Frederic Dembak, Franck Bardol and myself are happy to present this 9th edition of the Paris Machine Leanring Meetup this coming Wednesday where we should have four presentations and demos from Danny Bickson (GraphLab), Julien Darlay (LocalSolver), Laurent Revel ( Import.io ) and myself. Here are the abstracts:
From social networks, to protein molecules and the web, graphs encode structure and context, enable advanced machine learning, and are rapidly becoming the future of big-data. In this talk we will present the next generation of GraphLab, an open-source platform and machine learning framework designed to process graphs with hundreds of billions of vertices and edges on hardware ranging from a single mac-mini to the cloud.
LocalSolver is a useful tool for solving large-scale mixed-variable non-convex optimization problems. It was first designed for combinatorial problems and is now able to handle continuous decisions. Based on various optimization techniques (localsearch, linear programming, constraint programming...) LocalSolver scales up to millions of variables, running on standard computers. It provides high quality solutions in short running times using a mathematical description of the problem. Several machine learning problems can be viewed as continuous or combinatorial optimization problems where the objective is to find a model or a hypothesis minimizing a loss function that is evaluated on a sample set. For instance, feature selection is the selection of a minimal subset of attributes describing your data such that observations from different classes remain distinct. Another example is the K-means clustering problem that is looking for a partition of a set of observations into K clusters minimizing the distance between each observation and the center of its cluster. Both examples can be modeled as optimization problems and solved byLocalSolver. In this presentation we will describe what is under the hood of LocalSolver using examples from industrial optimization applications and classical problems from machine learning.
Demonstration of import.io easy scraper capabilities.
+ Advanced Matrix Factorizations and Machine Learning, Igor Carron
Many of the techniques used in Machine Learning rest on a set of advanced matrix factorizations. However, most of these factorizations are combinatorial in nature and can naively be solved through brute force and for small problems. In recent years, however, many heuristics have appeared that allow these factorizations to be computable for large scale problemes. In this talk, I will try to provide a panorama of these heuristics and some of their use in signal processing and machine learning.