Last week I attended PyData Berlin 2016. It was my first non-academic conference. I was not sure if it was going to be interesting or going beyond things I can see on the Internet anyway. But since I've never been to Berlin, had an open invitation by a friend of mine¹, and it's 6.5h by train² from Warsaw, I decided to go.

td;dr: It was worth my time and I really enjoyed the event.

Talks

Talks I have found the most interesting:

• Katharina Rasch, What every Data Scientist should know about data anonymization
  • anonymization by merging unique rows; sensitive information is culture-dependent
• Łukasz Czarnecki, Brand recognition in real-life photos using deep learning
  • if you use pre-trained convolutional neural network (like VGG_S), not that many samples are needed (~300 per class)
• Maciej Gryka, Removing Soft Shadows with Hard Data (research paper)
  • custom Random Forest can do wonders, even for problems typically suitable for CNNs (vide image colorization)
• Julia Evans, How to trick a neural network
  • a periodical reminder that learning by playing (and breaking) is the best way of learning
• Matthew Honnibal, Designing spaCy: A high-performance natural language processing (NLP) library written in Cython
  • after reading Sense2vec with spaCy and Gensim it was insightful to hear about the package design philosophy

Also, these were good:

• Daniel Kirsch, Functional Programming in Python
• David Higgins, Introduction to Julia for Python programmers
• Katharine Jarmul, Holy D@t*! How to Deal with Imperfect, Unclean Datasets
• Ian Ozsvald, Statistically Solving Sneezes and Sniffles (a work in progress)
• Olivier Grisel, Evolution of the pydata ecosystem
• Delia Rusu, Estimating stock price correlations using Wikipedia

And talks I missed, but I am sure were great:

• Wes McKinney, Python Data Ecosystem: Thoughts on Building for the Future
• Lev Konstantinovskiy, Practical Word2vec in Gensim (workshop)
• Maciej Jaskowski, Let's play Space Invaders!

If your beloved talk is not there, don't cry - most likely it was in a parallel session. (Also, in general topic selection and quality of presentation was good.)

I had a lightning talk: Teaching Machine Learning. I should write a blog post on it one day (especially on the 5-day data analysis summer school for sociology students and researchers, as now materials are in Polish). As for now, it is implicitly covered in Data science intro for math/phys background.

Other take-home lessons

• There are so many ex-physicists. I even heard Hey, I saw you... on ICPS³.
• Many methods are new, so it's crucial to learn them (and how can they be tailored to your tasks); sometimes being an expert with 10-year experience is physically impossible.
• The number of participants was optimal (200? I hate huge conferences).
• If other PyData events are of similar quality, it's not my last time there! :)

Links

• PyData Berlin 2016 Materials (mainly slides)
• Notes from my PyData Berlin keynote by Julia Evans
• If your have your own blog post or photo gallery from this event - mail me! :)
• UPDATE: There are videos from the talks!

Thanks to the organizers!