PPSN 2020 Workshop on Understanding Machine Learning Optimization Problems


Understanding Machine Learning and Optimization Problems (UMLOP)


-- to be held hybrid (onsite/online) at PPSN 2020 --



 


Free Access to the Workshop


The workshop will be made publicly available (for free). If you would like to attend the workshop, contact us via e-mail (e.g., kerschke@uni-muenster.de) and we will send you the link.


 


Schedule (Sunday, September 6, 09:00am - 10:30pm)


  • 09:00 -- 09:05 Marcus Gallagher (University of Queensland, Australia) (online) and Mike Preuss (Leiden University, The Netherlands) (onsite)
    • Introduction to UMLOP 2020
  • 09:05 -- 09:25 Katherine Malan (University of South Africa, South Africa) (online)
    • Recent Developments in the Landscapes of Neural Networks
    • Abstract: Neural network training can be viewed as an optimisation problem and the resulting search spaces can be analysed to help us understand why some training algorithms perform better than others on particular problems. There are, however, some critical differences between error landscapes when compared to fitness landscapes of black box optimisation problems. Some of these differences include unbounded search spaces, availability of gradient information, landscapes that depend on the choice of data set and loss function, difference between training and testing landscapes, and so on. This talk will briefly discuss some of these peculiarities of neural network error landscapes and outline recent advances in this exciting new area of research.
  • 09:25 -- 09:45 Marius Lindauer (Leibniz University of Hannover, Germany) (online)
    • Automated Deep Learning
    • Abstract: Automated Deep Learning (AutoDL) supports users, developers and researchers of deep learning by searching for well-performing architectures of neural networks and the hyperparameters to train them. In my talk, I give a short overview about the characteristics of the AutoDL problem itself and what we learned about it from our recent benchmark on learning curves (so-called LCBench)
  • 09:45 -- 10:05 Nacim Belkhir (Safran, France) (online)
    • Landscape of Real-World Black Box Problems: The Aeronautic Case
    • Abstract: With the growth of the air traffic, the aeronautic industry is drastically evolving by including numerical methods at the all level from simulation to automated control in the manufacturing chain. While several problems are usually solved using domain expert knowledge, many recent challenges rely on complex and computationally expensive tasks that are preferred to be tackled using data driven or black box formulations. In this talk, we give an overview of the aeronautic challenges focusing on black box problems.
  • 10:05 -- 10:25 Gerard van Westen (Leiden University, The Netherlands) (onsite)
    • Artificial Intelligence and Medicinal Chemistry -- Highlights, Challenges, and Opportunities
    • Abstract: Medicinal Chemistry is changing, the catalytic effect of data science on drug discovery cannot be denied. History dictates that these new tools will likely be a synergistic addition to the drug discovery process rather than a revolutionary replacement of existing methods (similar to throughput screening, combinatorial chemistry, or structure-based drug discovery).
  • 10:25 -- 10:30 Olivier Teytaud (Facebook AI Research, France) (online) and Pascal Kerschke (University of Münster, Germany) (onsite)
    • Q&A, Discussion, and Wrap-Up of UMLOP 2020

 


Scope of the Workshop


Although recent works have shown how machine learning (ML) methods can actually benefit from incorporating evolutionary computation (EC) strategies, there still exists only little knowledge exchange between the two communities.


Our workshop tries to reduce the gap between ML and EC by discussing common ML problems in more detail with the overall goal of improving the understanding of their specific structure. Moreover, we want to investigate how and when EC could be used to solve certain ML tasks. Thus, we are interested in questions such as:


  • What do the landscapes of ML problems look like?
  • Which ML problems are of benign/malign nature for EC approaches (i.e., when should one consider incorporating EC into ML methods)?
  • How similar are such problems to instances from popular EC benchmark suites?

 


Topics of Interest


In order to facilitate first analyses, we provide an initial set of ML problems (related to real-world applications) and invite participants to contribute their insights into these problems. Submission could contain -- but are definitely not limited to -- any of the following related issues:


  • applying exploratory landscape analysis to one of the aforementioned ML problems,
  • analyzing algorithm (including portfolio) performances with a focus on "why",
  • investigating existing problem solutions w.r.t. quality and/or diversity aspects,
  • discussing (the feasibility of) commonly used performance metrics within these domains, or
  • extending the provided set of test problems by contributing further optimization problems (from the ML domain), which are of relevance for real-world applications.

 


Workshop Organizers






 


 


Disclaimer Regarding Covid-19


This workshop is part of the PPSN 2020 in Leiden, The Netherlands. Due to the restrictions resulting from the Covid-19 pandemic, it will be held as a mixture of onsite and online talks.


 


Further Notes


Please help us raising awareness for this workshop and distribute the link to this website (http://www.erc.is/go/umlop2020).