Single-channel source separation for radio-frequency (RF) systems is a challenging problem relevant to key applications, including wireless communications, radar, and spectrum monitoring. This thesis addresses the challenge by focusing on data-driven approaches for source separation, leveraging datasets of sample realizations when source models are not explicitly provided. To this end, deep learning techniques are employed as function approximations for source separation, with models trained using available data. Two problem abstractions are studied as benchmarks for our proposed deep-learning approaches. Through a simplified problem involving Orthogonal Frequency Division Multiplexing (OFDM), we reveal the limitations of existing deep learning solutions and suggest modifications that account for the signal modality for improved performance. Further, we study the impact of time shifts on the formulation of an optimal estimator for cyclostationary Gaussian time series, serving as a performance lower bound for evaluating data-driven methods. ...

The recent success of deep learning is conditioned on the availability of large annotated datasets for supervised learning. Data annotation, however, is a laborious and a time-consuming task. When a model fully trained on an annotated source domain is applied to a target domain with different data distribution, a greatly diminished generalization performance can be observed due to domain shift. Unsupervised Domain Adaptation (UDA) aims to mitigate the impact of domain shift when the target domain is unannotated. The majority of UDA algorithms assume joint access between source and target data, which may violate data privacy restrictions in many real world applications. In this thesis I propose source-free UDA approaches that are well suited for scenarios when source and target data are only accessible sequentially. I show that across several application domains, for the adaptation process to be successful it ...

Technology advancements profoundly impact numerous aspects of life, including how we communicate and interact. For instance, hearing aids enable hearing-impaired or elderly people to participate comfortably in daily conversations; telecommunications equipment lifts distance constraints, enabling people to communicate remotely; smart machines are developed to interact with humans by understanding and responding to their instructions. These applications involve speech-based interaction not only between humans but also between humans and machines. However, the microphones mounted on these technical devices can capture both target speech and interfering sounds, posing challenges to the reliability of speech communication in noisy environments. For example, distorted speech signals may reduce communication fluency among participants during teleconferencing. Additionally, noise interference can negatively affect the speech recognition and understanding modules of a voice-controlled machine. This calls for speech enhancement algorithms to extract clean speech and suppress undesired interfering signals, ...

The unprecedented processing demand, posed by the explosion of big data, challenges researchers to design efficient and adaptive machine learning algorithms that do not require persistent retraining and avoid learning redundant information. Inspired from learning techniques of intelligent biological agents, identifying transferable knowledge across learning problems has been a significant research focus to improve machine learning algorithms. In this thesis, we address the challenges of knowledge transfer through embedding spaces that capture and store hierarchical knowledge. In the first part of the thesis, we focus on the problem of cross-domain knowledge transfer. We first address zero-shot image classification, where the goal is to identify images from unseen classes using semantic descriptions of these classes. We train two coupled dictionaries which align visual and semantic domains via an intermediate embedding space. We then extend this idea by training deep networks that ...

It takes more time to think of a silent scene, action or event than finding one that emanates sound. Not only speaking or playing music but almost everything that happens is accompanied with or results in one or more sounds mixed together. This makes acoustic event detection (AED) one of the most researched topics in audio signal processing nowadays and it will probably not see a decline anywhere in the near future. This is due to the thirst for understanding and digitally abstracting more and more events in life via the enormous amount of recorded audio through thousands of applications in our daily routine. But it is also a result of two intrinsic properties of audio: it doesn’t need a direct sight to be perceived and is less intrusive to record when compared to image or video. Many applications such ...

This thesis contributes to motion characterization through inertial and physiological signals captured by wearable devices and analyzed using signal processing and deep learning techniques. This research leverages the possibilities of motion analysis for three main applications: to know what physical activity a person is performing (Human Activity Recognition), to identify who is performing that motion (user identification) or know how the movement is being performed (motor anomaly detection). Most previous research has addressed human motion modeling using invasive sensors in contact with the user or intrusive sensors that modify the user’s behavior while performing an action (cameras or microphones). In this sense, wearable devices such as smartphones and smartwatches can collect motion signals from users during their daily lives in a less invasive or intrusive way. Recently, there has been an exponential increase in research focused on inertial-signal processing to ...

Auscultation with a stethoscope enables us to recognize pathological changes of the lung. It is a fast and inexpensive diagnosis method. However, it has several disadvantages: subjectiveness, i.e. the lung sound evaluation depends on the experience of physicians, can not provide continuous monitoring and a trained expert is required. Furthermore, the characteristics of the lung sounds are in the low frequency range, where the human hearing has limited sensitivity and is susceptible to noise artifacts. Exploiting the advances in digital recording devices, signal processing and machine learning, computational methods for the analysis of lung sounds have been a successful and effective approach. Computational lung sound analysis is beneficial for computer-supported diagnosis, digital storage and monitoring in critical care. Beside computational lung sound analysis, the recognition of acoustic contextual information is important in various applications. The motivation for recent research on ...

Music source separation is the task of separating musical sources from an audio mixture. It has various direct applications including automatic karaoke generation, enhancing musical recordings, and 3D-audio upmixing; but also has implications for other downstream music information retrieval tasks such as multi-instrument transcription. However, the majority of research has focused on fixed stem separation of vocals, drums, and bass stems. While such models have highlighted capabilities of source separation using deep learning, their implications are limited to very few use cases. Such models are unable to separate most other instruments due to insufficient training data. Moreover, class-based separation inherently limits the applicability of such models to be unable to separate monotimbral mixtures. This thesis focuses on separating musical sources without requiring timbral distinction among the sources. Preliminary attempts focus on the separation of vocal harmonies from choral ensembles using ...

The explosion of Deep Learning (DL) added a boost to the already rapidly developing field of Computer Vision to such a point that vision-based tasks are now parts of our everyday lives. Applications such as image classification, photo stylization, or face recognition are nowadays pervasive, as evidenced by the advent of modern systems trivially integrated into mobile applications. In this thesis, we investigated and enhanced the visual counting task, which automatically estimates the number of objects in still images or video frames. Recently, due to the growing interest in it, several Convolutional Neural Network (CNN)-based solutions have been suggested by the scientific community. These artificial neural networks, inspired by the organization of the animal visual cortex, provide a way to automatically learn effective representations from raw visual data and can be successfully employed to address typical challenges characterizing this task, ...

In the audio research field, the majority of machine learning systems focus on recognizing a limited number of sound events. However, when a machine interacts with real data, it must be able to handle much more varied and complex situations. To tackle this problem, annotators use natural language, which allows any sound information to be summarized. Automated Audio Captioning (AAC) was introduced recently to develop systems capable of automatically producing a description of any type of sound in text form. This task concerns all kinds of sound events such as environmental, urban, domestic sounds, sound effects, music or speech. This type of system could be used by people who are deaf or hard of hearing, and could improve the indexing of large audio databases. In the first part of this thesis, we present the state of the art of the ...

Air pollution remains a major global challenge, particularly in urban areas where high pollutant concentrations negatively impact public health and contribute to climate change. Among the various pollutants, biogenic volatile organic compounds (BVOCs) play a critical role in atmospheric chemistry, influencing the formation of secondary organic aerosols and ground-level ozone, affecting air quality and climate dynamics. Accurately estimating BVOC emissions at high spatial resolution is challenging due to the limitations of satellite observations and computational models. Additionally, forecasting nitrogen dioxide (NO2) concentrations in urban environments is vital for effective air quality management, yet existing models often struggle to capture complex spatiotemporal dependencies. The thesis aims to address these challenges by proposing novel deep learning (DL) frameworks to tackle two key tasks: (i) improving the spatial resolution of BVOC emission maps through super-resolution (SR) techniques and (ii) developing a robust model ...

Recently, significant advances in the field of automated face recognition have been achieved using computer vision, machine learning, and deep learning methodologies. However, despite claims of super-human performance of face recognition algorithms on select key benchmark tasks, there remain several open problems that preclude the general replacement of human face recognition work with automated systems. State-of-the-art automated face recognition systems based on deep learning methods are able to achieve high accuracy when the face images they are tasked with recognizing subjects from are of sufficiently high quality. However, low image resolution remains one of the principal obstacles to face recognition systems, and their performance in the low-resolution regime is decidedly below human capabilities. In this PhD thesis, we present a systematic study of modern automated face recognition systems in the presence of image degradation in various forms. Based on our ...

Modeling dynamical systems is a fundamental task in various scientific and engineering domains, requiring accurate predictions, robustness to varying conditions, and interpretability of the underlying mechanisms. Traditional data-driven approaches often struggle with long-term prediction accuracy, generalization to out-of-distribution (OOD) scenarios, and providing insights into the system's behavior. This thesis explores the integration of supervised disentanglement into deep learning models as a means to address these challenges. We begin by advancing the state-of-the-art in modeling wave propagation governed by the Saint-Venant equations. Utilizing U-Net architectures and purposefully designed training strategies, we develop deep learning models that significantly improve prediction accuracy. Through OOD analysis, we highlight the limitations of standard deep learning models in capturing complex spatiotemporal dynamics, demonstrating how integrating domain knowledge through architectural design and training practices can enhance model performance. We further extend our supervised disentanglement approach to high-dimensional ...

The effectiveness of machine learning (ML) in today's applications largely depends on the goodness of the representation of data used within the ML algorithms. While the massiveness in dimension of modern day data often requires lower-dimensional data representations in many applications for efficient use of available computational resources, the use of uncorrelated features is also known to enhance the performance of ML algorithms. Thus, an efficient representation learning solution should focus on dimension reduction as well as uncorrelated feature extraction. Even though Principal Component Analysis (PCA) and linear autoencoders are fundamental data preprocessing tools that are largely used for dimension reduction, when engineered properly they can also be used to extract uncorrelated features. At the same time, factors like ever-increasing volume of data or inherently distributed data generation impede the use of existing centralized solutions for representation learning that require ...

Various real-life data such as time series and multi-sensor recordings can be represented by vectors and matrices, which are one-way and two-way arrays of numerical values, respectively. Valuable information can be extracted from these measured data matrices by means of matrix factorizations in a broad range of applications within signal processing, data mining, and machine learning. While matrix-based methods are powerful and well-known tools for various applications, they are limited to single-mode variations, making them ill-suited to tackle multi-way data without loss of information. Higher-order tensors are a natural extension of vectors (first order) and matrices (second order), enabling us to represent multi-way arrays of numerical values, which have become ubiquitous in signal processing and data mining applications. By leveraging the powerful utitilies offered by tensor decompositions such as compression and uniqueness properties, we can extract more information from multi-way ...