MLSP 2020

IEEE International Workshop on

September 21–24, 2020 Aalto University, Espoo, Finland (virtual conference)

Accepted Special Sessions

Disclaimer: This does not guarantee inclusion of the special session to the program, if not sufficient papers are accepted. Special session submissions go through the same peer-review process as other conference submissions. Selection of reviewers will be done in cooperation with special session organisers.

Acoustic Signal Processing Using Machine Learning

Emanuël A.P. Habets and Sharon Gannot

Machine learning, and deep learning in particular, has revolutionized research in acoustic signal processing, achieving unprecedented performance in various tasks. Yet, it is often criticized for being uninterpretable. In recent years, there is a growing tendency in the machine learning community to develop algorithms that can be interpreted. The aim of this special session is to present recent advances in the field that are in the crossroad between the learning and the acoustic signal processing communities. Various applications are addressed, namely event detection with constraints on the number of microphones, acoustic scene classification, noise reduction using microphone arrays, joint audio and EEG processing for hearing aids applications, speech source separation and speaker localization. These papers, presented by leading researchers in the field from all over the world, apply a combination of “classical” array processing algorithms, e.g. multichannel Wiener filters and MVDR beamformers, and various learning techniques, e.g. deep networks, generative adversarial networks (GANs), variational Bayesian inference, and domain adaptation.

Robust and High-dimensional Statistical Learning for Signal Processing Applications

Esa Ollila and Frédéric Pascal

High-dimensional data has become ubiquitous in the Big Data era. At the same time, challenges due to high-dimensionality are well recognized by data scientists today. For example, data dimensionality often exceeds the sample size, a regime opposite to conventional statistical settings. Another challenge is that outliers are increasingly hard to detect from high-dimensional data. Hence statistical robustness (i.e., insensitivity to outliers and model imperfections) is an increasingly important criterion in the design of high-dimensional statistical learning procedures. This special session proposal is targeted on tackling these challenges and provides cutting-edge international research in the fields of high-dimensional statistics and robust statistical learning for signal processing applications from international experts in the field.

Deep Learning for Inverse Problems

Marco Prato and Samuli Siltanen

Inverse problems arise from needs to interpret indirect measurements. Sometimes the desired information is buried in the data in an unstable way, making the recovery task very sensitive to measurement noise and modelling errors. There is an active branch of mathematics devoted to overcoming ill-posedness and designing signal processing for robust extraction of information. The inverse problems setup is natural for machine learning as training sets are often available in the form of pairs of targets and data. However, ill-posedness causes difficulties for learning the target from the data, while learning the data from the target is typically straightforward. This special session is built around taking the best of both worlds. Combining inverse problems mathematics with deep learning techniques may lead to more accurate reconstruction results than either of the methodologies alone, where the mathematical part can additionally help to provide interpretability for the results.

Advances in Machine Learning and Signal Processing for Financial Technology

Che Lin, Anastasios Tefas, and Yeong-Luh Ueng

Financial technology (Fintech), a broad category that refers to the innovative use of technology in the design and delivery of financial services and products, has revolutionized the financial industry or, even more broadly, the service industry. With the popularization of mobile payments and other growing technologies for payment/shopping, a large amount of data is generated every day, and these data are available to companies and banks. However, the challenges lie in the retrieval of business insights and actionable strategies from this enormous amount of data. FinTech applications are a highly promising area for the use of machine learning and signal processing algorithms as these are are perfect technical vehicles for the extraction of business intelligence from the vast amount of customer behavioural data for FinTech applications.

Machine Leaning and Signal Processing for Autonomous Systems

Nazre Batool, Letizia Marchegiani, Sahar Abbaspour, and Jesper Rindom Jensen

Autonomous systems are one of the most important current trends in research and development: robots, driverless vehicles, unmanned aerial vehicles are becoming a more integral part of our society every day. Analysis and interpretation of the information from sensors plays a central role in autonomous systems. The sensor data obtained can contain irrelevant information and the amount of information collected from sensors can impair the process of utilizing such information. Therefore, developing techniques for processing sensor data is crucial and this is where machine learning and signal processing as core techniques can help with efficient and reliable data analysis. Autonomous system applications present a different set of constraints and challenges for machine learning and signal processing methods: availability of limited resources (both computational and energetic), necessity to deal in real-time with highly multi-modal data which might contain orthogonal information that needs to be fused and parsed in a sensible way, necessity to deal with uncertainty in a way that might prioritise safety over other performance measures, necessity to cope with real-world scenarios, which might be affected by unpredictable sources of noise and settings’ changes. These constraints can present an opportunity to encourage the development of machine learning and signal processing techniques towards their applicability to new scenarios and contexts, enhancing their robustness and reliability.

Special Session Call for Proposals

MLSP is seeking original, high quality proposals for Special Sessions, to be included in the technical program along with the regular track. Special Sessions are expected to address research in focused, emerging, or interdisciplinary areas of particular interest, not covered already by traditional MLSP sessions.

The proposals for Special Sessions should include the following information:

  1. Title of the special session
  2. Motivation for the proposed session, stating its relevance for MLSP
  3. Description of the session scope and main topics to be covered
  4. List of expected papers: sessions must have 6 papers
  5. Short bio and contact information of the organizers

The proposals will be evaluated based on their quality, timeliness, novelty, relevance to MLSP, potential to open new perspectives and/or bring together confluent areas, and further develop the field. Each Special Session will include a selection of six papers. The session organizers will be responsible for soliciting submissions.

Paper submissions to the Special Sessions should adhere to the same style as the regular papers, and shall follow the same time line and review process. The Special Session Chair will collaborate with the organizers in the assignment of reviewers to the submitted papers, the quality of which shall meet the same standards as regular papers. If fewer than the needed number of papers are selected, the Special Session will be canceled, and the accepted papers will be scheduled in regular sessions.

Proposals should be sent via email to:

with “MLSP2020 Special Session Submission” in the subject line.

Proposal deadline March 19, 2020.