Sunday, January 03, 2016

Paris Machine Learning Meetup: A Two Year and a Half Review

Happy New Year to y'all.

The Paris Machine Learning meetup now has 2973 members as of this writing. We have had more than 34 meetups so far. Here is a small review of what we have seen in the past year and half.  First and foremost, we are thankful to the sponsors and speakers who made this happen.

With regards to the rooms provided, the following companies and organizations did provide us with the luxury of hosting people all around Paris in rooms ranging from a 100 to 250 people. On behalf of Franck and myself, a big Thank You to them.
The following sponsors provided food and drinks for the networking event afterwards. Here again thank you to them:
All these sponsors have understood that a vibrant Machine Learning community is an important signal. Did we see any trends in these past two years and a half ? Yes. Here are some:

Deep Learning
Starting mid-2014 it became clear that deep learning was becoming important because it did not require people to know the black magic involved in training these deep architectures. These models with large amounts of coefficients has yielded a flurry of new actionable features that could not be guessed through traditional techniques. Because of how rich those new features are, deep learning techniques are beginning to supersede and extend current state of the art methods with very little inside knowledge in diverse engineering fields: As long as there is lots of data, lots of it. The sudden appearance of a rich set of coefficients in these deep representation has also yielded some interesting development in the realm of creativty and the arts.
Competitions (Kaggle/AutoML...), Challenges

Kaggle, the Uber of algorithms :-) , has been shown, along with other challenge platforms, to be a great equalizer among algorithms. The results of these challenges have unearthed algorithms that some people thought were passé or academically irrelevant (Random Forests). This is noteworthy since these successes have yielded a renewed interest in making sense of these algorithms. These challenges have also shown to the masses that deep learning wasn't just about specific academic datasets and was in fact pretty powerful as long as you had lots of data. The recent emergence of a greedy solver (XGBoost) is a signal we also picked up. These competitions are game changers not just in business but also in academic circles.

We asked a few nonprofit projects to tell us about their needs in terms of Machine Learning or provide a panorama of their fields. Some were obvious fit since they were academic, they asked questions that are currently not being addressed by companies such as: How do you learn in a different way ? or is there a master algorithm ? Others projects were less obviously connected to Machine Learning at first but that impression eventually dissipated during and after the networking events.


We also saw the rise of APIs for ML algorithms recently. As specialists, we wondered about how to go beyond those interfaces. However, the need for these sorts of approaches are currently very much needed.  

Companies/Startups and ML

We have had quite a few companies mentioning how they were using Machine Learning to improve their businesses. Some of these companies are transforming themselves while others looked like Machine Learning companies that fell into a specific fields. We also had a few local startups incuding some ambitious ones.

Social media

We started some social media accounts to provide a more direct connection between membrs of the group in between meetups. There is a group on LinkedIn, a Facebook page and a Google+ group. The idea is to provide a local scope of machine learning within Paris or France. The underlying reason for these group is to have a conversation on how Machine Learning is viewed and used by different stakeholders. We believe that many conversation need to happen. We also have a Twitter account and like to use the #MLParis hashtag. We also have a YouTube Playlist of most of the meetup videos . We have also streamlined our process with regards to producing the programs and have asked willing speakers to go and fill out this form. We have also set up a low frequency newsletter, go here if you want to receive it.

The Longer Views

Over the course of less than one year and half, Andrew, Yoshua and Russ provided us with an outlook of how Deep Learning evolved. In a different direction, Leon opened our eyes on the bigger picture of counterintiutive reasoning. Chris showed us that data science proves to be a way to jump from one field to another. Isabelle is organizing a competition set up to discover the mother of all agorithms and Lenka made some insightful connection between computational complexity of algorithms with large datasets and statistiscal physics. Philippe and Samim provided an outlook of the diverse fields in which ML is beginning to have an impact. Let us hope the rest of season 3 will provide additional insights.

The List

Technically, we had quite a few algorithm/development/tutorials talks ( SGD, Deep Learning, Kernel, NLP/Word Embeddings/Sentiment Analysis, Streaming algorithms, Bayes, Reinforcement Learning, Random Forest, Extreme Classification, Matrix Factorization, Computer Vision (SIFT;....), Model Composition, ELM/Shallow Neural Network, Nature Inspired), a whole slew of applications ( Telecoms, Fair Algorithms, Causality, Biology/Genomics, Robotics, Education, Botnet detection/Internet Security, Tax, HFT/RTB, Journalism, Animal Communications, Visualization , Finance, Creative AI, IoT, Hardware, Energy Use Data Centers), but also presentations and description of packages (Vowpal Wabbit, Scikit Learn, GraphLab/Dato, XGBoost, ML APIs, Rule Based ML, Add-ons) as well as a few announcements. Here is a more detailed listing:

General talks

General talks
Companies/Startups and general presentations in relation to ML

    Competitions (Kaggle/AutoML...), Challenges




    Deep Learning

    NLP/Word Embeddings/Sentiment Analysis
    Streaming algorithms

    Bayesian approaches
    Random Forest
    Matrix Factorization
    Computer Vision (SIFT;....)
    Model Composition
    Nature Inspired

    Botnet detection/Internet Security
     Creative AI
    Energy Use Data Centers:
     Rule Based ML

    ML and Data Science
    Here s what the archive page looks like:

    Related links

    All the slides and videos of the previous meetups can be found below:

    Season 3 (September 2015- June 2016)

    Season 2 (Sept 2014 - July 2015)

    Season 1 (June 2013 - July 2014)
    •  Epilogue Season 1 (July 2014 at DojoCrea)

    No comments: