Special Sessions

ICASSP 2014, along the tradition of previous ICASSPs, will house 13 high quality special sessions that complement the regular program with new or emerging topics of particular interest to the signal processing community. The aim of a special session is to provide an overview of the state-of-the-art, highlight current research directions and challenges in specific fields of signal processing.

Each Special Session will be a two-hour long oral session with six speakers. Special Sessions will be intermixed with regular sessions during the conference. Every speaker in a Special Session must submit a paper using the ICASSP 2014 paper submission guidelines, and by the regular paper submission deadline of 27 October 2013. Papers submitted to the Special Sessions will be reviewed by two reviewers selected by the organizers and one reviewer selected by the Conference Technical Program Committee.

The Organizers of the 13 accepted Special Sessions as well as the invited speakers will be required to register for the conference. Registration fees will not be waived, and ICASSP does not offer honoraria.

Sergio Barbarossa and Ananthram Swami,
ICASSP 2014 Special Session Chairs

List of Special Sessions

SS1 - Signal Processing for Big Data
Organized by: Nikolaos D Sidiropoulos

SS2 - Enhanced Radar Sensing in Harsh Environments Phenomenology, Advanced Signal and Image Processing, Demonstrations
Organized by: Yuriy V. Shkvarko

SS3 - Optimization algorithms for high dimensional signal processing
Organized by: Volkan Cevher and Mário A. T. Figueiredo

SS4 - Signal Processing for Cyber-Security and Privacy
Organized by: Holger Boche and Rafael F. Schaefer

SS5 - Seismic Signal Processing
Organized by: Leonardo Duarte, Daniela Donno, Renato R Lopes and João Romano

SS6 - Dictionary-based processing of single- and multi-channel audio
Organized by: Marco Liuni, Axel Roebel and Peter Balazs

SS7 - Joint Optimization of RF devices and Resource Allocation in Wireless Networks
Organized by: Jarmo Takala and Erik G. Larsson

SS8 - Social Nets: Learning and Optimization
Organized by: Georgios B. Giannakis and Gonzalo Mateos

SS9 - Array signal processing for radio astronomy: the SKA is the future
Organized by: Stefan J. Wijnholds, Alle Jan van der Veen and Alfonso Farina

SS10 - Dynamic Geometry Compression
Organized by: Dinei Florencio and Cha Zhang

SS11 - Signal Processing Techniques for Interference Alignment
Organized by: Constantinos B. Papadias and Tharmalingam Ratnarajah

SS12 - Deep Learning for Music
Organized by: Eric Battenberg, Erik Schmidt and Juan Bello

SS13 - Non-native Speech Processing
Organized by: Michael Johnson and Ricardo Gutierrez-Osuna

Abstracts of Special Sessions

SS1 - Signal Processing for Big Data
Organized by: Nikolaos D Sidiropoulos

On March 29, 2012, the White House announced the Big Data Research and Development Initiative to mobilize the research and development enterprise towards Big Data analytics for solving some of the world’s most pressing challenges. Similar initiatives are now well underway in Europe. Signal processing researchers can be important contributors to this endeavor, especially because Big Data analytics include multidimensional signal analytics that build upon a signals and systems fabric. Recognizing the need to engage the signal processing community in Big Data research early on, a special session is planned to bring together and highlight relevant research directions, and to help steer interest among ICASSP participants.

Unlike many web-mining researchers and professionals with a computer science background, most signal and systems engineers are trained to seek `exact' solutions, and dealing with up to modest data. Big Data is a disruptive paradigm, demanding insightful, data-centric approximation methodologies - a new mindset. Many engineers and optimization experts are familiar with distributed computation and optimization strategies, yet Big Data require robust decentralized (as opposed to distributed) processing to cope with node failures, with limited communication overhead; handling unstructured or disparate datasets; and adaptive and online solutions that can accomplish all the above with limited local memory and computational resources. These considerations naturally bring up important questions, such as:
• What are the right data-centric approximation methodologies to trade-off complexity for accuracy in massive decentralized signal and data analysis tasks?
• How can disciplined signal and data analysis algorithms be developed for big, unstructured or loosely structured data?
• How can we catalyze Big Data collaboration between scientists and engineers?
• How shall we educate engineers about Big Data?

The focus of this special session will be on signal processing theory and scalable (low complexity, storage, communication cost) algorithms for Big Data. Tools of interest include compressed sensing, computational linear algebra, optimization, dimensionality reduction, machine learning / data mining; and big data applications related to signal processing, or pertaining to very big signal / image / video data.

SS2 - Enhanced Radar Sensing in Harsh Environments Phenomenology, Advanced Signal and Image Processing, Demonstrations
Organized by: Yuriy V. Shkvarko

Due to the diversity, extremely large dimensionality, and statistical model uncertainty of the multi-scale remote sensing signal data and image sets provided by the latest-generation of the multimode active and passive radar sensing instruments operating in harsh sensing environments, there are no commonly accepted robust adaptive/collaborative intelligent signal/image processing architectures and techniques in the context of advanced computational support of different sensing missions that have rapidly created new signal/image processing challenges. Sensing in harsh environments is complicated due to the operational scenario uncertainties attributed for random perturbations of the signals in the turbulent propagation medium, possible distortions in the signal formation operators with the statistics usually unknown for the observer, multiplicative signal-dependent speckle noise, uncontrolled antenna vibration and carrier trajectory deviations in the case of SAR. The dominant idea of this ERS’2014 Special Session is to bring together the researches from the international radar signal/image processing and remote sensing communities to present and discuss the state-of-the-art developments of the new emerging trends and advances in multimode sensor signal and image processing with special emphasis on the radar sensing in harsh operational environments. ERS’2014 is devised to encompass the following emerging areas of the signal/image processing for multimode radar/SAR sensing/imaging instruments with emphasis on phenomenology and processing specifics peculiar for harsh/uncertain sensing environments:
• Modeling of Harsh Sensing Environments (Speckle, Data Degradations, Model Uncertainties);
• Experiment Design Frameworks for High Performance Radar Sensing;
• Parallel and Distributed Algorithm Design for Feature Enhanced Radar Sensing;
• Robust Adaptive Techniques for Feature Enhanced Radar Imaging in Harsh Sensing Environments;
• Cooperative Data Fusion Concepts and Techniques;
• Feature Extraction, Signature Fields Mapping and Image Understanding;
• Hardware/Software Co-Design for (Near) Real Time Feature Enhanced Sensing.

SS3 - Optimization algorithms for high dimensional signal processing
Organized by: Volkan Cevher and Mário A. T. Figueiredo

Recent advances in optimization are pointing the way to a new breed of algorithms for difficult signal processing problems that are defined in very high dimensions. This session aims at providing a snapshot of some emerging techniques, covering some of the basic theory that governs the underlying procedures, as well as illustrating new algorithms to tackle high dimensional optimization problems efficiently. The participants of the session are well-known experts in the interface between optimization and signal processing, and the algorithms to be presented will all be numerically demonstrated in signal processing applications for broader interest to the ICASSP community.

SS4 - Signal Processing for Cyber-Security and Privacy
Organized by: Holger Boche and Rafael F. Schaefer

In today’s communication systems there is an architectural separation between data-encryption and error-correction. The encryption module is based on cryptographic principles and abstracts out the underlying communication channel as an ideal bit-pipe. The error-correction module is typically implemented at the physical layer. It adds redundancy into the source bits in order to combat channel impairments or multiuser interference and transforms the noisy communication channel into a reliable bit-pipe. While such a separation based architecture has long been an obvious solution in most systems, a number of applications have emerged in recent years where encryption mechanisms must be aware of the noise structure in the underlying channel and likewise the error-correction and data-compression methods must be aware of the associated secrecy constraints required by the application. Such joint approaches can be studied by developing new mathematical models of communication systems that impose both reliability constraints as well as secrecy constraints. It is notable that this approach leads to guaranteeing information security are irrespective of the computational power the adversary and is a fundamental departure from current computation-based cryptographic solutions.

The study of even very simple models that simultaneously introduce both secrecy and reliability constraints can lead to surprising insights that can fundamentally change the way we approach the design of security mechanisms. This special session addresses recent developments in the broad area of cyber‐security and privacy for wireless networks and biometrics such as privacy aspects in biometric systems, secrecy in wireless systems, and secret key generation.

SS5 - Seismic Signal Processing
Organized by: Leonardo Duarte, Daniela Donno, Renato R Lopes and João Romano

A fundamental problem in geophysics is to estimate the properties of the Earth’s subsurface based on measurements acquired by sensors located over the area to be analyzed. Among the different methods to accomplish this task, seismic reflection is the most widespread and has been intensively applied for hydrocarbon exploration. The characterization of the subsoil using seismic reflection techniques is conducted by recording the wave field that is originated from the interaction between the environment under analysis and a seismic wave generated by controlled active sources (e.g. a dynamite explosion in land acquisition). Signal processing (SP) plays a fundamental role in seismic reflection. Indeed, in order to extract relevant information from seismic data, one has to perform tasks such as filtering, deconvolution, and signal separation. Originally, there was a close interaction between the signal processing and geophysics communities – for instance, important achievements in deconvolution and the wavelet transform were obtained in the context of seismic data. Nowadays, however, this interaction has been partially lost – as a consequence, geophysicists are not aware of the most recent SP methods, and, on other hand, the SP community is drawing weak attention to this interesting application. Given this panorama, the main goals of this special session are to shed some light on the research in seismic signal processing, and to broaden and reinforce collaboration between the signal processing and the geophysics research communities. With this goal in mind, the session comprises works on important theoretical and practical topics that arise in seismic signal processing.

SS6 - Dictionary-based processing of single- and multi-channel audio
Organized by: Marco Liuni, Axel Roebel and Peter Balazs

Representations of audio signals are often obtained with decompositions in terms of atomic functions with desired features, that ease the identification and elaborations of fundamental characteristics. Such atomic functions form sets, called dictionaries, that span the whole signal space or a given subspace, with variable redundancy. Since the main goal of a representation is to increase the readability of the observed phenomenon, the concept of sparsity plays a key role in the choice of a dictionary: it concerns the efficiency of a given representation, in relation with the aimed information. This concept leads to two complementary fundamental issues, determining two classes of applications: one is, given a privileged analysis domain, to define a sparsity measure and the related optimal representations; the second is, given an aimed notion of sparsity, to define an analysis domain where the signal's representation is highly sparse. This special session covers two related research directions, namely frame theory and dictionary learning, aiming to create connections between the communities working on these different problems.

SS7 - Joint Optimization of RF devices and Resource Allocation in Wireless Networks
Organized by: Jarmo Takala and Erik G. Larsson

Energy consumption is a critical aspect in green, sustainable technologies for cognitive radio terminals, which can connect to networks operating on different frequency bands with a variety of air interfaces. Flexibility and configurability at RF, baseband, and MAC layers, with cross-layer modeling and control, are important when exploiting the efficiency potential of spectrum sharing. Recently receiver configuration based on programmable paradigms has received attention but practical solutions for the radio frequency/radio components are still lacking, while programmable baseband computation and related design chains have been developed. The software based adaptive configuration of radio frequency chains is still in its infancy, but it is a key ingredient for frequency agile radios needed for cognitive devices and flexible RF spectrum use.

A major gap is the lack of models and comprehensive understanding of realistic configurable RF chains and radio modules. The ability to control unwanted emissions is a major implementation concern. Especially in battery-powered devices, high power-efficiency is very critical, which leads to driving the transmit power amplifier closer to its saturation region. This results in nonlinear intermodulation distortion leading to increased unwanted emissions and can violate the given spectrum emission limits. Digital predistortion is currently the de-facto solution in e.g. base station equipment in mobile cellular radio networks. However, it has limited use in the mobile terminals due to computational requirements, and power efficient amplification has been obtained through envelope elimination and restoration and/or envelope tracking power amplifier configurations. This has challenges with increasing instantaneous bandwidths and carrier aggregation type scenarios evolving currently e.g. in 3GPP LTE-Advanced mobile cellular radio.

The target of the special session is to explore signal processing algorithm challenges and solutions as well as the required architectures and implementations to enable cost and power efficient realization of cross layer optimization of multiuser wideband radios and systems.

SS8 - Social Nets: Learning and Optimization
Organized by: Georgios B. Giannakis and Gonzalo Mateos

Over the past decade there has been a growing interest on the complex connectedness of modern society. Noteworthy examples of this connectedness phenomenon include the explosive growth of the Internet, the popularized “six degrees of separation” characterizing social ties such as friendship, and the ability of news and epidemics to spread with unprecedented speed and intensity. As a result, the study of networks has increased dramatically with multidisciplinary efforts from researchers in physics, computer science, engineering, statistics, economics, and the biobehavioral sciences, just to name a few. As modern (social) networks grow in size and importance, while they become more complex and heterogeneous, there is an urgent need to advance a holistic theory of networks that in turn offers ample opportunities for signal processing (SP) research. In this context, the proposed special session aims at (a) delineating the analytical background and the relevance of SP tools to the modeling and analysis of social networks; and (b) introducing the SP community to the major challenges in understanding the collective behavior of emerging social-computational systems, as well as engineering “socially-intelligent” complex networks.

SS9 - Array signal processing for radio astronomy: the SKA is the future
Organized by: Stefan J. Wijnholds, Alle Jan van der Veen and Alfonso Farina

Array signal processing techniques are obtaining increasing attention in radio astronomy. This interest is driven by a new generation of radio telescopes based on phased array technology, in particular the Square Kilometre Array (SKA), an imaging array consisting of thousands of elements. Achieving the envisaged 74 dB imaging dynamic range requires high quality array signal processing techniques, that need to be computationally efficient to cope with the data deluge of ~10 TB/s generated by the SKA.

This special session provides the signal processing community with an overview of the challenges faced by the SKA, which is pushing the limits of imaging array performance and computational feasibility, as well as the routes being explored to overcome them. The presentations in this session cover high quality image reconstruction techniques, efficient array calibration techniques and RFI (Radio Frequency Interference) mitigation methods. In many cases, the strengths of these methods are demonstrated using actual data from current radio telescopes or technology demonstrators for the SKA. The session will be concluded by a talk by IBM Research, one of the industrial partners to the SKA project, on holistic design of signal processing systems for big data projects. These presentations give a nice overview of the wide variety of recent developments. These are not only relevant to the radio astronomy community, but have clear ties with radar applications, medical imaging applications and processing of Big Data. With this session, we hope to stimulate cross-fertilization between these fields.

SS10 - Dynamic Geometry Compression
Organized by: Dinei Florencio and Cha Zhang

With the advent of consumer grade depth cameras, there has been a surge of interest and availability of depth information. While in the past geometric information originated mostly from computer graphics, it is now more and more common to have geometry information originated from sensing the environment, e.g., from Kinect, stereo cameras, lidar, etc. While currently there is much work on compression of depth maps within the signal processing community, there could be significant advantages to compress the geometry directly. There has been some work in the computer graphics community targeted at mesh compression, but we believe that many of the techniques typically used in signal processing have the potential to greatly increase the compression efficiency. To encourage the research community to lend its expertise to this important problem, the Multimedia Technical Committe, with the support of the Signal Processing Society, chose one specific topic within this area (namely, compression of dynamic consistent meshes) and established a Dynamic Geometry Compression Competition , potentially the first in an annual series. This competition is loosely associated with the proposed special session, in the sense that it is within the scope of the special session, but the special session targets a more general problem, including direct compression of other types of dynamic geometry. We believe the special session will be an important event to increase the attention of the SPS community to this new research area, which we believe will become fundamentally more important in the next several years. Together with the Competition, as well as another special session planned for ICIP, we believe this will provide a foundation for significant further research in this topic, which we expect to flourish in the next several years.

SS11 - Signal Processing Techniques for Interference Alignment
Organized by: Constantinos B. Papadias and Tharmalingam Ratnarajah

Interference alignment (IA) is one of the most recent approaches for the pre- and post- processing of wireless signals that co-exist with each other in wireless communication applications. The approach is premised on the so-called “interference channel” scenario, wherein a number of (e.g. K) transmitter / receiver nodes (pairs) co-exist, and hence mutually interfere in their attempt to communicate with each other. The K communication transmitters are assumed to share information about their channels (both intended and unintended) but not their information data streams. Assuming Rayleigh i.i.d. channels, the IA approach promises the attainment of a sum rate that corresponds to half the degrees of freedom of the interference channel. Despite the high promise, a number of important challenges regarding the use of IA-type techniques in wireless interference channels have yet to be addressed. This special session brings forth a collection of papers that address precisely some of these open issues, with emphasis on the signal processing aspects of IA. These include, on the theoretical side: the development and analysis of adaptive coding / modulation techniques for IA with imperfect CSI; the development of IA-type techniques that have good sum rate behavior in the low-to-moderate SNR regime; and the performance analysis of IA techniques in the presence of uncoordinated interference, such as in heterogeneous network environments. On the more practical side, the performance evaluation of IA techniques when used within wireless standards (such as the IEEE802.11ac); the evaluation of the impact of imperfect Channel State Information in over-the-air transmissions; and, finally, the implementation of IA techniques on terminals with compact antenna arrays that use mixed digital / analog processing will also be addressed. Overall, the session will provide a number of novel contributions, both in terms of technique development and performance evaluation, that we feel advance the state-of-the-art in signal processing for the interference channel

SS12 - Deep Learning for Music
Organized by: Eric Battenberg, Erik Schmidt and Juan Bello

Interest in deep learning has exploded in recent years as it has transitioned from a once controversial line of research into the de facto standard in many areas of machine perception. While such methods are now state of the art in image and speech recognition, deep architectures and feature learning have only recently become areas of intense interest in the fields of music signal processing and music information retrieval. The application of deep learning techniques to music audio present unique challenges that set it apart from ongoing efforts in other fields. Music signals are time series where events are organized in musical - rather than real - time on a grid that constantly changes as a function of rhythm and expression. They typically combine multiple voices highly synchronized in time and overlapping in frequency, according to complex arrangements that include short and long-term temporal dependencies. The universe of possible arrangements is nearly unlimited, with wide variations according to musical tradition, style, composer and interpretation. This complexity, unfettered by any standard computational approaches to the many tasks in music signal processing, presents a unique set of problems that are well-suited to the high levels of abstraction afforded by the perceptually and biologically motivated aspects of deep learning. The proposed session attempts to provide a much-needed forum for an in-depth discussion of these challenges, as well as for the presentation of novel experiments and domain specific adaptations to deep learning in music.

SS13 - Non-native Speech Processing
Organized by: Michael Johnson and Ricardo Gutierrez-Osuna

Non-native speech presents unique challenges as it can deviate substantially from the acoustic and prosodic norms of a language. These challenges have led to significant work in speech processing in such areas as pronunciation analysis, model adaptation, and accent and dialect classification, and their application to computer-assisted pronunciation training. This work crosses several different research domains, including acoustics and speech processing, language modeling, and articulatory phonetics and linguistics. This special session focuses on bringing researchers together across these diverse areas together to promote cross-fertilization of ideas and energize further research in non-native speech processing. The session will cover multiple aspects of the problem, including pronunciation modeling and scoring, ASR model adaptation, accent conversion/modification, perceptual evaluation of accents and speaker similarity, and computer assisted pronunciation training.