Federated Learning (FL) revolutionizes distributed machine learning, enabling clients to learn collaboratively while keeping data private. In contrast, Decentralized Federated Learning (DFL) offers direct communication between clients without a central server, improving fault tolerance and network efficiency, but communication overhead remains a challenge. To address this, we propose a new scheme called Low Huffman-coded Delta Quantization (LHDQ) which achieves a remarkable quantization rate of $\frac{5}{3}$ bits per parameter. We evaluate LHDQ within the DFL architecture under two proposed transmission protocols and compare it against conventional quantization schemes under various communication channel conditions. Despite a slight reduction in accuracy, LHDQ offers compelling advantages as alleviating communication bottlenecks, reducing transmitted bits, and accelerating training and convergence processes.

The classification of Electrocardiograms (ECG) sig-nals is vital for diagnosing and monitoring cardiovascular diseases. Recently, there has been increasing interest in leveraging efficient machine learning models for ECG classification. Op-timizing computational resources in this domain is crucial for real-time analysis and reducing power consumption. This paper explores two approaches to enhance the efficiency of ECG classification on edge devices. Firstly, a smart lead-switching mechanism intelligently switches between a high-power model and a low- power model. Secondly, Apache Tensor Virtual Machine (TVM) is employed to optimize the implementation of the developed models on the Jetson Nano single-board computer. Experimental results demonstrate noticeable achievements, including up to 96% reduction in FLOPS and 95% reduction in inference time. Our results hold significant promise for improving ECG classification on edge devices, enabling early detection of abnormalities, and enhancing patient care.

In the era of big data, new learning techniques are emerging to solve the difficulties of data collection, storage, scalability, and privacy. To overcome these challenges, we propose a distributed learning system that merges the hybrid edge-cloud split-learning architecture with the semi-supervised learning scheme. The proposed system based on three semisupervised learning algorithms (FixMatch, Virtual Adversarial Training, and MeanTeacher) is compared to the supervised learning scheme and trained on different datasets and data distributions (IID and non-IID) and with a variable number of clients. The new system could efficiently utilize the local unlabeled samples on the client side and gave a performance encouragement that exceeds 30% in most cases even with small percentage of labelled data. Additionally, certain Split-SSL algorithms showed performance that was on par with or occasionally even better than more resource-intensive algorithms, although requiring less processing power and convergence time.

This work addresses the problem of 3D-localizing and enhancing the speech of one main speaker in noisy multi-speaker hospital environments using a multi-channel microphone array. In our model, we propose conducting speaker localization using a machine learning model based on convolutional recurrent neural networks (CRNN) followed by minimum variance distortionless response (MVDR) beamforming. In addition, to ensure that our speech enhancement module is adaptive when deployed in different environments, we trained a meta learning model. Firstly, in the localization step, an estimation of the direction of arrival (DOA) in the elevation and azimuth planes is executed. This is conducted in a 3D space with the presence of noise, reverberation, and up to two more speakers. Using estimated DOA, the MVDR beamformer then enhances the speech of the main speaker. In order to test our model, we adopted and simulated a real-world problem where the objective was to enhance the speech of a clinician in a noisy intensive care unit (ICU) with the presence of other speakers. Furthermore, in order to validate our model, we adopted a speech-to-text module to evaluate the word error rate. Moreover, we implemented our algorithm on hardware using commercially available components and tested it in a real environment. Results showed that our model outperforms other machine learning and non-machine learning algorithms. Finally, we used our trained Meta Learning model to show that our model can adapt to new environments while maintaining high performance after retraining with only a few-shot recordings.

Electrocardiography (ECG) is a standard diagnostic tool for evaluating the overall heart's electrical activity and is vital for detecting many cardiovascular diseases. Classifying ECG recordings using deep neural networks has been investigated in literature and has shown very good performance. However, this performance assumes that the training data is centralized, which is often not the case in real-life scenarios, where data resides in multiple places and only a small portion of it is labeled. Therefore, in this work, we propose an ECG classification system that focuses on preserving data privacy and enhancing overall system efficiency. We analyzed the complexity of previously proposed deep learning-based models and showed that the temporal convolutional network-based models (TCN) were the most efficient. Then, we built on the TCN models a modified split-learning (SL) system that achieves the same classification performance as the basic SL but reduces the communication overhead between the server and the client by 71.7% as well as reducing the computations at the client by 46.5% compared to the original SL system based on the TCN network. Finally, we implement semi-supervised learning in our system to enhance its classification performance by 9.1%-15.7%, when the training data consists only of 10% labeled data. We have tested our proposed system on a test IoT setup and it achieved satisfactory classification accuracy while being private and energy efficient for green-AI applications.

Each year, pneumonia causes harm to millions of people worldwide. The chest radiograph remains the most important tool for diagnosing pneumonia and many other thoracic diseases. Many deep-learning approaches have been advocated in recent years, taking advantage of large public medical datasets and powerful computational resources to support radiologists in their examinations. However, the vast amount of data available is rarely utilized. In this work, we propose two novel attention-based mechanisms for fusing multimodal and multi-view data in a medical context, respectively. A unimodal multi-view classifier for chest X-rays and a multimodal classifier are developed, inter alia, operating these mechanisms to classify pneumonia using chest radiographs in combination with other clinically derived modalities. Our unimodal multi-view model outperforms other state-of-the-art approaches on the MIMIC-IV dataset in the AUROC metric by at least 9.2%, achieving an AUROC score of 81.5%. Our attention-based multimodal model, namely Radiopaths, achieved an AUROC score of 87.3%, further increasing the multi-view model's performance by 5.8%, while acquiring robustness to missing modalities and scalability.