In recent years the use of network analysis has seen an increase across multiple fields. Using features derived from network analysis in a random forest model has already shown promising results in...Show moreIn recent years the use of network analysis has seen an increase across multiple fields. Using features derived from network analysis in a random forest model has already shown promising results in an experiment of the Netherlands Labour Inspection to predict risk scores of violating Dutch labour laws for Dutch companies. In order to see if this is the best way for the NLA to use network analysis this method has been compared with using node classification, a deep learning method that uses sampling and aggregation to label nodes of a network based on its neighbours and the network structure. In this study node classification has shown to perform better at predicting risk scores for Dutch companies than a random forest model with network features does. A simulation study was done on the node classification method to test its robustness and has shown that it is important that there are enough labels in the training set in order for the method to perform well and that the quality of these labels influences the extent to which the model overfits. If the data scientist of the NLA choose to use node classification in future projects it is important that they make an ethical selection of the variables to use for prediction and that they ensure that there are enough labels in the dataset for the node classification model to perform well.Show less
The flood of archaeological remote sensing data in present times calls for digital solutions which can reduce the time and cost required to manually analyse these vast amounts of data. In recent...Show moreThe flood of archaeological remote sensing data in present times calls for digital solutions which can reduce the time and cost required to manually analyse these vast amounts of data. In recent times, Deep Learning techniques based on Covolutional Neural Networks for auto- mated detection of archaeological objects, are fast gaining traction due to the potential they hold. However, much of these studies remain re- stricted to detection of discrete objects with uniform morphologies. Thus, there lies a gap in the use of these methodologies for mapping of larger archaeological systems, which can contribute immensely to landscape archaeology, and our knowledge of human cultural activity in the past. This thesis attempts to make this shift by using a CNN-based in- stance segmentation methodology to detect individual plots of large Celtic Field systems. It was implemented on LiDAR data from the Veluwe region in the Netherlands, using the Mask R-CNN model. The results show that the methodology has the ability to not only detect field plots present in the landscape, but also outline their exact shape. These results when embedded in a wider framework can con- tribute greatly to archaeological prospection and our understanding of the archaeological landscape in the Veluwe.Show less
In this thesis paper the possibility of using Convolutional Neural Networks (CNN) for wavefront reconstruction is studied. This is tested with the Generalised Optical Differentiation Wavefront...Show moreIn this thesis paper the possibility of using Convolutional Neural Networks (CNN) for wavefront reconstruction is studied. This is tested with the Generalised Optical Differentiation Wavefront Sensor (g-ODWFS), a nonlinear pupil plane wavefront sensor that uses spatially varying polarisation rotators in the focal plane. Network architectures are developed for modal reconstruction in the Zernike and actuator basis. Both simulations and experimental work show that using a CNN for wavefront reconstruction improves the estimation of large phase aberrations as compared to the currently used linear method. This increases the convergence rate of the adaptive optics system, which is demonstrated in closed loop simulations.Show less
Deep reinforcement learning has solved the game of Go, along with all other board games. Can it also be applied to real-world use cases? This research combines a literature study and experimental...Show moreDeep reinforcement learning has solved the game of Go, along with all other board games. Can it also be applied to real-world use cases? This research combines a literature study and experimental evaluation, focusing on the case of automation for tele-operated robotics. This is necessary because tele-operation of robots is slow and cumbersome. Classical robotics solutions are expensive, and limited in precision, but deep reinforcement learning provides an opportunity for learning visuomotor skills using partial information.Show less
