Sunday, June 15, 2014

Europe Wide Machine Learning Meetup and Paris Machine Learning #12: Season 1 Finale, Andrew Ng and More...

Franck and I were initially discussing what a season finale for the Paris Machine Learning Meetup would look like. After some back and forth discussion, an idea emerged: Since the remote presentations have been a great add-on for the whole season, what about asking a big name in ML to join us and share some insights. But that was too easy: what about sharing that exchange with three other European meetups since most of us are in the same time zone ? Deal.

We then contacted Martin Jaggi in Zurich, then Daniel Nouri and Andreas Muller in Berlin and finally Jurgen Van Gael, Jacqueline Forien and Martin Goodson in London who all said yes. Ok this is good, what about the big name ? We did not go far on our common list, in fact we stayed on the first name: Andrew Ng. Quite a few of our meetup participants have taken courses at Coursera, so Andrew was a natural choice. The Email was sent and Andrew Ng was kind enough to answer positively in less than a week. Bingo !

Since then, Andrew has gotten a position as Chief Scientist at Baidu, coincidence?! I think not :-)

So today or tomorrow (whenever you read this blog entry) i.e. Monday June 16th, there will be four Machine Learning Meetups across Europe and they will all share a session with Andrew. The streaming will be here:

The YouTube Link is here. The streaming is for the Paris meetup, which means that it starts at 7:00PM Paris time with Andrew coming up at 7:30PM Paris time.

We aim for a twitter hashtag to be: #MLEurope and #MLParis

it looks smooth and all but as I am writing this there are still many ways Murphy's law could strike. Fingers crossed.

Each of the meetups have their own programs aside from Andrew's presentation. Here they are:


Naturally, all the presentations are available on the Paris Meetup archive page.

18h30+ Ouverture des portes. Badging, Pizzas, Champagne,....

19h00+ Introduction: Franck Bardol, Igor Carron

19h05+ Bastien Legras, Google Cloud Platform Solution Engineer and Francois Sterin, Principal, Global Infrastructure
How Google Uses Machine Learning And Neural Networks To Optimize Data Centers, and how you can benefit from it with Google Cloud Platform"

19h30+ Andrew Ng (Coursera, Baidu Chief Scientist), Deep Learning: Machine learning and AI via large-scale neural networks 

20h10+ Muthu Muthukrishnan, On Sketching

20h30+ Yaroslav Bulatov (Google SF), CNN and Google Streetview.

20h45+ Camille Couprie (IFPEN), Semantic scene labeling using feature learning (Joint work with Clément Farabet, Laurent Najman and Yann LeCun)
In this talk, we address the problem of assigning an semantic category to every pixel of an image, or video. We introduce a model architecture that allows us to learn hierarchies of multi-scale features while being computationally efficient. We present results on different datasets, including one that contains depth information, which may be handled in our trainable model very easily. As the output predictions may be noisy in videos, we propose a temporal smoothing method using minimum spanning trees, providing an efficient solution for embedded, real-time applications.
21h00+ Sam Bessalah, Abstract algebra for stream learning.
A quick introduction into some common algorithms and data structures to handle data in streaming fashion, like bloom filters, hyperloglog or min hashes. Then in a second part how abstract algebra with monoids, groups or semi groups help us reason and build scalable analytical systems beyond stream processing. 

All the slides are already here.


Andrew Ng, co-founder of Coursera, director of the Stanford AI Lab, soon Chief Scientist at Baidu (via video, starting at 19:30)

• Apache Giraph for Applications in Machine Learning & Recommendation Systems, Maria Stylianou, HPC Software Engineer at Novartis (start 19:00)

Abstract: Over the last years, companies have turned to big data analytics to better understand their customers and drive their services according to customers' needs. In many cases, data is represented in graphs, for instance, describing user connections in a social network or user-item ratings in an online retailer. So far, Hadoop has been the swiss army knife of analytics, but has proven to be inefficient for graph mining and machine learning algorithms. This gave rise to a new generation of processing systems designed for this type of analytics. Apache Giraph is a representative of such systems. In this presentation, we showcase Giraph, its model and characteristics. We then demonstrate its suitability with machine learning algorithms and finally walk through an example algorithm for recommendation systems built on top of Giraph. 
• Smarter than Smart Sharpen - Advances in Image Deconvolution in Digital and Computational Photography, Michael Hirsch, Postdoc at University College London and the Max Planck Institute in Tübingen (start around 20:30) 

Abstract: Image Deconvolution is a key area in signal and image processing, that is receiving an ever increasing interest from the academic as well the industrial world due to both its theoretical and practical implications. It involves many challenging problems, including modeling the image formation process, formulating tractable priorsincorporating generic image statistics, and devising efficient methods for optimization. This renders image deconvolution an intriguing but also intricate task, which has recently seen much attention as well as progress in both the image and signal processing but also the computer vision and graphics community.
In this talk I will present some recent advances, that not only help sharpen blurry photos but might also change the design of future cameras.

We'll also keep you updated about the outcome of the Germany-Portugal soccer game of course!


1. Deep Learning: Machine learning and AI via large-scale neural networks (Andrew Ng, 40min) 

The co-founder of Coursera, director of the Stanford AI Lab and general Machine Learning legend is going to join us (as well as three other European ML meetups) via video.
2. Deep Neural Nets Study Group (Marcel Ackermann, 10min)

Marcel is going to present the new Deep Learning study group here in Berlin.

3. Kaggle contests (Abhishek Thakur, 5min) 

4. Structured Prediction and PyStruct (Andreas Mueller, 45min) 

Structured Prediction is a generalization of classification and regression to structured output spaces like sequences or trees. Unlike standard classification algorithms, structured prediction methods can exploit correlation in output variables.
The general form of a structured predictor f is f(x) = argmax_y g(x, y) where g is a compatibility function between an output structure y and the input x. Most methods for learning structured prediction are closely related to graphical models, which also have the constraint of g(x, y) being a probability distribution over y (in the conditional case).
The talk will give an introduction into structured prediction methods and the library PyStruct that implements many common learning algorithms in python. 

A special edition of London Machine Learning Meetup where we join forces with our London, Paris, Berlin and Zurich colleagues to hear from Andrew Ng, known to many of us for his Coursera introductory Machine Learning course, and recently announced as Baidu's chief scientist. 

Andrew Ng will present Deep Learning: Machine Learning and AI via large-scale neural networks during 40 minutes.

We are delighted to be able to announce that Andrew's webcast will be followed with a special panel of UK-based academics who will each present for 15 minutes and allow plenty of time for Q&A. 

Our panelists:

David Barber, Author of 'Bayesian Reasoning and Machine Learning', Reader in Computational Statistics and Machine Learning at University College London.

Deep and Shallow
David will discuss aspects related to deep and shallow learning. He will show data compression results that outperform standard deep learning methods and at a fraction of the computational cost.

David Duvenaud, Final year Ph.D. candidate at Cambridge University

Visualizing Priors on Deep Functions 
David will discuss properties of deep, infinitely-wide neural nets giving rise to deep Gaussian processes and deep kernels. David will visually explore these relatively simple models, and show what happens when one performs dropout in Gaussian processes.

No comments: