AutoML 2015 workshop @ ICML 2015

samedi 11 juil. 2015, 08:30 → 18:00 Europe/Paris

Lille Grand Palais

Balázs Kégl (LAL), Frank Hutter (University of Freiburg)

The web site of the event:

https://sites.google.com/site/automlwsicml15/

Please submit the questions you would like to raise in the panel discussion at this site.

08:30 → 10:00 Session 1

08:30 Invited Talk: Open Research Problems in AutoML

Orateur: Rich Caruana (Microsoft Research)

Slides CaruanaAutoMLWorkshopICML2015rev4.pdf

09:10 Invited Talk: Bandits and Bayesian optimization for AutoML

Complex optimization and decision making tasks are beginning to play an increasingly crucial role across a wide variety of scientific fields. This is becoming more and more evident as entire research programs are being automated. In this talk I'll describe a set of methods, known as Bayesian optimization, which provide a very sample efficient approach to this problem. Much of the gains of these methods are obtained by building a posterior model of a function during optimization in order to efficiently explore its surface. I will further describe a number of advanced search mechanisms and models and show how these can be used for automating Machine Learning problems. Finally, I will also briefly provide links to related bandit literature.

Orateur: Matthew Hoffmann (University of Cambridge)

Slides slides_bayesian_optimization.pdf

09:50 Poster Spotlights 1

5 spotlights of 2 minutes each

10:00 → 10:30 Coffee break

10:30 → 12:00 Session 2

10:30 Invited Talk: Algorithm Recommendation as Collaborative Filtering

Orateur: Michele Sebag (CNRS)

Slides AutoML_CollaborativeFiltering_15_Sebag.pdf

11:10 Poster spotlights 2

9 spotlights of 2 minutes each

11:30 1st Poster Session

12:00 → 14:00 Lunch Break

14:00 → 16:00 Session 3

14:00 Invited Talk: Recursive Self-Improvement

Most machine learning researchers focus on domain-specific learning algorithms. Can we also construct meta-learning algorithms that can learn better learning algorithms, and better ways of learning better learning algorithms, and so on, restricted only by the fundamental limitations of computability? In 1965, J. Good already made informal remarks on an intelligence explosion through such recursive self-improvement (RSI). I will discuss various concrete algorithms (not just vague ideas) for RSI: 1. My diploma thesis (1987) proposed an evolutionary system that learns to inspect and improve its own learning algorithm, where Genetic Programming (GP) is recursively applied to itself, to invent better learning methods, meta-learning methods, meta-meta-learning etc. 2. RSI based on the self-referential Success-Story Algorithm for self-modifying probabilistic programs (1997) was already able to solve complex tasks. 3. My self-referential deep recurrent neural networks (since 1993) run and inspect and change their own weight change algorithms. Back in 2001, my former student Hochreiter (now prof) already had a practical implementation of such an RNN that meta-learns an excellent learning algorithm, at least for a limited domain. 4. The Goedel machine (2006) is the first RSI that is mathematically optimal in a particular sense. Will RSI finally take off in the near future?

Orateur: Juergen Schmidhuber (IDSIA)

Slides recursive-self-improvement.pdf

14:40 Invited Talk: Automatically constructing models, and automatically explaining them, too.

How could an artificial intelligence do statistics? It would need an open-ended language of models, and a way to search through and compare those models. Even better would be a system that could explain the different types of structure found, even if that type of structure had never been seen before. This talk presents a prototype of such a system, which builds structured Gaussian processes regression models by combining covariance kernels to build a custom model for each dataset. The resulting models can be broken down into relatively simple components, and surprisingly, it's not hard to write code that automatically describes each component, even for novel combinations of kernels. The result is a procedure that takes in a dataset, and outputs a report with plots and English descriptions of the different types of structure found in that dataset.

Orateur: David Duvenaud (Harvard University)

Slides automl_(4).pdf

15:20 2nd Poster Session

16:00 → 16:30 Coffee break

16:30 → 18:00 Session 4

16:30 Invited Talk: OpenML: A Foundation for Networked & Automatic Machine Learning

OpenML is an online machine learning platform where scientists can automatically log and share data sets, code, and experiments, organize them online, and collaborate with researchers all over the world. It helps to automate many tedious aspects of research, is readily integrated into several machine learning tools, and offers easy-to-use APIs. It also enables large-scale and real-time collaboration, allowing researchers to build directly on each other's latest results, and track the wider impact of their work. Ultimately, this provides a wealth of information for building systems that learn from previous experiments, to either assist people while analyzing data, or automate the process altogether.

Orateur: Joaquin Vanschoren (Eindhoven University of Technology)

Slides OpenML_AutoML.pdf

17:10 AutoML Challenge

Orateur: Marc Boulle (Orange)

17:30 Panel Discussion: Next steps for AutoML

Panelists: Marc Boulle, Rich Caruana, David Duvenaud, Matthew Hoffmann, Juergen Schmidhuber, Michèle Sebag, Joaquin Vanschoren.