[T01] EEG Signal Processing and Machine Learning

Abstract

Identification of brain states such as sleep, pain, mental fatigue, and emotions, abnormalities such as autism, cognitive impairment, and ageing, diseases such as seizure, and responses to various stimuli highly applicable to brain-computer interfacing, is a popular, hot, and demanding topic in signal processing and machine learning domains. The first step to achieve this is to understand the brain function and the manifestation of brain mal functioning. The second step is to design the necessary algorithms to highlight such functions and model the changes or abnormalities, and the third step is to score, identify, or classify the brain state or severity of the problem. Tensor factorization, beamforming, connectivity/causality estimation, evaluation of brain dynamics, distributed sensor networks (with application to multi-subject or hyperscanning BCI for rehabilitation purposes), detection of event related potentials (ERPs), and other methods to detect the biomarkers will be briefly referred to. The focus will be on application of tensors, deep learning, and regularized beamforming to epileptic seizures. Detection and localization of interictal epileptiform discharges (IEDs) as well as so-called delayed responses to deep brain stimulation (DBS) from scalp, intracranial, and joint scalp-intracranial EEG recordings will be emphasized. Also, it will be demonstrated how the uncertainty in data labelling can be incorporated into the IED detection algorithm.

[T02] From Statistical to Causal Inferences for Time-Series and Tabular Data

Abstract

The vast majority of generated data are longitudinal as they are being produced by all modern systems, which are becoming more and more intelligent and autonomous. These systems require monitoring various quantities in order to infer the past as well as future changes in their inner state, and also to detect the associated internal and external events. The longitudinal data are often represented as multivariate time-series, and are predominantly stored as tabular data in Excel spreadsheets, CSV files and other types of log-files. The traditional statistical methods for processing longitudinal data rely on Gaussianity, Markovian property, and auto-regressive modeling. These methods are attractive for their low-complexity and full interpretability, however, they fail to describe more complex spatio-termporal statistical dependencies, and their assumptions often cannot be satisfied. It is only very recently when the deep learning techniques started to be effective also for processing time-series data and capture their small-scale and large-scale dependencies. The most promising strategy is to combine the traditional statistical methods with more sophisticated deep learning methods, although there are still many unsolved challenges. For instance, the unique properties of tabular data make any statistical learning to be rather difficult. Or, there are procedures for embedding categorical variables in high-dimensional vector spaces, but it ignores cases of categorical random processes. Moreover, it is often important to infer causal explanations from data, i.e., to relate causes to effects to understand why a certain event occurred, or which event is very likely to occur in a near future. The two main objectives of this tutorial are: (1) explain the key ideas in performing statistical inferences for time-series and tabular data involving traditional statistical vs. modern deep-learning methods, (2) explain how to perform causal inferences for time-series data by formulating the relevant statistical inference problems, and how to recognize that causal relationships cannot be identified from given data.

[T03] Human-Centric RF Sensing: Pose Estimation, ECG Monitoring and Self-Supervised Learning

Abstract

This tutorial contributes to the evolving domain of human-centric RF sensing technologies by integrating deep learning methodologies to address human sensing tasks. Amidst rapid advancements in RF and deep learning technologies, our work is timely and essential for advancing non-intrusive, privacy-respecting human-centric applications. This tutorial presents state-of-the-art techniques in multi-person 3D pose estimation and ECG monitoring, facilitated by large-scale, real-world RF sensing datasets. Additionally, we delve into novel signal processing and deep learning integrations, such as temporal-spatial attention and self-supervised learning for unlabeled RF data, highlighting their impact on enhancing RF sensing capabilities. Our review also introduces cutting-edge methodologies and offers comprehensive insights to spur further research in RF-based human sensing technologies. This resource aims to bridge current gaps and guide the signal processing community through the burgeoning landscape of RF sensing applications, thereby catalyzing innovation and expanding the utility of RF sensing in real-world scenarios.

[T04] Emerging Topics for Speech Synthesis: Versatility and Efficiency

Abstract

Speech synthesis is an essential technology that enables natural speech communication between humans and robots. With the advancement of deep learning technologies, the naturalness of synthetic speech has remarkably improved and the range of reproducible voice qualities is becoming increasingly diverse. This improvement comes from the development of massive speech corpora and sophisticated machine learning algorithms, such as deep generative modeling and self-supervised learning. In this tutorial, we introduce the foundational technologies that support state-of-the-art speech synthesis methods and discuss the aspects that need further research for additional advancements. We first give a lecture on the basics of speech synthesis. We then survey work on deep learning-based speech synthesis from the perspective of data, learning, and evaluation. Specifically, we focus on two aspects towards building foundation models of speech synthesis: versatility and efficiency. Finally, we discuss some possible applications of recent speech synthesis technologies such as conversational agents and the impact of these applications on human society.