Paris Machine Learning Meetup Newsletter September 2016 [in French]
- Les prochains meetup #1 (14 septembre) et 2 meetups Hors série: #1 Stan — 20 Septembre — et #2 H20 Workshop le 21 septembre
- franceisAI 16–18 Septembre
- Corrélation n’est pas causalité
- Autres meetups et autres news
- Commission Parlementaire — OPECST, votre opinion nous interesse.
- Horse Workshop in London 19 Septembre, Londres/London
- 8 weeks course : Machine Learning with texts in Python
- Data Science Game, 10 & 11 september in Paris
- Kaggle data set
- Facebook Artificial Intelligence (Deep Mask)
- Baidu Research (Deep Learning Framework Paddle)
- Présentation d’un talk — Sponsoring d’un meetup
1 Les 3 Meetups Paris Machine Learning de Septembre, @ParisMLgroup
Parce que vous avez été gentils cet été et que le meetup a maintenant plus de 3900 membres, ce n’est pas un mais trois meetups que le Paris Machine Learning meetup vous réserve en Septembre (en plus de FranceisAi et des autres superbes meetups Paris NLP,…). Comme on dit en Anglais “No good deed goes unpunished” :-)
- Le premier meetup de la saison aura lieu le mercredi 14 septembre au Dojo:
Si vous voulez sponsoriser le catering du networking event (pizzas & drinks) du premier meetup de l’année, contactez Igor ou Franck.
Program
Nous devrions avoir au moins deux présentations, si vous voulez en faire une, n'hésitez pas a remplir ce formulaire: ici.
David Klein,
Title: 'Deep Learning for Global Biodiversity Monitoring'
Abstract: Healthy ecosystems with intact biodiversity provide human societies with valuable services such as clean air and water, storm protection, tourism, medicine, food, and cultural resources. Protecting this natural capital is one of the great challenges of our era. Species extinction and ecological degradation steadily continues despite conservation funding of roughly U.S. $20 billion per year worldwide. Measurements of conservation outcomes are often uninformative, hindering iterative improvements and innovation in the field. There is cause for optimism, however, as recent technological advances in sensor networks, data management, and machine intelligence can provide affordable and effective measures of conservation outcomes. I will present several working case studies using our system, which employs deep learning to empower biologists to analyze petabytes of sensor data from a network of remote microphones and cameras. This system, which is being used to monitor endangered species and ecosystems around the globe, has enabled an order of magnitude improvement in the cost effectiveness of such projects. This approach can be expanded to encompass a greater variety of sensor sources, such as drones, to monitor animal populations, habitat quality, and to actively deter wildlife from hazardous structures. I present a strategic vision for how data-driven approaches to conservation can drive iterative improvements through better information and outcomes-based funding mechanisms, ultimately enabling increasing returns on biodiversity investments.
Stephane Senecal, Orange
Titre : « Les techniques de Machine Learning au cœur du succès d’AlphaGo »
Résumé :
The game of Go has long been viewed as the most challenging of classic games for artificial intelligence owing to its enormous search space and the difficulty of evaluating board positions and of selecting moves. After a brief review of the game of Go itself, we will focus on how researchers at Google DeepMind introduced a new approach to computer Go that uses “value networks” to evaluate board positions for predicting game issue and “policy networks” to select moves.
These models are in fact deep neural networks and are trained or learnt by a novel combination of supervised learning from human expert games, and reinforcement learning techniques from games of self-play. Basic concepts of deep neural networks models and of reinforcement learning framework will be explained.
Google DeepMind researchers also introduce a new search algorithm that combines Monte Carlo simulation techniques with value and policy networks. Using this search algorithm, and without any lookahead search (as for classical computer Chess), their program AlphaGo defeated the human European Go champion by 5 games to 0 and a world-class human Go champion by 4 games to 1. This is the first time that a computer program has defeated human professional players in the full-sized game of Go, a feat previously thought to be at least a decade away.
Détails et inscriptions ici :Paris Machine Learning Meetup #1 Season 4
- Le deuxieme meetup sera centré sur Stan http://mc-stan.org/
Ce meetup est organisé avec Dataiku et aura lieu le Mardi 20 septembre. Il faut s’inscrire sur leur meetup pour cet evenement.
Program
Speaker: Eric Novik
Title: Introduction to Bayesian Inference with Stan and R
Program
Speaker: Eric Novik
Title: Introduction to Bayesian Inference with Stan and R
Stan is a modern, high-performance probabilistic programming language that interfaces with R, Python, Julie, Matlab, and Stata. Eric Novik, one of the organizers of the NYC Bayesian Data Analysis meetup and a founder of Stan Group, will be in town to give a brief talk on Stan.
We will build up a simple Stan model from scratch to demonstrate various parts of the Stan program and will also present a multi-level, overdispersed Poison model that can be used for pricing products in a retail setting. We fit the latter model with the rstanarm package, which can be thought of as a fully Bayesian equivalent to lme4.
After the talk, Eric and Stan core developers Daniel Lee and Michael Betancourt will answer questions from the audience and offer some thoughts on the future of statistical inference and Bayesian computing.
Détails et inscriptions ici : Paris Machine Learning Meetup: Hors Série #1 Season 4
- Le troisième meetup sera un Workshop H2O (Scalable Machine Learning with H2O) autour de leurs outils (http://h2O.ai ). Le meetup se fera chez Murex aura lieu le Mercredi 21 septembre
Jiqiong Qiu organizes the meetup with us.
Presentation of H2O -H20.ai- by Jo-fai Chow and Jakub Háva
1. Introduction to Machine Learning with H2O (Jo-fai Chow - 30 mins)
In this talk, I will walk you through our company (H2O.ai), our open-source machine learning platform (H2O) and use cases from some of our users. This will be useful for attendees who are not familiar with H2O.2. Project “Deep Water” (H2O integration with TensorFlow and other deep learning libraries) (Jo-fai Chow - 30 mins)
The “Deep Water" project is about integrating our H2O platform with other open-source deep learning libraries such as TensorFlow from Google. We are also working GPU implementation for H2O.In this talk about the motivation and potential benefits of our recent project named “Deep Water”.3. Sparkling Water 2.0 (Jakub Háva - 45 mins)
Sparkling Water integrates the H2O open source distributed machine learning platform with the capabilities of Apache Spark. It allows users to leverage H2O’s machine learning algorithms with Apache Spark applications via Scala, Python, R or H2O’s Flow GUI which makes Sparkling Water a great enterprise solution. Sparkling Water 2.0 was built to coincide with the release of Apache Spark 2.0 and introduces several new features. These include the ability to use H2O frames as Apache Spark’s SQL datasource, transparent integration into Apache Spark machine learning pipelines, the power to use Apache Spark algorithms via the Flow GUI and easier deployment of Sparkling Water in a Python environment. In this talk we will introduce the basic architecture of Sparkling Water and provide an overview of the new features available in Sparkling Water 2.0. The talk will also include a live demo showing how to integrate H2O algorithms into Apache Spark pipelines – no terminal needed!
Détails et inscriptions ici :Paris Machine Learning Meetup: Hors Série #2 Season 4
Important: Pour venir, il vous suffit de vous inscrire quand les meetups sont ouverts (en général une petite semaine avant). Igor et moi avons décidé de ne pas faire de liste d’attente parce que cela ne marche pas. Si il y a trop d’inscrits par rapport à la salle, seuls les premiers arrivés et les présentateurs pourront rentrer. Comme toujours, nous mettons toujours un système de Streaming en place et surtout toutes les présentations devraient être sur nos archives AVANT le meetup. Suivez le hashtag #MLParis pour savoir si il y a encore des places si vous pensez que vous serez en retard. Nous demandons à toute la communauté de jouer le jeu et a chacun de changer son RSVP en “non” pour que les autres puissent se dire s’ils ont une chance de pouvoir rentrer. En aucun un RSVP “oui” ou un RSVP “non” ne vous assure une place dans la salle. Le dynamisme de ce meetup est entre vos mains et la façon dont vous gérer vos RSVP nous permettra d’organiser ces événements uniques. Un très grand merci d’avance!
2. AI Labs (Franceis AI Septembre 16th-18th)
Antoine Dusséaux nous envoie le message suivant:
Si vous voulez changer le monde avec l’IA, rejoignez-nous à Paris les 16, 17 et 18 septembre prochains pour le plus important weekend startup sur l’intelligence artificielle afin d’y concrétiser vos projets de startups les plus fous.Le AI LABS c’est 54 heures, 100 talents sélectionnés, des experts renommés sur l’IA, un bootcamp avant l’événement, un lieu d’exception (#cloud.paris) et une expérience inoubliable. Pour plus d’infos et inscription >> www.ailabs.tech(Si vous souhaitez devenir ambassadeur ou volontaire pour l’événement, faites-le nous savoir !)
3. Corrélation n’est pas causalité
Après qu’Andrew Ng ait accepté de participer en fin de la saison 1 du Paris Machine Learning meetup, Baidu lui offrait la place de Chief Scientist d’un des plus gros moteurs de recherche du monde :-)
Sander Dieleman était d’abord un présentateur au Paris Machine Learning meetup en saison 2 . puis Google lui a offert un job à Deepmind et il a ainsi fait partie de l’équipe qui a construit AlphaGo et a permis de changer notre vision du monde. D’ailleurs, il raconte son histoire dans le blog de Kaggle :-)
Le CTO Movidius devait parler pour le meetup du hardware de la saison derniere au Paris Machine Learning mais n’avait pas pu pour des raisons d’emploi du temps. Movidius fait maintenant partie d’Intel depuis hier. Nous remarquons que c’est la deuxième entreprise qui était au line-up de ce meetup et qui s’est fait racheter. Nervana a en effet été acquise par Intel il y a un mois pour 350M$ :-)
Certains voient une relation de cause a effet entre la participation de certains speakers au meetup et leur trajectoire exceptionnelle après leur prestation au meetup. Igor et moi ne sommes pas sur, il nous faudrait plus de speakers, plus de salles, et plus d’entités qui sponsorisent nos événements de networking pour être vraiment sûr. Can you help ?
4. Autres meetups et autres news
Michael Benesty, un des orateurs de la saison 3 sur les algorithmes “Fair and Ethical” vient d’écrire sur Linkedin sur le fait que La qualité de l’anonymisation des décisions de justice par machine learning baisse de façon dratique en fonction du nombre de mentions à anonymiser
5. OPECST
Comme nous vous l’avions dit dans la dernière newsletter, l’Office parlementaire d’évaluation des choix scientifiques et technologiques (OPECST) du parlement Français a commencé à travailler sur l’intelligence artificielle en vue de rédiger un rapport. Les deux rapporteurs sont Madame la sénatrice Dominique Gillot, ancienne ministre, et Monsieur le député Claude de Ganay. Franck et moi avons mis en place un formulaire Google qui vous permet de donner vos impressions sur le sujet. Ce formulaire sera lisible directement par le rapporteur. Nous vous demandons de mettre un moyen de vous joindre au cas ou le rapporteur estime que vos écrits doivent être présenté et entendu par le parlement. Il serait optimal si vous mettiez vos avis avant le 10 septembre 2016. Le formulaire se trouve ici: https://goo.gl/forms/gNdyEiwTgCmvG7il2
6. Horse 2016 in London — free one day workshop
Si vous êtes à Londres le 19 Septembre, Bob Sturm un des speakers de la saison 3 organise un workshop:
HORSE 2016 is a free one-day workshop (with free coffee & nibbles and a free lunch) that will explore issues surrounding “horses” and “Potemkin villages” in applied machine learning. One of the most famous “horses” is the “tank detector” of early neural networks research (https://neil.fraser.name/writing/tank): after great puzzlement over its success, the system was found to just be detecting sky conditions, which happened to be confounded with the ground truth. Humans can be “horses” as well, e.g., magicians and psychics. In contrast, machine learning does not deceive on purpose, but only makes do with what little information it is fed about a problem domain. The onus is thus on a researcher to demonstrate the sanity of the resulting model; but too often it seems evaluation of applied machine learning ends with a report of the number of correct answers produced by a system, and not with uncovering how a system is producing right or wrong answers in the first place.
7. 8-week course: Machine Learning with Text in Python
Hello Paris Machine Learning!
I wanted to let you know about an upcoming course I’m teaching called “Machine Learning with Text in Python.”
In this 8-week Master Course (September 28 — November 22), you’ll learn how to build effective machine learning models using text-based data to solve your own data science problems. Topics include:
- Feature extraction from unstructured text using the scikit-learn library
- Model building and model evaluation
- Using Natural Language Processing techniques to improve your models
- Feature engineering from messy data sources using regular expressions (regex)
- Advanced machine learning techniques (randomized search, pipelines, model stacking, etc.)
This is an online course, but it’s limited to 50 students and has the “feel” of a classroom course: There are instructor-led webcasts twice a week, optional readings and homework assignments (with instructor feedback), a Slack team for interacting with the instructors and other students, and a private Kaggle competition to practice what you’ve learned. I’m an experienced data science classroom instructor, and I bring that same energy to the online environment!
Click here to learn more about the course, read comments from previous students, and watch a sample of the course content: http://www.dataschool.io/learn/
I’m holding an info session about the course on September 13: http://ccst.io/e/text-course
Feel free to get in touch with me, and I’d be happy to answer your questions!
Best,
Kevin Markham
Founder, Data School
8. Data Science Game
International student challenge
Final stage 10 & 11 september in Paris
Details here : http://www.datasciencegame.com
9. Kaggle released an open data set section
More here : https://www.kaggle.com/datasets
10. Facebook AI Research team released DeepMask and SharpMask
DeepMask and SharpMask are new state-of-the art deep learning algorithms that enable our computer vision systems to better detect different objects in an image, and also precisely delineate and label them.
Github repo of DeepMask/SharpMask:https://github.com/facebookresearch/deepmask
Research paper : http://arxiv.org/pdf/1603.08695v2.pdf
11. Baidu deep Learning Framework
Chinese tech giant Baidu released its Deep Learning Framework. Baidu announces Paddle at Baidu World in Beijing last week
Full release of the documentation is due for Sep 30th.
12. Présentation d’un talk — Sponsoring d’un meetup
Vous voulez présenter ? Il faut remplir ce formulaire.
Pour nous accueillir ou sponsoriser un event (catering : food / softs / drinks)
Contact : Franck and Igor
Toutes les dates des meetup de la saison 4: www.meetup.com/Paris-Machine-learning-applications-group
Liked this entry ? subscribe to Nuit Blanche's feed, there's more where that came from. You can also subscribe to Nuit Blanche by Email, explore the Big Picture in Compressive Sensing or the Matrix Factorization Jungle and join the conversations on compressive sensing, advanced matrix factorization and calibration issues on Linkedin.
No comments:
Post a Comment