Background Electroencephalography (EEG) using a dry electrode cap is currently being investigated as a pre-hospital stroke triage instrument. Developing an algorithm for automatic interpretation of the EEG signals is challenging, considering the amount of artefacts often in the signal. Ideally, an algorithm should be capable of distinguishing between different artefact types to determine the appropriate action: whether to correct them, reject them, or consider their potential predictive value. Neural networks have demonstrated their value for these types of classifications in wet EEG data. However, this approach requires enormous amount of data and dry EEG data is sparse.
Objective This study aims to develop a multi-class artefact classification model for dry EEG data using transfer learning.
Methods First, a convolutional neural network (CNN) for multi-class (clean, eye movement, muscle activity and electrode artefact) classification was developed. Wet electrode EEG recordings from a publicly available dataset were used, containing data of 213 patients (Part I). Second, this model was implemented and transfer learned for multi-class (clean, pulse artefact, muscle activity and artefact) classification of dry EEG recordings using data of 13 subjects (Part II). The models were trained using annotated multi-channel input. Model performances were evaluated on unseen test data using accuracy, area under the receiver operating characteristic curve, F1-score, precision, and recall. 
Results The pre-trained multi-class model achieved an overall accuracy of 74.8%. The fine-tuned model was able to correctly differentiate between the classes with an accuracy of 71.2%, with the best performance for the classes muscle activity (AUC 0.92, F1-score 0.80) and artefact (AUC 0.94, F1-score 0.80). 
Conclusion Transfer learning enabled the development of a good performing multi-class artefact classification model specifically tailored for dry EEG data, even though available data was limited. The developed model could assist in assessing appropriate action for different artefact types in dry EEG data interpretation algorithms.

The human brain is arguably the most complex system we know of. For centuries, understanding the brain and the way it works has been of great concern to scientists. In 1952, the first mathematical model of a neuron was developed by Hodgkin and Huxley. This work pioneered the field of computational neuroscience, which studies the organization of the brain and the way it processes information. Modeling and simulation of neurons and small brain regions at the cellular level offer great insights into the details of brain mechanics, however, they are far away from clinical applications.

The Virtual Brain (TVB) is a simulation framework that takes a different approach to brain modeling. TVB reduces complexity at the micro level to achieve the macro level of brain modeling and simulation. This approach in brain modeling is referred to as Large-scale Brain Network or Neural-Mass Modeling. Additionally, TVB incorporates individual brain-imaging data with the models to achieve personalized, patient-specific digital brain twins. The digital brain twins can be used in clinical settings such as neurosurgeries and real-time brain-interface systems. However, building and utilizing such large-scale brain models requires high performance and low latency in terms of computation.

Today, heterogeneous architectures, due to their enhanced performance, energy efficiency, and better flexibility are being utilized increasingly in computing systems ranging from low-power edge devices to high-performance cloud infrastructures. Versal Adaptive Compute Acceleration Platform (ACAP) is a high-performance, heterogeneous computing platform developed by Xilinx/AMD. Versal ACAP offers an array of vector processors, specialized in artificial intelligence (AI) and signal processing workloads, next to the traditional programmable logic with extensive memory resources. The large-scale neural-mass simulation problem at hand has heterogeneous computational needs, which makes it interesting to see how it maps to the Versal ACAP. In this work, we explore the benefits and challenges of using heterogeneous systems (specifically, the Versal ACAP) in the high-performance workload of large-scale neural-mass simulation.

We designed and implemented a dataflow-style system for large-scale brain network simulation on the Versal ACAP. Two different versions of this system, namely AIE-Only and Heterogeneous, were developed in order to explore the capabilities of the Versal ACAP when accelerating our application. Compared to the current GPU version of TVB, the Heterogeneous implementation performed on average around 4× slower in terms of throughput. However, the Versal ACAP Heterogeneous implementation delivered around 13× lower latency, two orders of magnitude better energy efficiency, and around 2× better power efficiency compared to the GPU version of TVB. Although the Heterogeneous implementation falls short in terms of throughput compared to the GPU version, its lower latency and energy efficiency make it suitable for real-time applications that might use large-scale neural-mass modeling such as virtual surgery or brain interface devices.

Complex mathematical models are used in computational neuroscience to stimulate brain activity to understand the biological processes involved. The simulation of such models is computationally costly, and thus highperformance
computing systems are selected as a potential solution to increase performance.
This thesis aims to implement a new versatile, multi-GPU eHH simulator (mgpuHH), explore its performance and make general observations on performance scalability over different modeling and cluster configuration properties. This work offers a multinode multi-GPU solution that offers excellent
scalability performance due to how the simulator is constructed, with the use of OpenMPI and CUDA. The simulator is configured with JSON configuration files, containing the neural descriptions and simulatorspecific settings. Consequently, enabling a userfriendly environment, for the neuroscientists, without the need of recompiling or understanding the source code. The gap junction calculations are identified as the critical function bottlenecking performance of the simulator. Therefore, an algorithm tailored to utilize GPU performance is implemented to decrease wallclock time for these specific calculations. For internode
communication, OpenMPI can be configured in two ways. Eiter share all possible compartments potentials with every node in the network or only share the compartments potentials to nodes that need them. These methods rely internally on MPI Allgather and Alltoallv respectively. When available, GPUDirect, NVlink, and RDMA are supported. The implementation hides communication overhead, when possible, by concurrently executable compute kernels. A neuron model from the Inferior Olivary Nucleus is selected for benchmarking. Reported results go up to 32 Nodes with a total of 64 GPU cards. The design shows linear weak and strong scaling within the experimental setups for intranode and internode scalability. With this simulator, networks over 10 million cells become available to model on largescale GPU clusters, setting a new standard for eHH simulations. Comparisons against related work on CPU and FPGAs have been conducted, a 100x speedup is achieved versus a single cpu threaded solution. Furthermore, a 2x speedup is achieved over an FPGA solution (flexHH) and 10 fold over a multithreaded CPU (GenEHH, with 128 threads) solution, both reported speedups are for a fully connected network with 7000 IO cells.