - summer school focused on medical image processing, including ML applied
to medical data
● RAAI - conference on AI related topics (NLP, Vision, Information Retrieval, ML)
● AI Meet-ups and other self organised meetings: these are generally informal free
events where people with interest in AI gather to discuss and network
This list is not exhaustive. Similar initiatives are happening in other EE countries also. Whilst
these initiatives are highly beneficial for the local communities, there is still a lack of
communication between different research groups in EE and between EE and the rest of
the world.
Here are some possible recommendations that we believe would make a high positive
impact for the EE community (as well as other underrepresented regions):
- Currently, top conferences have different registration fees for students vs
non-students, which are very beneficial. They could also have different registration
fees based on the country of affiliation. Such a measure could be negligible to the
overall budget, given the strong involvement of international companies in top
conferences through generous sponsorships.
- Top conferences that tend to be sold out could change the first-come-first-served
registration policy, to control the distribution of attendants in terms of affiliation.
- Small and medium conferences should consider more EE countries as potential
locations; see at the end of this report a more detailed argument on this.
- Experts in the field from well-developed countries could seek more to establish
collaborations with EE groups.
- Renowned hubs could attempt to provide internships/fellowships aimed at students
and faculty from regions that are underrepresented.
Organisation rules:
● The school is a non-profit event.
● The school must be organised in an Eastern European country, preferably on an
yearly basis.
● The school gathers participants, speakers, and sponsors from all over the world, all
having in common the passion and interest for ML / AI.
● The selection of participants will favor students and practitioners from Eastern
Europe. However, this does not exclude participants from the rest of the world. On
the contrary, they are welcome and needed in order to build collaborations between
the EE and rest of the world. We aim to have a significant percentage of participants
not being from EE.
● The selection of participants strives to be gender-balanced.
● A significant percentage of speakers must be affiliated in Eastern Europe, to highlight
the local research and help setting up collaborations.
● Gender balance among speakers is highly desired.
● The school fees must be as low as possible, for the school to be accessible.
The first edition of TMLSS was organised in Cluj-Napoca, at Babes-Bolyai University (UBB),
Faculty of Economics and Business Administration (FSEGA).
Accommodation for participants was provided at the UBB Economica II dorms, in walking
distance to FSEGA for roughly 70€ for the entire school (6 nights), two persons per room.
School fees:
● 100€ for students
● 150€ post-doc or faculty
● 300€ other (e.g. industry)
The school provided scholarships
(registration fee waiver, free
accomodation in the dorms and
travel grants) need-based.
The school ensured catering for the
whole week (breakfast, lunch, dinner,
coffee breaks for 6 days), excluding
2 dinners.
The organisers of TMLSS2018 were: Doina Precup (McGill University and DeepMind), Luigi
Malago (RIST), Razvan Florian (RIST), Razvan Pascanu (DeepMind), Viorica Patraucean
(DeepMind).
The local partner institutions in organising the school were the Romanian Institute of
Science and Technology (RIST) and Babes-Bolyai University (UBB).
RIST provided logistic support for several aspects including visa letters, sponsorship
contracts, management of the budget, online payment system for collecting registration
fees, etc. UBB provided lecture and lab rooms for free and dorm accommodation at its
standard price for external students. Besides helping with logistics aspects, 12 students and
researchers from RIST and UBB respectively, joined the organisation efforts as volunteers
and teaching assistants.
Various international and local companies sponsored the event at different levels; see
below the list per sponsorship tier. Additionally, a number of individuals made donations to
the school: Anna Bortsova, Anthony Phalen, Cosmin Paduraru, David Khosid, David
Warde-Farley, Razvan Pascanu, Mihaela Rosca, Tudor Leu + anonymous donors.
Importantly, these donations were made through the Benevity platform and Google
matched the donations.
The first edition of the school focused on Deep Learning and Reinforcement Learning, and
gathered 19 experts in the field listed below. In terms of ratios, 4 out of 19 (~21%) speakers
were affiliated to institutions from EE. 6 out of 19 (~31%) speakers were female. Overall, the
speakers were affiliated to 11 different institutions.
15 speakers held lectures attended by all participants (i.e. single track) and 5 speakers held
practical sessions in parallel rooms (~ 30 participants per room), same topic being taught in
all rooms.
The slides of almost all the lectures were made available to participants. The code for lab
session was released on github in the hope that it can provide a starting point for teaching
materials in other educations programs.
The schedule spread across 6 days, starting on Monday morning and ending on Saturday
evening. Each day consisted of 3 lectures of 100 minutes each, an industry keynote, and one
lab session. The school had 3 social events: the welcome reception on Monday evening, a
half-day trip on Thursday afternoon at Turda salt mine, and the Gala Dinner on Friday
evening. Additionally, on Tuesday and Wednesday evenings there were poster and industry
stands sessions. A panel discussion was held on Thursday, having Dumitru Erhan as
moderator and Doina Precup, Razvan Pascanu, Nicolas Heess, Luigi Malago, Maria-Florina
Balcan, and Ulrich Paquet as panelists. Participants were able to send their questions in
advance or ask questions live.
The lectures covered a wide range of topics: Deep Learning for Computer Vision, RNNs for
NLP, Learning Theory, Maths of Neural Networks, Unsupervised Learning and Generative
Models, Basics of RL and Deep RL, Continuous Control, Robotics, and GraphNets.
The school provided catering for the whole week (breakfast, lunch, dinner, coffee breaks),
except dinners on Thursday and Saturday. On Tuesday, Wednesday and Friday, the lunch
was at a nearby restaurant.
School timeline
The school was announced in early December 2017, during NIPS conference, when the
school website tmlss.ro went public. Announcement emails were sent on various academic
mailing lists (in the fields of ML, AI, Computer Vision, Neuroscience etc.). The application
period started on January 24th, with application deadline on March 30th. Selection
notifications were released in two rounds, on April 13th and April 27th. The timeline was
decided such that candidates have enough time to prepare their submissions (about 9
weeks), and that there is enough time after the selection notifications to make travel
arrangements (about 10 weeks).
Application procedure
Interested individuals were required to fill in an application form to provide contact details,
affiliation details, diversity-related details and a statement of research interests. Additionally,
they were required to upload a CV and, optionally, an abstract of an ML-related project.
Submitting an abstract increased significantly the chances of being accepted. The abstract
submission was not made compulsory to not prevent beginners from applying to the
school. Reference letters were not required, as they can be cumbersome to obtain
sometimes. We preferred to use the Research statement section instead to extract basically
the same information that a reference letter would provide.
Selection criteria
The selection procedure took into account various aspects, all meant to provide answers to
two questions: “How can the school help the career of this applicant?” and “How can this
applicant help the school?”. All in all, we strived to have a balance between experienced
participants and less experienced but with potential, as we believe that the teaching is not
only from speakers to participants, but also between participants. Once all the applications
were evaluated from this perspective, they were ranked taking into account also diversity
aspects.
Every application received at least 3 and at most 5 reviews. In the first round of notifications,
applicants who received at least 3 Accept scores were notified to be Accepted, applicants
who received no Accept score were notified as Not Accepted, and applicants who received
at least one Accept score were put on a waiting list and their applications were discussed
individually by the selection committee. Exception from this schedule was made for
applications from countries like India, Iran, were visa procedure was lengthier. In these
cases, the Accept/Not Accept decision was made in the first round, after discussions within
the selection committee.
We received 827 valid applications. The preliminary budget and planning was done to
accommodate around 80 participants. Due to the large number of applications, the number
was increased as far as the budget and overall concept of the school allowed. Eventually,
the school had 146 full participants: 98 accepted through selection, 31 from sponsoring
companies, and 18 from local partners; the latter included teaching assistants and
volunteers. In addition, there were 20 speakers and organizers, and 3 members of staff, for
an overall total of 169 persons. Besides the full participants, about 20 faculties and PhD
students affiliated to the local universities in Cluj-Napoca attended the lectures as well. The
next figures give details about statistics in terms of gender, affiliation, and career level,
before and after selection for the participants accepted through selection only.
Affiliation Statistics
As can be observed from these figures and the detailed maps on the next page, the ratios
before and after selection are similar, and a significant percentage of participants (35.4%) are
not from EE, showing that the school welcomes participants from all over the world.
Gender Statistics
Before selection
(827 applications)
After selection
(98 accepted applications)
These figures show the number of applicants and of accepted participants based on gender
and country of affiliation. In both EE and the rest of the world, it can be observed that the
number of female applicants is much smaller compared to male, with a slightly better ratio
for EE.
The breakdown of applicants and participants based on career level shows that the ratios
before and after selection were roughly preserved, except for the “Other (industry)”
category. This shows our commitment to support students and academic researchers, who
are most affected by the lack of funding in research.
Additionally, the ratio of faculty and post-doc increased after selection.
Due to the large number of applications, we were not able to provide individual feedback
for any application, be it accepted or not accepted. We did encourage and provided
individual advice to interested people who contacted us before the application deadline
asking about the suitability of a particular project.
In general, the selection committee wanted to see that the candidate was passionate about
Machine Learning, that the candidate had put effort into preparing the application, hence
was truly interested in attending the school, and that the candidate had the necessary basic
knowledge to understand or at least grasp the concepts taught at the school.
Last, but not least, we are aware that the selection process can be improved and we strive
to do so for the next edition taking into account the valuable feedback that we received
from the applicants this year. Namely, we aim to improve the guidelines and the application
evaluation process. We hope that the applicants who were not accepted this year will apply
to next year’s edition. Our policy is to favour applicants who haven’t attended the school in a
previous edition.
Income
The main part (~75%) represented sponsorships from international and local companies. The
other sources of income were donations from individuals (8.6%) and registration fees from
participants (16.4%).
Expenses
According to the no-profit rule, all the income of this year was spent in its entirety on
scholarships, catering, and other expenses (e.g. welcome bags, t-shirts, toiletries for dorms,
reusable water bottles, banners, pocket guides, etc.), as shown below. Importantly, this year,
most of the speakers’ travel expenses were covered by themselves, on their research
grants or as company expenses. Also speakers donated their time, and were not paid by the
school to lecture.
32 scholarships were awarded. These included registration fee, accommodation and travel
grant that covered fully or partially the travel costs. Due to the large number of scholarship
requests, we limited the travel grant awarded to any one participant, to be able to support
as many participants as possible. The travel grants limits were 70 EUR for participants
travelling from Romania, 250 EUR for participants travelling from Europe, and 750 EUR for
participants travelling from outside Europe. These limits were set as to cover fully or at least
75% of the overall travel cost. All the participants that were accepted and required a
scholarship received one, to cover fully or partially the costs of attending the school.
Note that the amount received from registration was smaller than the amount put in
scholarships. Hence the organisation costs of the school were supported exclusively
through sponsorships and donations.
Shortly after the end of the school, we sent out feedback forms to all participants (including
sponsors) to collect their opinions regarding: the overall organisation of the school, most
appealing and useful aspects of the school, overall schedule, quality and difficulty of
lectures and practical sessions, quality of industry keynotes, quality of poster sessions,
opportunities to network and social events, catering and accommodation.
This feedback is essential for shaping future editions, to keep the positive aspects and
improve the less positive ones flagged by participants. 64% of participants (including
sponsors) filled in the survey.
Motivation to apply for the school (multiple answers allowed): 89.8% choice of topics,
86.4% lecturers, 44.3% location, 27.3% low fees.
Most useful parts of the school (multiple answers allowed): 83% listening to the lectures,
72.7% networking with participants (getting to know new people), 64.8% lab sessions, 47.7%
getting to know personally the speakers, 33% presenting my work during poster sessions.
Quality of lectures: 51.1% excellent, 40.9% very good, 6.8% good, 1.1% should be improved.
Quality of lab materials: 47.7% excellent, 37.5% very good, 12.5% good, 2.3% should be
improved.
Quality of posters: 40% excellent, 51.8% very good, 7.1% good, 1.2% should be improved
Were there enough networking opportunities: 61.4% plenty, 28.4% enough, 10.2% not
enough
Rating of the social events: 84.1% excellent, 14.8% very good, 1.1% should be improved
Rating of the catering: 46% excellent, 41.4% very good, 9.2% good, 3.4% should be improved
Rating of dorms: 50% excellent, 35.7% very good, 11.9% good, 2.4% very poor
How did you find out about the school? 18.8% mailing lists, 12.9% facebook, 9.4% twitter,
8.2% meetup, 50.7% other sources (friends, colleagues, google search etc)
Lecture
“This was one of the best academic events I have
attended so far,[...] HUGE THANKS [...] TMLSS 2018
was one amazing event with a great value for
almost negligible price […] I think that TMLSS
2018 was a huge contribution to the Eastern
European ML community. Especially to people
that can't afford to apply to similar events in
North America (due to travel cost) such as
myself.” -- Lukas Martak
Poster session
“Thank you so so much for the amazing
experience. As a postdoc at Max Planck Institute
(Tuebingen), I still learned quite a bit this time;
really grateful for the organizing effort. I'd happy
to help if you guys need any in the future :) It was
a blast in Romania!!!” -- JJ Zhu
Poster session
“I was immensely impressed by how humble and
approachable all the lecturers were. In a full
week of interacting with all these accomplished
scientists I never heard one of them bragging
about their accomplishments or the field in
general. It shows that people at the forefront of
DL & RL research do not necessarily embrace the
hype around it.” (blogpost) -- Maria Rigaki
*Quotes are taken from feedback with no alteration. Emphasis is ours, and we used […] to shorten the quote.
Poster session
“Thank you very much again for organizing such
a wonderful event! I really enjoyed being in such
a live, friendly environment for a week and I would
definitely miss everything there. The organizers
cared a lot about the event and we had their
support all the time. From scientific point of view,
it was a really great summer school and we
learned a lot.” -- Masha Asadi
Coffee break
Coffee break
“There were no hype or buzzwords, but the
lecturers were people who know today’s
limitations, needs and questions and get things
done step by step, with an engineering mindset. A
fundamental feeling with which I left is that it is not
only about learning, but sharing the experience
and knowledge gained to empower others to
create and to strengthen the community and
motivation.” -- (blogpost) Dana Axinte
The feedback we received was overwhelmingly positive, as can be seen in the summary
and the quotes above. We are grateful for the kind words and the encouragements. They
are an important motivator for us to continue with the project. At least two blog posts have
appeared about the summer school (here and here), both presenting the school in a very
positive light. Some of the lab materials have already been used in two other workshops:
here and here.
We will focus next on the less positive feedback we received, as this is important for
improving future editions. The main negative comments can be split in a few categories,
and we will try to reply to each category.
There were several complaints regarding the sound in the first days and the air conditioning
of the lecture and lab rooms throughout week. While we did manage to improve the sound
after a couple of days, we couldn’t do anything about the air conditioning system of the
venue. We will definitely pay more attention in the future to the quality of the classroom.
Regarding catering, most parts of it were excellent as the participants acknowledged, but
the breakfast and coffee were indeed sub-par. We will do our best to improve this, as much
as the budget allows.
The schedule was indeed very dense and diverse in topics, as our aim was to maximise the
quantity of high-quality knowledge the participants are exposed to during the school. Given
the cost and investment from both participants and organizers, it seemed like a wasted
opportunity to not fully utilise the available time. We will, however, try to improve the
schedule for next edition. Among the suggestions we are considering are longer breaks,
possibly three poster sessions instead of two, more networking opportunities (possibly in
the form of small group discussions).
Most of the provided lab materials used tensorflow and Sonnet, a public library developed
by DeepMind. Most of the participants were not familiar with Sonnet, hence some frustration
appeared due to this additional challenge that they had to face in solving the lab exercises.
The justification for choosing Sonnet comes from the fact that preparing the lab materials
required considerably more effort and synchronisation from the lab instructors compared to
preparing lectures. Since the instructors were all familiar with Sonnet and the materials
were all prepared in their free time, Sonnet was a natural choice to reduce the amount of
work. We are grateful for the significant amount of time that they donated before and
during the school. The main issues regarding Sonnet were the lack of documentation and
sufficient getting-started tutorials.
One other important criticism concerned the selection process, its lack of transparency and
a perceived forced diversity. We refer the readers to section Selection process, for details
about the selection criteria. The selection process was far from trivial, as our aim was and
will be to encourage potential while also acknowledging expertise. We will do our best to
improve the selection procedure as much as possible, possibly by allowing a longer period
for reviewing applications.
Additionally, we strive to maximise the impact we can make in the local education system.
To this end, we are considering the following directions:
- Set up an advisory board formed of local and international experts, having a
consultative role. Its members will provide high-level guidance on the school
organisation using their experience and knowledge of the local and international
ecosystem, with the goal of maximising the impact of the school on the local
community. The members will be elected on an yearly basis.
- Communication in the EE community - we plan to set up a moderated mailing list of
active labs in Eastern Europe working on ML and AI. We hope this will facilitate
communication among the labs by making them aware of each other and help build
collaborations. Additionally, this mailing list could be used to post announcements
about ML related events, jobs posts etc. While similar systems exists within different
EE countries, our hope is to improve awareness and communication between EE
countries.
- Support local faculties to enhance their teaching materials. How: set up a network
(e.g. mailing list) where they can share teaching materials; invite international experts
to join the network and ask for support in improving these materials; the materials
from the summer school will be shared for these purposes; discuss the possibility to
set up series of invited lectures given by well known experts in the world that could
come and teach in local universities; discuss possibilities for local faculties to be
invited to visit established research groups in the world.
For a total number of 200 persons involved, we count roughly 20 speakers (similar to
TMLSS2018), 15 volunteers and teaching assistants, 25 participants from sponsors, and 140
participants from selection (compared to 98 participants for TMLSS2018)..
We consider roughly that all the expenses with the school organisation scale linearly with
the number of participants. This leads to a required budget of 78.800 EUR.
However, for this year’s edition, most of the speakers were self-funded. For the next edition,
we need to make budget provisions for speakers’ travel costs, a minimum of 10.000 EUR.
Speakers will not be paid to speak, these funds cover only flight and accommodation costs.
We aim to not increase the registration fees. Keeping similar percentages of students
(72.9%) vs. faculty & post-doc (19.8%) vs. industry (7.3%), among the participants accepted
from selection, the expected income from registrations would be 17.400 EUR. After taking
into account the number of scholarships awarded whose recipients do not pay registration
fees (this year ~33%), this leaves 14.100 EUR from registration.
The rest of 75.900 EUR until 90.000 EUR would need to be covered from sponsorships
and donations.
This is a clear win-win situation, and we hope conference or workshop organisers will start
considering the countries in the EE region as candidates for locating their events. Equally,
potential international sponsoring companies should see the high value for money of such
events: with much smaller sponsorships compared to events in Western Europe or US, they
can make a much larger impact on the participants’ experience, which in turn generates
positive publicity.
On a longer term, besides a positive impact on education, such initiatives could impact also
the local economy, by encouraging tourism and promoting destinations that are not
otherwise popular in a time when well known tourist locations (e.g. Venice, Barcelona etc)
struggle more and more with the large number of tourists..
Source: https://www.expatistan.com
*To calculate each country's price index, first a value of 100 is assigned to a central reference
country, here Czech Republic. Then the Price Index of every other country is calculated by
comparing their cost of living to the cost of living in the Czech Republic; e.g. if a country has a Price
Index of 134, it means that living there is 34% more expensive than living in the Czech Republic.