Spiking Neural Networks (SNNs) have been widely studied as a computational model due to their sparse spiking patterns, asynchronous behavior, and event-oriented processing style. It enables energy-efficient and low-latency processing ideal for real-time tasks, edge computing, and neuromorphic hardware implementations.

However, efficient training of deep SNNs remains a significant challenge, as spikes can easily vanish or lead to an explosion in deep layers if the membrane potentials or the network weights are not properly initialized. The problem of deriving effective initial weights is addressed for deep, feedforward Artificial Neural Networks (ANNs), but still remains underexplored for SNNs. Though the existing ANN-based methods show some benefits for SNNs, the unique computational dynamics of SNNs are often neglected.

Recent studies suggest that the recurrent SNN is a dynamical system that exhibits a critical state, where spikes propagate across multiple time steps, leading to rich information representations, and is believed to be beneficial for neural computation.
Building upon this perspective, this thesis proposes a novel, bio-inspired, criticality-driven initialization method tailored for deep feed-forward SNNs.
We demonstrate the presence of a phase transition in our models, shifting from subcritical to supercritical dynamics.

Moreover, we apply a Hamming distance-based method to assess the network's ability to discriminate between different inputs. The results show that the system effectively maps distinct inputs to corresponding distinct outputs over very deep layer depth at criticality, as indicated by the minimized normalized mean square error (NMSE) between the input and output Hamming distance matrices. Furthermore, we demonstrate that while subcritical and supercritical regimes result in spike vanishing and explosion, respectively, operating at criticality allows spikes to propagate stably throughout the network.

To evaluate the proposed initialization method empirically, we conduct experiments using feedforward SNNs on standard benchmarks such as MNIST and Fashion-MNIST, and further extend the framework to convolutional SNNs for the more complex CIFAR-10 dataset. The results show that, for all datasets, our criticality-based initialization enables higher accuracy and faster convergence in low-latency settings compared to other baseline initialization methods.

Neuromodulation is a promising treatment methodology for disorders of the brain. However, current clinical applications are primarily focused on a rudimentary form of neuromodulation, exclusively targeting individual regions in the brain. This has led to limitations in effective treatment options for patients suffering from neurological and neuropsychiatric diseases, with brain stimulation yielding a success rate of only 50% in 50% of patients after five years. In recent years, it has been strongly theorized that most brain disorders are emergent properties of pathological network connectivity rather than dysfunction of isolated brain regions. Targeting brain connectivity via neuromodulation is, therefore, a promising approach for improving treatment efficacy. This study aims at bridging this therapeutic gap by advancing treatment strategies towards network-based neuromodulation. To achieve this goal, a computational framework was designed and implemented that analyzes brain activity to identify biomarkers of abnormal connectivity networks. Furthermore, software pipelines for the construction of patient-specific,
virtual brain models were designed and implemented which enable analysis of simulated neural activity and connectivity networks, and enable in silico neuromodulation. The framework is tested and validated using an EEG dataset consisting of neural-activity recordings belonging to patients diagnosed with neuropathic pain and a separate healthy control group. Our computational framework shows high ability in identifying disease-specific abnormalities in network connectivity and in constructing pathological network hubs, with accuracies up to 98.6% in explaining neuropathic pain. Patient-specific,
brain-model construction is also able to accurately represent empirical patterns of connectivity in simulation. Furthermore, the simulated neuromodulation with personalized virtual brains demonstrates the ability to accurately model dynamic modulation of brain connectivity networks. These findings illustrate the effectiveness of our overall framework and highlight the potential for integration into clinical decision-making and closed-loop implantable neuromodulation systems, enabling more effective treatment of brain disorders.

Generating arbitrarily long, temporally consistent, and visually realistic street-view videos presents a formidable challenge at the intersection of computer vision and graphics. Existing methods, while capable of producing high-quality individual frames or short sequences, often struggle to maintain coherence and realism over extended durations. Parallel generation approaches, though effective in ensuring frame-to-frame consistency, face limitations in generating videos of arbitrary length and adapting pre-trained text-to-image models for the video domain. In response to these challenges, we propose a novel sequential approach that draws inspiration from recent advancements in 3D-aware image generation and object reconstruction. Our model, built upon the foundations of state-of-the-art techniques like MagicDrive, incorporates innovative conditioning mechanisms to leverage the visual context of previous frames. By establishing strong temporal dependencies and ensuring smooth transitions between frames, our approach enables the creation of coherent and controllable video sequences of any desired length. Through extensive experimentation on the challenging nuScenes dataset, we demonstrate the effectiveness of our sequential generation framework. Our model not only achieves competitive performance compared to existing parallel methods but also offers enhanced flexibility and computational efficiency. The ability to generate arbitrarily long videos opens up new possibilities for applications such as autonomous vehicle simulation, virtual reality training, and urban planning, where realistic and diverse visual data is crucial. Furthermore, our research contributes to a broader understanding of the trade-offs between sequential and parallel generation paradigms, highlighting the potential of sequential approaches for addressing the limitations of current methods.

Whispering, characterized by its soft, breathy, and hushed qualities, serves as a distinct form of speech commonly employed for private communication and can also occur in cases of pathological speech. The acoustic characteristics of whispered speech differ substantially from normally phonated speech and the scarcity of adequate training data leads to low automatic speech recognition (ASR) performance. This project aims to build an ASR system that can recognize both normal and whispered speech and discover which acoustic characteristics of whispered speech have an impact on whispered speech recognition.
In my study, I use signal processing techniques that transform the spectral characteristics of normal speech to those of pseudo-whispered speech, called pseudo-whispered-based data augmentation. I enhance an End-to-End ASR system by incorporating pseudo-whispered speech and state-of-the-art (SOTA) data augmentation methods, speed perturbation and SpecAugment, yielding an 18.2\% relative reduction in word error rate compared to the strongest baseline.
Results for the accented speaker groups in the wTIMIT database show the best results for US English. Further investigation uncovers that the lack of pitch in whispered speech has the largest impact on the performance of whispered speech ASR.

Nowadays, accurate vehicle classification plays a critical role in Advanced Driver Assistant Systems (ADASs), autonomous driving systems, and traffic monitoring systems. The benefits of utilizing additional polarimetric information in road target classification have been revealed in the literature. This thesis investigates the polarimetric characteristics of multi-class vehicles and explores new features contributing to vehicle classification, using a labeled street-way database extracted from the PARSAX S-band polarimetric Frequency Modulated Continuous Wave (FMCW) radar. The vehicle classes involved in this thesis are sedan, sedan with extended luggage bin, mini-van, small truck, and large truck.

Three calibration algorithms are proposed and validated for calibrating the labeled street-way database to ensure feature quality. The first channel calibration algorithm removes the channel-specific amplification factors and biases due to the non-ideal and non-identical electronic devices in the four polarimetric channels of the PARSAX radar. The second phase compensation algorithm compensates the phase difference between the H- and V-polarized channels, which is caused by the time shift between the transmitted H- and V-polarized signals. The last antenna pattern compensation algorithm resolves the power degradation in the measurements due to the PARSAX radar beam width limitation.

Based on the calibrated labeled street-way database, multiple polarimetric features are extracted from the Polarization Scattering Matrices (PSMs), coherency and covariance matrices using the eigenvalues/eigenvectors decomposition methods. These matrices represent either the central bodies or the whole bodies of the vehicles. In each case, the eigenvalues and eigenvectors are analyzed to indicate the vehicles' reflection amplitudes/power and polarization basis. Furthermore, these features are evaluated, and most of the amplitudes/power-based features show great classification capabilities. However, all vehicles have a similar polarization basis, which does not have a contribution to vehicle classification. In addition, the target length and eigenvalues of the covariance matrix of detection cells are also extracted as potential features. The feature evaluation results show that the target length and the first eigenvalue of the covariance matrix of detection cells have classification capabilities, while the second eigenvalue does not contribute to vehicle classification.

Neuromuscular blocking agents (NMBAs) are commonly employed in anesthesia to facilitate intubation and improve surgical working conditions. Understanding their pharmacokinetics (PK) and pharmacodynamics (PD) is crucial for optimizing their administration. PK describes drug absorption, distribution, metabolism, and elimination, while PD links drug concentration to its pharmacological effects. Integrating PK and PD information through PKPD modeling offers insights into the relationship between drug concentration and effect. Recent advancements in machine learning (ML) have shown promise in modeling the PD of anesthetics, offering potential benefits over traditional PKPD models. This study aimed to develop an automated data logger for recording neuromuscular transmission (NMT) measurements and rocuronium infusion data during surgery. A secondary goal was to predict TOF ratios using both traditional PKPD models and newer ML techniques.
The experimental setup involved development of data logger software. Data preprocessing consolidated data, removed outliers, and applied interpolation for missing values. Machine learning models, including linear regression, decision trees, and extreme gradient boosting, were trained and evaluated using double Leave-One-Group-Out cross-validation. Additionally, a traditional PKPD model estimated pharmacokinetic parameters based on patient characteristics and rocuronium administration data. Model performance was assessed using metrics such as Root Mean Squared Error (RMSE), Normalized RMSE (NRMSE), R-squared (R²), and the Pearson correlation coefficient.
From March 23 to June 20, 2023, a prospective observational study at Leiden University Medical Center (LUMC) included 42 patients in the operating room. Data were collected from three distinct monitors using data logger software. The collected data was divided into TOF-Cuff and GE NMT monitor subsets, excluding continuous rocuronium infusion records.
In terms of model performance, machine learning models displayed suboptimal results when applied to GE NMT monitor data, indicated by high RMSE and low R² values. In contrast, basic and optimized PKPD models exhibited better predictive capabilities. Similar trends were observed in the performance evaluation of TOF-Cuff data, with machine learning models less effective compared to PKPD models.
An in-depth analysis of NRMSE revealed outliers, mainly in the optimized PKPD model. Cumulative distribution plots highlighted performance variations across subjects, particularly in the TOF-Cuff results.
The dataset comprised 42 subjects undergoing surgical procedures, but the effective sample size for analysis was limited due to deep neuromuscular blockade, sensor placement issues, and data loss. Traditional pharmacokinetic-pharmacodynamic (PKPD) models, based on data from 423 patients, outperformed machine learning models in predicting Train of Four (TOF) ratios, as the latter faced overfitting challenges with a smaller dataset. Future directions suggest collecting more extensive data (ideally closer to 100 subjects) to improve machine learning model performance and possibly include features like time until full neuromuscular blockade recovery. Additionally, identifying the most critical features behind machine learning predictions can help streamline computational methods. Overall, refining the software, increasing data, and feature analysis can enhance machine learning-based neuromuscular blockade prediction.
In conclusion, this study introduced a novel data logger software for recording neuromuscular blockade data during surgery at LUMC. While machine learning approaches fell short in approximating TOF ratios, future research with expanded datasets and more comprehensive feature analysis holds promise for the development of more robust machine learning models.

Deep learning has been widely implemented in industrial inspection, such as damage detection from images. However, training deep networks requires massive data, which is hard to collect and laborious to annotate, especially in the aviation scenario of aircraft engines. To alleviate the demand for annotated data, we create BladeSynth - a large synthetic image dataset for detecting damage from aircraft engines, and empirically evaluate the transferability of state-of-the-art Scaled-YOLOV4 from synthetic to real world by pre-training on synthetic data and fine-tuning on real data. Our experiments show that pre-training on synthetic data improves the performance in damage detection in aircraft engine images.

With the advent of offshore wind farms, the research into the various phenomenon that affects their performance is vast and detailed. But the effect of a particular phenomenon, atmospheric gravity waves (AGWs), on wind farm performance is limited. AGWs are oscillations of the airflow due to an imbalance in the buoyancy and gravity forces, generated by topographical or meteorological obstacles in neutral or stable surface atmospheric conditions.

AGWs are frequent over offshore regions and affect offshore wind farms as the event occurs over a large area. Detecting them through satellite images is easy by an eye test, but not so much when viewed digitally through meteorological data. Weather data can be obtained from reanalysis data which combines past weather forecasts with observational data assimilation.

This project aims to develop machine learning models that detect AGWs in satellite images and detect AGWs from atmospheric conditions, such as temperature and wind speed profile with height. The models learn using the reanalysis data and satellite images. The same satellite images are used to label the reanalysis data so that the model is taught to pick out gravity waves in the case of having no satellite image. Thus the final objective of the project is to train a model to detect an AGW event, based solely on reanalysis data. The trained model is then used to predict the percentage of time an AGW occurs or will occur over a chosen wind farm site.

The recent advances in wireless communication, micro-fabrication, and integration have led to the rise of multi-agent networks such as wireless sensor networks, drone swarms, and satellite arrays. These multi-agent networks comprise multiple agents, which cooperate to solve a problem collectively, beyond the individual knowledge of a single agent. Multi-agent networks are also known for improving overall system performance, especially in terms of scalability, robustness, and flexibility.

Multi-agent networks are tasked with a variety of tasks e.g., target tracking, surveillance, traffic control, and environmental monitoring. Traditionally these tasks are solved using distributed state-space filtering methods, however under the assumptions of linear models with underlying Gaussian noise. However, in general, the models could be Non-linear, and the underlying stochastic noise is not necessarily Gaussian in nature. To address this limitation, Distributed particle filters (DPF) have been investigated in recent years. Depending on the quantities exchanged in the network, DPFs can be categorized as weight-based, likelihood-based, or posterior-based.

One of the key challenges for effective implementation of the DPF is the need for approximation strategies to reduce the burden on communication and computation resources to suit engineering applications. To this end, we review two state-of-the-art DPF, (a) one based on applying the Graph Laplacian (GL) approximation to weights, and (b) the other using the Gaussian mixture model (GMM) to represent the local posteriors. These two algorithms significantly reduce the communication burden; however, the main limitation is the synchronization and heavy computational burden, which places high demands on hardware and software implementation of agents.

To improve on these existing methods, we propose the use of Gaussian process (GP) enhanced resampling algorithm, which reduces the particle set size, while ensuring an acceptable filter performance. Based on this GP-enhanced resampling scheme, we propose a novel Distributed particle filtering algorithm. Secondly, based on the proposed GP-enhanced sampling, we propose several metaheuristic optimization algorithms (i.e., genetic algorithm and firefly algorithm) to fuse the information across the whole network to improve the estimation performance (i.e., to seek the global optimal particle set as we can treat particles as “parents” or “fireflies”).

We compare our proposed algorithms and study their performance against the state-of-the-art DPFs in terms of time, space, and communication complexity, for a target tracking application. We use time-averaged root mean square error (ARMSE), which is used to quantify the performance of each DPF. Our numerical results demonstrate the superiority of our proposed algorithms under the condition of limited communication and computational resources, and still maintaining the optimality of particle filter.

For example, to achieve a certain ARMSE for a target tracking application, the traffic for existing GL and GMM based DPFs, is 12000 scalars/second/agent and 600 scalars/second/agent respectively, while our proposed GP-enhanced resampling and firefly algorithm uses only 900scalars/second/agent. In terms of the computation runtime, the GL and GMM based approaches take 40 and 300 seconds (about 5 minutes) respectively, while our proposed algorithm takes just 1.5 seconds, and thus shows significant improvement. Finally, we propose research directions to further improve our proposed algorithms in the future.

This thesis studies the Canonical Polyadic Decomposition (CPD) constrained kernel machine for large scale learning, i.e. learning with a large number of samples. The kernel machine optimization problem is solved in the primal space, such that the complexity of the problem scales linearly in the number of samples as opposed to scaling cubically in the dual space. Product feature maps are applied to transform the input data. The weights are constrained to be a CPD, so the number of weights scales linearly in the number of features. The CPD introduces a nonlinearity, so nonlinear optimization must be applied.
It is studied in which situation it is more advantageous to apply iterative all-at-once opti- mization compared to Alternating Least Squares (ALS) to solve the CPD constrained kernel machine problem. Specifically, all-at-once gradient descent methods are studied. An efficient analytical algorithm for the all-at-once gradient is derived. Furthermore, it is shown that automatic differentiation (AD) can also be applied, but it is found to be slower than the analytical method.
The selection of a step size is found to be challenging. It is shown that the magnitude of the gradient of the mean squared error (MSE) term decreases for an increasing number of features. As a result, selecting the step size becomes more difficult for more features. To overcome this, the Line search and the Adam method are studied. A general expression for the exact line search solution is derived. It can be applied to compute the optimal step size for any step direction and any number of features. However, the Adam method performs better in terms of loss after training, convergence and the training run time. The mini-batch Adam method is used to evaluate the performance of all-at-once optimization for large scale learning.
It is found that the Adam method no longer performs well for data sets with around 16 features or more, likely due to the decrease in the magnitude of the gradient of the MSE term. On large scale data sets with fewer features, the Adam method outperforms ALS in terms of run time until convergence while achieving similar training and validation losses. The Adam method reached convergence on a data set with 11 million samples within ten minutes. Furthermore, it is shown that the scaling of the run time of the Adam method in terms of the feature map order and the CP-rank is more than an order lower than the scaling of ALS when the methods are run on a GPU. This makes to Adam method more suitable for more complex models.

The issue of securing microchip designs against hardware attacks has grown in magnitude as more and more embedded systems are deployed in hostile environments, where security measures have to be taken to prevent attackers from accessing unwanted information.
The first step in solving this problem is gaining awareness of the security vulnerabilities in a design, which can be done through a fault injection campaign. Current solutions tackle the issue by either requiring silicon manufacturing for every prototype, which is expensive, or simulating faults using a model of the device under test and of the possible faults being injected, which takes a considerable amount of time. Some hybrid solutions have been developed to improve this aspect (namely, by using FPGAs to implement the device and injecting faults on it instead), but they still require some form of specialized hardware to operate.
Moreover, a gap still remains between the results from these tools and the work necessary to update the design and mitigate the vulnerabiliies found. Results usually reference cells in the netlist, while the development is mostly done in a high-level hardware descriptive language.
This thesis proposes two improvements to the currently existing workflow: first, it explores the effectiveness of equivalence checking in tracing individual gates in the netlist representation back to the RTL lines that generated them, and second, it builds on the simulation-based fault injection approach by introducing a formal framework to prove the presence or absence of successful faults.
The results show that equivalence checking can significantly increase the number of cells recognized as results of individual RTL lines of code, enabling a designer to better pinpoint which components should be hardened against fault injection attacks. In terms of fault identification, the framework described can reduce the number of faults to be simulated down to 2% of the number necessary to exhaustively check a design for possible vulnerabilities, greatly speeding up existing simulation-based approaches.

Performing joint tracking and classification is the ultimate goal for many radar-based applications. For example, in indoor monitoring scenario, it is important to know the target's position as well as the related activities performed by that target. However, the literature often treats the joint problem independently due to its complexity. As a consequence, the dependencies and requirements between the tracking and recognition system are neglected. The main objective of this thesis work is to build two connectable systems for tracking and classifying human activities with radar sensors and address the problems caused by the mutual requirements through system designs. To achieve it, this thesis work proposed a multiple target tracking (MTT) system and a human activity recognition (HAR) system based on a distributed IR-UWB radar sensor network. The result shows that the MTT system is able to track multiple extended targets and extract the Doppler characteristics, and the HAR system provides an end-to-end solution for data fusion and activity classification. In summary, this work provides a foundation for their combination and shows improvements in both tracking and classification compared to the state-of-the-art.