Search
mlsplogo MLSP2017
IEEE International Workshop on
Machine Learning for Signal Processing

September 25-28, 2017  Roppongi, Tokyo, Japan

Special Sessions

New Extensions and Applications of Non-Negative Audio Modeling
Non-negative matrix factorization (NMF) has attracted a lot of attention after being proposed as a powerful approach for audio source separation and is still one of the de facto standards for many audio processing problems. While a black- box machine learning approach based on deep neural networks has recently proved powerful for various supervised audio processing tasks, the non-negative audio modeling approach still remains attractive for unsupervised blind signal processing problems. In addition, since it is a generative approach, it can be convenient in semi-supervised scenarios where only a limited amount of training data is available. Over the years, many variants of NMF have been proposed in the field of audio processing with the aim of adapting the regular NMF model to audio-specific problems. These include phase-aware extensions, multichannel extensions, convolutive extensions, combinations with the source-filter model, the introduction of new divergence measures, otpimization methods, models with deep architectures and combinations with neural networks. The aim of this special session is to bring together new ideas on extensions of the non-negative audio modeling and their applications.

This special session focuses on various extensions of the NMF models designed to solve audio-specific problems, including monaural and multi-channel audio source separation, voice conversion, audio event detection, music transcription and speech analysis.

Organized by

Hirokazu Kameoka, Communication Science Laboratories NTT, Japan
Alexey Ozerov, Technicolor Research and Innovation, France
Cédric Févotte, Institut de Recherche en Informatique de Toulouse (IRIT), France
Paris Smaragdis, University of Illinois at Urbana-Champaign, USA
Machine Learning for Computational Imaging
There has been growing interest in recent years in data-driven approaches to solving various problems in computational imaging. Data-driven signal modeling techniques such as dictionary learning, low-rank models, etc., have been recently demonstrated to provide promising performance in image reconstruction problems in magnetic resonance imaging, computed tomography, computational photography, microscopy, etc., especially in scenarios involving limited (e.g., compressed sensing) or corrupted data. There has also been very recent interest in exploiting deep learning methods for various imaging applications including inverse problems.

This special session will bring together papers on advanced machine learning approaches to various imaging problems, with a particular focus on inverse problems aka image reconstruction problems. The approaches presented in the session will exploit sophisticated signal models and image reconstruction models based on low-rank, sparsity, tensor and manifold structures, graphical, and convolutional models, etc., together with modern large-scale optimization techniques for learning and incorporating the models in imaging applications. Diverse imaging modalities and systems will be covered, with a focus on computational approaches.

The session would bring together leading researchers and field experts in computational imaging, biomedical imaging, machine learning, image processing, compressed sensing, and optimization, and facilitate substantive and cross-disciplinary discussions on cutting-edge data-driven imaging and image reconstruction approaches.

Main Focus Topics of the Session:

  • Computational imaging, Inverse problems, Image processing, Image reconstruction, Magnetic resonance imaging, Computed tomography, Dynamic imaging, Optical imaging, Computational photography, Compressed sensing
  • Sparse signal models, Dictionary Learning, Low-rank models, Tensor models, Manifold models, Convolutional models, Deep learning, Other machine learning techniques
  • Optimization methods, and Performance guarantees for data-driven techniques

Organized by

Saiprasad Ravishankar, University of Michigan
Brendt Wohlberg, Los Alamos National Lab
Stamatios Lefkimmiatis, Skoltech
Jong Chul Ye, KAIST, South Korea
Deep Learning for Speech Enhancement
Learning-based approaches for speech enhancement have gathered interest recently. Enhancement refers to dereverberation and/or denoising of speech signals to improve SNR, perceptual quality, intelligibility or the accuracy of automatic speech recognition. Conventional methods use statistical models for enhancement however their model parameters are not pre-trained from data, instead they are either estimated from a given speech utterance or fixed based on expert information. Recent learning-based methods which involve NMF, statistical models such as GMM-HMM or deep learning methods rely on training data of varying sorts. Deep learning based methods have been shown to be superior to other alternatives in recent literature. We would like to propose this session to further explore deep learning for speech enhancement and we will seek to accept papers that would propel the field further. It would also help inform MLSP community to the recent developments in the field.

Scope: Deep learning, multi-task learning, ensemble learning, speech enhancement, speech separation, dereverberation, denoising, single channel, multiple channels, feature design, neural network architectures, loss functions, joint enhancement and recognition networks, beamforming networks, model generalization, post-processing, integration of deep learning and other approaches, customized applications

Organized by

Hakan Erdogan, Microsoft
Jun Du, University of Science and Technology of China, China
Powered by CONWIZ, © Copyright 2017 | Privacy Policy and Terms of Use | Page maintained by Jan Larsen