Search
mlsplogo MLSP2016
IEEE International Workshop on
Machine Learning for Signal Processing

September 13-16, 2016  Vietri sul Mare, Salerno, Italy

Special Sessions

Bayesian Machine Learning for Neural Signal Processing
Bayesian statistics is a major branch of the field of statistics that uses probability to express the degrees of belief towards the real world, and integrates (based on the Bayes rule) the evidence in the data and a priori knowledge to make inferences. Compared with the conventional statistical methods, Bayesian statistics deals with multiple levels (e.g., models and parameters) of uncertainties in a unified framework and yields probabilistic solutions that are more useful for prediction than point estimates. Bayesian machine learning is an emerging research area that integrates Bayesian inference, optimization, Monte Carlo sampling, and machine learning techniques for data analysis. With the advent of fast and powerful computing resources and development of efficient optimization methods, the past two decades have witnessed the vast use and development of advanced Bayesian methods for analysis of neural and physiological signals, such as EEG, MEG, fMRI, calcium imaging, neuronal population spike trains. It is our belief that Bayesian machine learning will play an increasingly important role in signal processing applications.

This special session attempts to welcome and foster novel ideas, new applications, and outside-of-the-box thinking along this line of research. Active researchers are invited to make presentations on their recent innovative work and to share their thinking with the Machine Learning for Signal Processing (MLSP) audiences who have special interests in Bayesian statistics, signal processing and machine learning as applied to the aforementioned research fields. The central theme of this special session focuses on “Bayesian machine learning” and its application to “neural signal processing”. The speakers of this special session are selected to cover several representative research areas. The topics of this special session will include but are not limited to: state-space model, approximate Bayesian inference, brain imaging, change-point detection, neural decoding and brain-machine interface (BMI).

Organized by

Dr. Zhe (Sage) Chen, NYU, USA
Dr. Luca Citi Univ.Essex, UK
Advances in Gaussian Processes for Machine Learning and Signal Processing
Gaussian processes provide a powerful and expressive learning framework for machine learning and signal processing. In the last decade, many flavors of GPs have been developed, and the framework has been extensively adopted by practitioners to tackle many real problems. Signal processing has benefited as well, and now we find novel developments and applications in adaptive filtering, automatic control, bioengineering, marketing and communications, just to name a few.

The framework allows us to deal with the particular characteristics of the signal, the noise, include prior knowledge (both empirical and mechanistic) about the problem, and simultaneously find structure and discover patterns in the problem at hand.

Recently, deep GPs have also reported competing results with deep neural networks, and large scale implementations of current GPs have made it possible to tackle a plethora of real-world problems. This special session aims to provide an overview of recent advances in Gaussian processes, discussing challenges and opportunities in the field with special focus on signal processing applications.

Session scope and main topics

We aim to cover a vast field of theoretical and applied problems within the GPs. Theoretical developments are expected in the areas of structure/adaptive/on-line learning, causal inference, kernel/covariance design, large scale GPs, uncertainty estimation and propagation, heteroscedastic noise, multi-output/multi-task learning, and probabilistic numerics. Applications including, but not limited to: signal/image processing, marketing, neuroscience, bioengineering bioinformatics, geostatistics, communications, spatial analysis, time-series analysis, and econometrics.

Organized by

David Luengo, Universidad Politécnica de Madrid, Spain
Gustau Camps-Valls, University of Valencia, Spain
Simo Särkkä, Aalto University, Finland
Computational Methods for Audio Analysis
Nowadays, computational methods are largely used to face complex modelling, prediction, and recognition tasks in different research fields. One of these fields is represented by the analysis of audio signals, which finds application in communication, entertainment, security, forensics and health to name a few. In fact, with the rapid growth and omnipresence of digitized multimedia data, the processing, analysis, and understanding of such data by means of automated methods has become a central issue in machine learning and associated areas of research.

It is indeed of great interest for the scientific community to understand the effectiveness of novel computational methods for audio analysis, which has recently gained a particular attention, supported by many special issues, special sessions at conferences, and challenges proposed on such a topic. Also, an IEEE Task Force on Computational Audio Processing was recently launched with its effort devoted to computational intelligence techniques applied to audio processing. For these reasons, we believe that a special session on computational methods for audio analysis can be of great interest for the MLSP community, which might present the most recent and exciting advances on this topic.

Scope and main topics

The typical methodology adopted in this environment can be applied to different kinds of audio signals, from music to speech, from sound to acoustic data. In the “music” case study, music information retrieval is the major topic of interest, with many diverse sub-topics therein; for “speech”, we can immediately refer to speech/speaker recognition, but also the many diverse topics intimately related to the computational analysis of speech signals (affective computing and language processing, just to name a few); for “sound”, event detection/identification are recognition are the major aims; in the case of “acoustic”, acoustic fingerprint/signature and acoustic monitoring have lately seen a big interest in the field. The use of computational methods may also allow to provide a characterization of an audio stream by directly analyzing raw data, which seems to be a popular choice recently. Moreover, cross-domain approaches have been recently investigated to exploit the information contained in diverse kinds of environmental audio signals.

The aim of this session is therefore to provide the most recent advancements on the application of novel computational methods to a wide range of audio analysis tasks.

Potential topics include, but are not limited to:

  • Machine Learning Algorithms for Speech and Audio Analysis
  • Cross-domain Audio Analysis
  • Deep Learning for Audio Applications
  • Audio Source Separation and Localization
  • Audio-based Security Systems and Surveillance
  • Reinforcement Learning for Audio
  • Music Information Retrieval
  • Sound and Speech Recognition for Content Analysis
  • Speech and Audio Forensics
  • Computational Methods for Wireless Acoustic Sensor Networks
  • Speech and Speaker Analysis and Classification
  • Sound and Novelty Detection and Recognition
  • Machine Learning for Speech Enhancement
  • Intelligent Audio Interfaces

Organized by

Michele Scarpiniti, Sapienza University of Rome, Italy
Danilo Comminiello, Sapienza University of Rome, Italy
Tuomas Virtanen, Tampere University of Technology, Finland
Mark Plumbley, University of Surrey, UK
Stefano Squartin, Università Politecnica delle Marche, Italy
Paris Smaragdis, University of Illinois at Urbana-Champaign, USA
Machine Learning Methods for Big Data Privacy Protection
An emerging direction of research in big data mining and data analysis for non-military purposes has recently been the development of methods for collecting information while at the same time respecting user privacy. For instance, a school district may want to release an aggregate information, such as the number of A students in the district, but it needs to do so without having access to the individual info, i.e., whether a particular student is an A student or not. As another example, a face recognition system may be required to classify a person as a man or a woman without being able to identify the person’s name, ethnicity, or occupation.

Machine learning offers the essential array of tools for data mining. The increasing need for privacy protection introduces novel ML problems and demands for new approaches that are quickly getting the attention of the research community. In the privacy protection context, the question is how to perform machine learning tasks, such as classification and clustering, with masked data, i.e. without having complete access to the precise user data records.

Sscope and main topics

In the privacy preserving scenario new feature engineering methods need to be employed in order to transform the data in such a way that privacy is maintained while the utility of the data is maximized. Linear and nonlinear (kernel and/or deep learning) methods are explored for data representation and compression so that the identity of the original data is protected and classification performance suffers minimum reduction. Also Privacy-Enhanced tools and applications (e.g. PE face- or voice-recognition) are very important areas attracting the attention of the research community.

The aim of this special session is to bring together scientists presenting the latest ideas and concepts pertaining to this emerging field of privacy-preserving machine learning.

Topics include:

  • Privacy Preserving Data Analysis
  • Privacy Preserving Data Mining
  • Kernel methods for Privacy Preserving
  • Deep Learning for Privacy Preserving
  • Privacy enhanced Face Recognition
  • Compressive Privacy
  • Privacy-enhanced applications

Organized by Konstantinos Diamantaras, TEI of Thessaloniki, Greece

Powered by CONWIZ, © Copyright 2015 | Privacy Policy and Terms of Use | Page maintained by Jan Larsen