The conference will feature two panel discussions:
- Panel Session 1: Evaluation: SiSEC 2010 and remaining challenges
- Panel Session 2: The future of latent variable analysis and signal separation
Video of the Panel Session
We will first report the outcomes of the second community-based Signal Separation Evaluation Campaign (SiSEC 2010). SiSEC 2010 involved 7 audio datasets including real-world noisy and dynamic environments and, as a novel feature compared to the first edition, 3 biomedical datasets including fMRI and EEG data. The campaign featured 18 entrants and 33 algorithms for the audio tasks and 3 entrants for the biomedical tasks. We will present all the datasets, tasks, and achieved results and answer comments or questions by the audience.
We will then discuss the remaining open problems for the evaluation of source separation algorithms in various application areas. For example, evaluating source separation in the context of EEG or MEG is still particularly difficult because a) in any real measurement the ground truth is not known, and b) any simulation is prone to miss important aspects of real data. Meanwhile, in the context of audio, objective evaluation measures and stable measures for short recordings are still required. A set of auditory-motivated measures was used within SiSEC 2010.
We will compare the results obtained with these measures and with previous energy ratio-based objective measures. Of course, we will also discuss the questions raised by the audience about remaining challenges in audio/biomedical source separation and other areas.
All of your comments will make the panel a success, and open a bright future for source separation. Please prepare your questions!
|ID paper||Title and authors|
|55||The 2010 Signal Separation Evaluation Campaign (SiSEC2010): - Audio source separation -|
|Shoko Araki, Alexey Ozerov, Vikrham Gowreesunker, Hiroshi Sawada, Fabian Theis, Guido Nolte, Dominik Lutter, Ngoc Duong |
|60||The 2010 Signal Separation Evaluation Campaign (SiSEC2010): - Biomedical source separation -|
|Shoko Araki, Fabian Theis, Guido Nolte, Dominik Lutter, Alexey Ozerov, Vikrham Gowreesunker, Hiroshi Sawada, Ngoc Duong|
Chair: Paris Smaragdis (UIUC and Adobe, USA)
In this panel we plan to have an open discussion on the future of component analysis and signal separation. As these fields mature, we start to borrow ideas from other disciplines and find new application domains which keep presenting us with new and exciting research directions. We expect to hold a discussion on both technical issues, such as the use of priors, non-linear models, mixture-aware learning, evaluation in blind applications, etc; but also on higher-level concepts that help connect this community to other established areas such as control theory, neuroscience, etc. Now that we are starting to see commercial-grade algorithms come out, we also want to reevaluate the application domain of our community and discuss new areas to expand to.
Audience participation is strongly encouraged, and everyone is welcome to share their thoughts. Especially before the panel, we strongly encourage all participants of the conference to submit any points they want to discuss to the panel chair.