Most machine learning (ML) methods are based on numerical mathematics (NM) concepts, from differential equation solvers over dense matrix factorizations to iterative linear system and eigen-solvers. As long as problems are of moderate size, NM routines can be invoked in a black-box fashion. However, for a growing number of real-world ML applications, this separation is insufficient and turns out to be a severe limit on further progress.

The increasing complexity of real-world ML problems must be met with layered approaches, where algorithms are long-running and reliable components rather than stand-alone tools tuned individually to each task at hand. Constructing and justifying dependable reductions requires at least some awareness about NM issues. With more and more basic learning problems being solved sufficiently well on the level of prototypes, to advance towards real-world practice the following key properties must be ensured: scalability, reliability, and numerical robustness. Unfortunately, these points are widely ignored by many ML researchers, preventing applicability of ML algorithms and code to complex problems and limiting the practical scope of ML as a whole.

Description and Motivation

Our workshop addresses the abovementioned concerns and limitations. By inviting numerical mathematics researchers with interest in *both* numerical methodology *and* real problems in applications close to machine learning, we will probe realistic routes out of the prototyping sandbox. Our aim is to strengthen dialog between NM and ML. While speakers will be encouraged to provide specific high-level examples of interest to ML and to point out accessible software, we will also initiate discussions about how to best bridge gaps between ML requirements and NM interfaces and terminology; the ultimate goal would be to figure out how at least some of NM's high standards of reliability might be transferred to ML problems.

The workshop will reinforce the community's awakening attention towards critical issues of numerical scalability and robustness in algorithm design and implementation. Further progress on most real-world ML problems is conditional on good numerical practices, understanding basic robustness and reliability issues, and a wider, more informed integration of good numerical software. As most real-world applications come with reliability and scalability requirements that are by and large ignored by most current ML methodology, the impact of pointing out tractable ways for improvement is substantial.

General Topics of Interest

A basic example for the NM-ML interface is the linear model (or Gaussian Markov random field), a major building block behind sparse estimation, Kalman smoothing, Gaussian process methods, variational approximate inference, classification, ranking, and point process estimation. Linear model computations reduce to solving large linear systems, eigenvector approximations, and matrix factorizations with low-rank updates. For very large problems, randomized or online algorithms become attractive, as do multi-level strategies. Additional examples include analyzing global properties of very large graphs arising in social, biological, or information transmissing networks, or robust filtering as a backbone for adaptive exploration and control.

In the current era of web-scale datasets, high throughput biology and astrophysics, and multilanguage machine translation, modern datasets no longer fit on a single computer and traditional machine learning algorithms often have prohibitively long running times. Parallelized and distributed machine learning is no longer a luxury; it has become a necessity. Moreover, industry leaders have already declared that clouds are the future of computing, and new computing platforms such as Microsoft's Azure and Amazon's EC2 are bringing distributed computing to the masses. The machine learning community has been slow to react to these important trends in computing, and it is time for us to step up to the challenge.

While some parallel and distributed machine learning algorithms already exist, many relevant issues are yet to be addressed. Distributed learning algorithms should be robust to node failures and network latencies, and they should be able to exploit the power of asynchronous updates. Some of these issues have been tackled in other fields where distributed computation is more mature, such as convex optimization and numerical linear algebra, and we can learn from their successes and their failures.

The goals of our workshop are:

1. To draw the attention of machine learning researchers to this rich and emerging area of problems and to establish a community of researchers that are interested in distributed learning.
2. To define a number of common problems for distributed learning (online/batch, synchronous/asynchronous, cloud/cluster/multicore) and to encourage future research that is comparable and compatible
3. To expose the learning community to relevant work in fields such as distributed optimization and distributed linear algebra.
4. To identify research problems that are unique to distributed learning.

Organizers

• John Langford   (Yahoo Research)
• Ofer Dekel  (Microsoft Research)
• John Duchi  (UC Berkeley and Google Research)
• Alekh Agarwal  (UC Berkeley)
• Lawrence Cayton  (MPI Tuebingen)

Program Committee

• Ron Bekkerman - LinkedIn
• Misha Bilenko - Microsoft    Ran Gilad-Bachrach - Microsoft
• Guy Lebanon - Georgia Tech
• Ilan Lobel - NYU
• Gideon Mann - Google
• Ryan McDonald - Google
• Ohad Shamir - Microsoft
• Alex Smola - Yahoo!
• S V N Vishwanathan - Purdue
• Martin Wainwright - UC Berkeley
• Lin Xiao - Microsoft

Today's data-driven society is full of large-scale datasets, e.g., images from the web, sequence data from the human genome, graphs representing friendship networks, time-series data of stock prices, speech corpora of news broadcasts, etc. In the context of machine learning, these datasets are often represented by large matrices representing either a set of real-valued features for each datapoint or pairwise similarities between datapoints. Hence, modern learning problems in computer vision, natural language processing, computational biology, and other areas often face the daunting task of storing and operating on matrices with thousands to millions of entries. An attractive solution to this problem involves working with low-rank approximations of the original matrix. Low-rank approximation is at the core of widely used algorithms such as Principle Component Analysis, Multidimensional Scaling, Latent Semantic Indexing, and manifold learning. Furthermore, low-rank matrices appear in a wide variety of applications including lossy data compression, collaborative filtering, image processing, text analysis, matrix completion and metric learning.

In the past several years, there has been a growing body of research devoted to developing randomized methods to efficiently and accurately generate low-rank approximations to large matrices. This topic has been studied in a wide range of disciplines including numerical linear algebra, theoretical computer science and machine learning. Although many of these techniques have extensive theoretical guarantees and empirical success on a limited number of datasets, there nonetheless remains a gap between these (mostly theoretical) advances and practical implementations of large-scale machine learning problems. Within the past few years, however, there have been significant advances in making these theoretical tools practical for large-scale machine learning. For example, high- quality numerical implementations have been developed, ensemble methods have been introduced to improve performance, and these techniques have been used to solve outstanding problems in human genetics and computer vision.

In this workshop, we aim to survey recent developments with an emphasis on usefulness for practical large-scale machine learning problems and to provide a forum for researchers to discuss several important questions associated with randomized low-rank approximation techniques, including: What are the state-of- the-art approximation techniques? How does the heterogeneity of data affect the randomization aspects of these algorithms? Which methods are appropriate for various machine learning tasks? How do these methods work in practice for large-scale machine learning tasks? What is the tradeoff between numerical precision and time/space efficiency in the context of machine learning performance, e.g., classification or clustering accuracy? In summary, we are interested in exploring the impact of low-rank methods for large-scale machine learning. We will study new algorithms, recent theoretical advances and large- scale empirical results, and more broadly we will motivate additional interesting scenarios for use of low-rank approximations for learning tasks.

Workshop Organizers

• Arthur Gretton, University College London - Gatsby
• Michael Mahoney, Stanford University
• Mehryar Mohri, Courant Institute (NYU) and Google Research
• Ameet Talwalkar, UC Berkeley

Program Committee

• Alexandre d'Aspremont, Princeton University
• Christos Boutsidis, Rensselear Polytechnic Institute
• Kamilika Das, NASA Ames Research Center
• Maryam Fazel, Unversity of Washington
• Michael I. Jordan, UC Berkeley
• Sanjiv Kumar, Google Research New York
• James Kwok, Hong Kong University of Science and Technology
• Gunnar Martinsson, University of Colorado Boulder

Computational photography (CP) is a new field that explores and is about to redefine how we take photographs and videos. Applications of CP are not only "everyday" photography but also new methods for scientific imaging, such as microscopy, biomedical imaging, and astronomical imaging, and can thus be expected to have a significant impact in many areas. There is an apparent convergence of methods, what we have traditionally called "image processing", and recently many works in machine vision, all of which seem to be addressing very much the same, if not tightly related problems. These include deblurring, denoising, and enhancement algorithms of various kinds.

• What do we learn from this convergence and its application to CP?
• Can we create more contact between the practitioners of these fields, who often do not interact?
• Does this convergence mean that the fields are intellectually shrinking to the same point, or expanding and hence overlapping with each other more?

Besides discussing such questions, the goal of this workshop is two-fold:

• (i) to present the current approaches, their possible limitations, and open problems of CP to the NIPS community, and
• (ii) to foster interaction between researchers from machine learning, neuro science and CP to advance the state of the art in CP.

The key of the existing CP approaches is to combine (i) creative hardware designs with (ii) sophisticated computations, such as e.g. new approaches to blind deconvolution. This interplay between both hardware and software is what makes CP an ideal real-world domain for the whole NIPS community, who could contribute in various ways to its advancement, be it by enabling new imaging devices that are possible due to the latest machine learning methods or by new camera and processing designs that are inspired by our neurological understanding of natural visual systems. Thus the target group of participants are researchers from the whole NIPS community (machine learning and neuro science) and researchers working on CP and related fields.

Research on Multiple Kernel Learning (MKL) has matured to the point where efficient systems can be applied out of the box to various application domains. In contrast to last year's workshop, which evaluated the achievements of MKL in the past decade, this workshop looks beyond the standard setting and investigates new directions for MKL.

In particular, we focus on two topics:

1. There are three research areas, which are closely related, but have traditionally been treated separately: learning the kernel, learning distance metrics, and learning the covariance function of a Gaussian process. We therefore would like to bring together researchers from these areas to find a unifying view, explore connections, and exchange ideas.
2. We ask for novel contributions that take new directions, propose innovative approaches, and take unconventional views. This includes research, which goes beyond the limited classical sum-of-kernels setup, finds new ways of combining kernels, or applies MKL in more complex settings.

Organizers

• Marius Kloft, UC Berkeley
• Ulrich Rückert, UC Berkeley
• Cheng Soon Ong, ETH Zürich
• Alain Rakotomamonjy, University of Rouen
• Sören Sonnenburg, FML of the Max Planck Society / TU Berlin
• Francis Bach, INRIA / ENS

Discrete optimization problems and combinatorial structures are becoming increasingly important in machine learning. In the endeavour to process more and more complex data, one may quickly find oneself working with graphs, relations, partitions, or sparsity structure. In addition, we may want to predict structured, sparse estimators, or combinatorial objects such as permutations, trees or other graphs, group structure and so on.

While complex structures enable much richer applications, they often come with the caveat that the related learning and inference problems become computationally very hard. When scaling to large data, combinatorial problems also add some further challenges to compact representations, streaming and distributed computation.

Despite discouraging theoretical worst-case results, many practically interesting problems can be much more well behaved (when modeled appropriately), or are well modeled by assuming properties that make them so. Such properties are most often structural and include symmetry, exchangeability, sparsity, or submodularity. Still, not all of these structures and their application ranges are well understood. The DISCML workshop revolves around such structures in machine learning, their applications and theory.

Organizers

• Andreas Krause, ETH Zurich
• Jeff A. Bilmes, University of Washington
• Stefanie Jegelka, UC Berkeley

The workshop will bring together researchers from vision, learning and related areas to present and discuss the recognition of spatio-temporal motion of people across a broad range of application areas ranging from sign language recognition through to gesture and activity.

Programme Committee

• Aleix Martinez, Ohio State University
• Andrew Zisserman, Oxford, UK
• Bastian Leibe, RWTH Aachen University
• Bernt Schiele, TU Darmstadt, Germany
• Bjorn Stenger, Toshiba Research, UK
• Christian Vogler, ILSP, Greece
• Christophe Collet, IRIT, France
• Edmond Boyer, INRIA, France
• Fernando De la Torre, CMU, USA
• Hermann Ney, RWTH Aachen University, Germany
• Iain Matthews, Disney Research, USA
• Ivan Laptev, INRIA, France
• Josef Kittler, Uni of Surrey, UK
• Justus Piater, Université de Liège, Belgium
• Mark Everingham, Uni of Leeds, UK
• Matthew Turk, Uni of California, USA
• Milos Zelezny, West Bohemia, Czech
• Nicu Sebe, University of Trento, Italy
• Petros Maragos, NTUA, Greec
• Philippe Dreuw, RWTH Aachen University, Germany
• Piotr Dollar, Caltech, USA
• Rama Chellappa, Uni of Maryland
• Richard Bowden, Surrey University, UK
• Shaogang Gong, QMUL, UK
• Stan Sclaroff, Boston University, USA
• Thomas B. Moeslund, AAU, Denmark
• Thomas Deselaers, ETH Zurich, Switzerland
• Vassilis Athitsos, Uni of Texasm Arlington, USA
• Vittorio Ferrari, ETH Zurich, Switzerland

This Stream at EURO 2010 (European Conference on Operational Research) will consist of the following 5 sessions (each with 3/4 talks):

• Mathematical Programming Approaches for Classification problems
• Bioinformatics Applications of Machine Learning
• Machine Learning to help people with disabilities
• Machine Learning for Multiple Sources
• Neural Network Applications

In statistics and machine learning, the evaluation of algorithms typically relies on their performance on data. This is because, in contrast to a theoretical guarantee (e.g. a consistency result), it is in general not possible to prove that an algorithm performs well on a particular (unseen) data set. Therefore, it is of vital importance that we ensure the reliability of data-based evaluations. This requirement poses a wide range of open research problems and challenges. These include

1. the lack of a ground truth to validate results in real-world applications,
2. the high instability of empirical results in many settings,
3. the difficulty to make statistics and machine learning research reproducible,
4. the general over-optimism of published research findings due pre-publication optimization of the algorithms and publication bias.

This workshop brings together scientists from statistics, machine learning, and their application fields to tackle these challenges. The workshop serves as a platform to critically discuss current shortcomings, to exchange new approaches, and to identify promising future directions of research.

Organizing Committee

• Nicole Krämer, WIAS Berlin
• Anne-Laure Boulesteix, LMU Munich

Pattern Analysis and Statistical Learning cover a wide range of technologies and theoretical frameworks, and significant activity in the past years has resulted in a remarkable convergence and many advances in the theory and principles underlying the field.

Bringing these technologies to real world demanding applications is however often treated as a separate problem, one that does not directly affect the field as a whole. It is instead important to consider the field of Pattern Analysis as fully including all issues involved with the applications of this technology, and hence all issues that arise when deploying, scaling, implementing and using the technology.

We call for constributions in the form of Demos, Case Studies, Working Systems, Real World Applications and Usage Scenarios. Challenges may stem from the violation of common theoretical assumptions, from the specific types of patterns and noise arising in certain scenarios, or from the problem of scaling up the implementation of state of the art algorithms to real world sizes, or from the creation of integrated software systems that contain multiple pattern-analysis components.

We are also interested in new application areas, where Pattern Analysis has been deployed with success, and in issues involving the visualisation and delivery and exploitation of the patterns discovered by PA technologies. Systems working in noisy and unstructured environments and situations are particularly interesting.

The goal is to discuss and reward work aimed at making theory useful and relevant, without requesting the researchers to propose new theoretical methods, but rather requesting to show how they solved the many challenges related to applying these methods to real world scenarios, or how they benefited other fields of research. Getting ideas to work in real scenarios is what this is about.

Workshop Organisers

• Nello Cristianini, University of Bristol
• Tom Diethe, University College London
• John Shawe-Taylor, University College London