ICML Workshop on Numerical Methods in Machine Learning: Call for Contributions

ICML Workshop on Numerical Methods in Machine Learning: Call for Contributions

Posted by Rebecca Martin on Tue, 31/03/2009 - 00:00

International Conference on Machine Learning (ICML)
Workshop on Numerical Mathematics in Machine Learning
June 18, 2009. Montreal, Canada
Deadline for abstract submissions: April 27, 2009


Most machine learning (ML) algorithms rely fundamentally on concepts of numerical mathematics. Standard reductions to black-box computational primitives do not usually meet real-world demands and have to be modified at all levels. The increasing complexity of ML problems requires layered approaches, where algorithms are components rather than stand-alone tools fitted individually with much human effort. In this modern context, predictable run-time and numerical stability behavior of algorithms become
fundamental. Unfortunately, these aspects are widely ignored today by ML researchers, which limits the applicability of ML algorithms to complex problems.

Our workshop aims to address these shortcomings, by trying to distill a compromise between inadequate black-box reductions and highly involved complete numerical analyses. We will invite speakers with interest in *both* numerical methodology *and* real problems in applications close to machine learning.
While numerical software packages of ML interest will be pointed out, our focus will rather be on how to best bridge the gaps between ML requirements and these computational libraries. A subordinate goal will be to address the role of parallel numerical computation in ML.

Examples of machine learning founded on numerical methods include low level computer vision and image processing, non-Gaussian approximate inference, Gaussian filtering / smoothing, state space models, approximations to kernel methods, and many more.

The workshop will comprise a panel discussion, in which the invited speakers are urged to address the problems stated above, and offer individual views and suggestions for improvement. We highly recommend active or passive attendance at this event. Potential participants are encouraged to contact the organizers beforehand, concerning points they feel should be addressed in this event.

Invited Speakers:

Inderjit Dhillon University of Texas, Austin
Michael Mahoney Stanford University
Jacek Gondzio Edinburgh University, UK

[Further speaker to be confirmed]


Potential short talks / posters should aim to address:

- Raising awareness about the increasing importance of stability and predictable run-time behaviour of numerical machine learning algorithms and primitives
- Stability and predictable behaviour as a criterion for making algorithm choices in machine learning
- Lessons learned (and not learned) in machine learning about numerical mathematics. Ideas for improvement
- Novel developments in numerical mathematics, with potential impact on machine learning problems

Contributions will be considered only if a clear effort is made to analyze problems that arise, and if choices of algorithms, preconditioning, etc. are clearly motivated. For reasons stated in "Motivation", submissions that apply numerical methods in a black box fashion, or that treat numerical techniques without motivating the use for machine learning, cannot be considered. The usual "smoothing over problems" conference paper style is discouraged, and naming and analyzing problems is strongly encouraged.

Potential Subtopics (submissions are not limited to these):

A- Solving large linear systems
Arise in the linear model/Gaussian MRF (mean computations), nonlinear optimization methods (Newton- Raphson, IRLS, ...)
- Preconditioning, use of model structure.
Our main interest is on semi-generic ideas that can be applied to a range of machine learning real-world situations

B- Novel numerical software developments relevant to ML
- Parallel implementations of LAPACK, BLAS
- Sparse matrix packages

C- Approximate eigensolvers
Arise in the linear model (covariance estimation), spectral clustering and graph Laplacian methods, PCA
- Lanczos algorithm and specialized variants
- Preconditioning

D- Exploiting matrix/model structure, fast matrix-vector multiplication
- Matrix decompositions/approximations
- Multi-pole methods
- Signal-processing primitives (e.g., variants of FFT)

F- Parallel numerical computation for ML

G- Other numerical mathematics (ODEs, PDEs, Quadrature, etc.) focusing on machine learning

Submission Instructions:

We invite submissions of extended abstracts, from 2 to 4 pages in length (using the ICML 2009 style file). Criteria for content are given in "Topics". Submissions should be sent to suvadmin@googlemail.com

Accepted contributions will be allocated short talks or posters. There will be a poster session with ample time for discussion. Short talk contributions are encouraged to put up posters as well, to better address specific questions.

Important Dates:

Submissions due: April 27, 2009
Author notification: May 11, 2009
Workshop date: June 18, 2009

Matthias W. Seeger MPI Informatics / Saarland University, Saarbruecken
Suvrit Sra MPI Biological Cybernetics, Tuebingen
John P. Cunningham Stanford University (EE), Palo Alto

We acknowledge financial support through the PASCAL 2 Initiative of the European Union.