Temporal Segmentation: Perspectives from Statistics, Machine Learning, and Signal Processing Workshop

Data with temporal (or sequential) structure arise in several applications, such as speaker diarization [FSJW08b, FDH08], human action segmentation [ZTH08], network intrusion detection [TRBK06], DNA copy number analysis [LXZ08], and neuron activity modelling [Y07], to name a few.

A particularly recurrent temporal structure in real applications is the so-called change-point model [BH92], where the data may be temporally partitioned into a sequence of segments delimited by change-points, such that a single model holds within each segment whereas different models hold accross segments. Change-point problems may be tackled from two points of view, corresponding to the practical problem at hand: retrospective (or "a posteriori"), aka multiple change-point estimation [F06], where the whole signal is taken at once and the goal is to estimate the change-point locations [BKLMW09], and online (or sequential), aka quickest detection [PH09], where data are observed sequentially and the goal is to quickly detect change-points. We refer to these classes of tasks as temporal segmentation.

An extensive literature has developed in these two viewpoints, in both the statistics (and probability) community [L01], and in the signal processing community [K98]. Many of the optimal algorithms proposed in this literature were developed under rather restrictive assumptions, however: parametric models for distributions, low-dimensional multivariate data, and, in the online case, perfect knowledge of the pre- and post-change distributions.

In applications such as human action segmentation or speaker diarization, data are large-scale, expensive to label, and high-dimensional, therefore requiring approaches that can tackle more complex situations in temporal segmentation. Recent years have witnessed new approaches with broader applicability, essentially by proposing unsupervised [XWSS06, ZTH08], nonparametric [FSJW08b], and scalable temporal segmentation algorithms [FL07, FDH08].

The purpose of this workshop is to bring together experts from the statistics, machine learning, signal processing communities, to address a broad range of applications from robotics to neuroscience, to discuss and cross-fertilize ideas, and to define the current challenges in temporal segmentation. We intend to encourage discussions on the following particular issues: How can traditional statistical approaches for temporal segmentation, essentially generative, be extended to discriminative approaches, allowing us to deal with high-dimensional data? How well do unsupervised approaches for temporal segmentation perform with respect to supervised ones? What are the main statistical and computational issues that arise when addressing large-scale (long) data signals?

Organizers

• Zaid Harchaoui (primary organizer, zaidh@andrew.cmu.edu)
• Stephane Canu
• Olivier Cappe
• Arthur Gretton
• Alain Rakotomamonjy
• Jean-Philippe Vert