The efficacy of endovascular therapy in large vessel occlusion (LVO) of the anterior circulation is dependent to a high degree on the selection of patients who are likely to benefit from this procedure. To this end, functional outcome prediction based on clinical parameters is an active area of research. In the preoperative screening of LVO patients, CT-Angiography (CTA) imaging is commonly acquired. We compare the functional outcome prediction performance of multiple deep learning based classifiers with multiple conventional methods, including the clinically validated MR PREDICTS decision tool. Using a dataset composed of 1929 preprocedural CTA images combined with clinical data, we compare a clinical baseline model with an imaging based pipeline and a combined pipeline. For the imaging model backbone we train various state-of-the-art architectures (Med3D, Vision Transformer, Voxel Transformer). These models are used to predict dichotomized modified Rankin Scale score 90 days after mechanical thrombectomy. Binary classifier outcomes are quantified using Area-Under the receiver operating characteristic Curve (AUC). The activation maps of the best performing image based model are further investigated using the GradCAM++ posthoc visualization method. Combining clinical features with information extracted from CTA images does not significantly improve the performance of functional outcome prediction methods compared to the baseline model. The information extracted from the images does not seem to
be complementary to the clinical features. This multimodal technique can however replace radiologically derived biomarkers, as its performance is non-inferior.

Mind wandering is a phenomenon that is used to describe moments where a person's attention appears to shift away to something that is not related to the primary task, which can have a negative influence on the task performance. In this research, the aim is to create a viable algorithm that can automatically detect mind wandering based on eye movement from the Mementos data set. A method to automatically detect mind wandering could be used in online education in order to help students study more effectively, for example. The Mementos data set contains, unlike previous research, videos captured in an uncontrolled environment using inexpensive equipment. Features based on fixations and saccades were used to create two algorithms which were able to perform better than chance, having an average AOC-ROC of 0.63 and 0.59, as well as having an average F1 score of 0.046 and 0.041 compared to the chance-based model with an F1 score of 0.029.

Spending time in front of screens has become an inescapable activity, which might be interrupted by unrelated external causes. While automatic approaches to identify mind-wandering (MW) have already been investigated, past research was done with self-reports or physiological data. This work explores automated detection utilizing solely facial expressions from Mementos data, which comes in the form of webcam recordings, where participants react to music videos. The recordings are annotated with labels indicating perceived MW. Video responses are turned into time series by first extracting facial characteristics, which are encoded with Facial Action Coding System (FACS). Temporal information is represented with 170 temporal features. Classification is conducted with support vector machines (SVM) through a data-level approach and an algorithm-level approach, first by synthesizing data and second by adding class weights to SVM. Both approaches are evaluated with metric scores insensitive to imbalanced data. On average,
results show that detection performs marginally better than by chance. However, the evaluation metric values vary across multiple classification runs, thus the prospect of using the Mementos dataset for automatic MW detection based on only facial expressions is not promising.

Mind-wandering happens when one's current train of thought, related to a specific task, is interrupted, due to internal disconnected thoughts. This phenomenon is highly subjective, and its detection is really important due to the internal understanding of the human mind that can be obtained. Several methods have been used in order to detect mind-wandering, such as thought probes, self-reports, electrophysiological measures or even eye-based tracking methods, however the detection of mind-wandering solely from sound has not been researched about. Therefore, this study is going to investigate if automatic detection of mind-wandering through sound is feasible. In this work, this question is tackled through a machine learning approach, where a linear SVM model is trained through acoustic features. Two methods of oversampling are considered, due to the high data imbalance between the classes, and these two are evaluated and compared. The approach is evaluated through different metrics, such as recall, precision, F1-score and accuracy, but also a comparison with other techniques is done. Results of this work show that sound as its own is not a reliable way of automatically detecting mind-wandering. These results however, might be implementation specific, as the ground truth values of mind-wandering were created through the means of perceived mind-wandering, and the techniques of random oversampling and SMOTE were also used. These could be causes of unreliability of the research. Future work should should take this into consideration and also apply this approach to a different data set, to assess its feasibility.

Mind wandering occurs when a person’s attention unintentionally shifts away from their current thought or task. Being able to automatically detect cases of mind wandering can assist applications with attention retention, and help people with maintaining focus. Many methods have been tested to deal with mind-wandering detection, but they are mainly conducted in controlled environments.
There also has been little study into the usefulness of learned features from neural networks. This paper is focused on showcasing the effectiveness of
using neural network-generated features as input for classification models. Specifically, using ResNet to generate features which are then used as input by supervised learning models for classification. These features and models were used to classify mind wandering in the Mementos data set, outside of a controlled environment or differently put as “In-the-wild“. The study shows that the extracted features could not be used to accurately detect mind wandering based on the F1-Score (Macro) measure. The results can be attributed to data imbalance, low amount of data, lack of dataset-tailored pre-processing operations, and indiscriminate features. To improve on the study, more data collection is advised and the usage of methods like re-sampling and data augmentation to deal with data imbalance. And lastly, experimentation with neural network training and transforming the data into a time series format to better represent the temporal information from the data.

The aim of this research is to discuss if it is possible or feasible enough to detect Mind-wandering of individuals using their hand and body movements from video recordings. The basis for this research is “Mementos”[9] data set, containing over 2000 recordings of people watching music videos. During experiment videos from data set were used to create software, that would automatically determine Mind-wandering. Results of the study have shown that body and hand movements are useful for detection of the phenomenon, however are not self-sufficient indicators for reliable detection. This study has found a reproducible methodology for automatic detection of Mind-wandering from pre-recorded videos, which is contributing towards exploration of this complex phenomenon in the field of Computer Science.

With the prospects of decentralized multi-agent systems becoming more prevalent in daily life, automated negotiation agents have made their place in these collaborative settings. They are an approach to promote communication between the agents in reaching solutions that are better for all involved.

Recent literature has shown great potential in using machine learning, particularly model-free deep reinforcement learning like Proximal Policy Optimization (PPO), to develop more performant automated negotiation strategies. This work focuses on using information from the opponent's sequence of offers in a bilateral negotiation to further improve a baseline PPO agent. This involves extracting and representing information from the opponent's sequence of offers into a state vector with a fixed dimension to modify the input to the agent's policy, and then comparing the utilities this modified agent achieves to the baseline PPO agent. Since there is a large variety of numerical measures to represent a sequence of offers, an ablation study is conducted to investigate the effectiveness of each.

The modified agents consistently reached solutions that had higher social welfare, although the agent's own utility did not improve or diminish significantly in comparison to the base PPO agent.

This paper introduces a strategy for learning opponent parameters in automated negotiation and using them for future negotiation sessions. The goal is to maximize the agent’s utility while being consistent in its performance over various negotiation scenarios. While a number of reinforcement learning approaches in the field have used Q-learning, this paper uses the newer Proximal Policy Optimization algorithm. Machine learning has been used in opponent modeling, classifying opponents, and learning strategies, but there have been few attempts to store and re-use this information. In an experimental setup, it is shown that this approach outperforms a baseline in terms of individual utility.

Automated negotiation agents can highly benefit from learning their opponent’s preferences. Multiple algorithms have been developed with the two main categories being: heuristic techniques and machine learning techniques. Historically, heuristic techniques have dominated the field, but with the recent development in the field of machine learning, this is no longer true. The main goal of the paper is to compare these two techniques quantitatively using the Pearson correlation of bids. The models that were chosen as the heuristic and machine learning baseline are the Smith and the Perceptron models, respectively. Our results show that the two baselines have similar performance. This leads us to conclude that machine learning algorithms have caught up with their heuristic counterparts. Furthermore, we have also found a statistically significant correlation between the Perceptron model’s accuracy and the seen bid space.

Recent developments in applying reinforcement learning to cooperative environments, like negotiation, have brought forward an important question: how well can a negotiating agent be trained through self-play? Previous research has seen successful application of self-play to other settings, like the games of chess and Go. This paper explores the usage of self-play within the training of a negotiating agent and determines if it is possible to successfully train an agent purely through self-play. The results of the experimentation show that a training stage using self-play can match or even exceed an approach using a set of training opponents. By using multiple self-play opponents, the average utility can be further improved by introducing more variance during training. In addition, using a combination of both self-play and training opponents leads to a hybrid approach that performs better than either of the two techniques separately.

The domains of the negotiation can vary significantly. It is possible that a domain is very cooperative, where both agents can receive a high utility; the opposite is also possible, where the domain is very competitive and the agents cannot both get a high utility. In the same manner, the agents can have different strategies leading to a complicated problem with no obvious solution.

This research seeks to represent the differences in negotiation domains to improve a machine learning based agent to help the agent generalize these domains. To achieve this several ways of representing the domain have been explored.
First is the shared domain information. With this representation, the agent uses information about the amount of issues, values and possible bids there are. Second is the private domain information, in this representation, the agent uses different calculations to get a view of how favorable the domain is in terms of utility. Last is the derived information, this is the representation where the agent learns about the domain by interaction with the environment or the opposing agent.

From the experiments, a conclusion could be made that a part of these representations had a positive impact on the final utility of the agent. The shared domain information had a considerable improvement over the base agent with the features having a non-negligible impact on the negotiation. The derived information also had a considerable impact on the final outcome.

This thesis mainly studies the causality in natural language processing. Understanding causality is key to the success of NLP applications, especially in high-stakes domains. Causality comes in various perspectives such as enable and prevent that, despite their importance, have been largely ignored in the literature. In view of the lack of a dataset that can be used for causality-related research, in this thesis, we first build a first-of-its-kind, fine-grained causal reasoning dataset - FineCR, that contains new causality relations such as enable and prevent, with the help of human annotators. Our dataset contains human annotations of 25K cause-effect event pairs and 24K question-answering pairs within multi-sentence samples, where each can contain multiple causal relationships. To study current NLP models' ability to deal with the causality-related dataset and to figure out the problems that still exist, we define a series of NLP tasks based on FineCR, including causality detection, causality event extraction and causality question answering. Our experimental results with state-of-the-art deep learning models prove that there is still much room for improvement on those causal reasoning tasks. We found that those models have different shortcomings for different tasks. For the causality detection task, current classification models are easily affected by keywords, while the model cannot accurately extract the events for the causality event extraction task. And for the causality question answering task, it is sometimes difficult for the model to find the corresponding answer due to its inability to understand the semantics well. Those discoveries indicate the need to design better solutions to event causality research. In conclusion, our novel datasets and tasks provide a challenging benchmark for evaluating models' causal ability, and the experimental results shed light on future directions for improving neural language models.

Problem Statement. The responsibility problem in the AI field has been taken seriously. The point of the problem is how to prevent AI systems from making moral decisions for whom they cannot be held accountable. One of the directions to address the responsibility problem is Meaningful Human Control (MHC). Most of the work focused on theoretical definition and measurement ex- plorations and few researchers investigate it by combining application scenarios. Research Question. One of the previous works identified that explanations could be used to improve the subjective feeling of MHC when the effect of con- trol is delayed. Therefore, we wanted to study what type of explanations could facilitate the improvement and how explanations achieved that. Method. We conducted an expert study to obtain advice on selecting explanation types by a ranking question questionnaire and structured questions. Due to the findings that information sharing could improve Situation Awareness (SA), we conducted a pilot study and a user study to study the effect of explanations on different levels of SA and the human feeling of MHC. We used the Situation Awareness Global Assessment Technique (SAGAT) and a Five-point Likert scale question- naire to measure the effect of explanations on SA and MHC. Moreover, we compared the effect of sub-explanations we used in the user study (consequen- tial and counterfactual explanations) with a questionnaire. Results. It was shown that the explanations help to get more overall SA and SA in the pro- jection level. It was also shown that higher SA scores are associated with a better feeling of MHC. However, our result showed that the explanations have no significant effect on the subjective experience of MHC. The findings also in- dicated that a high frequency of playing computer games can result in a good subjective experience of control. Comparisons of the sub-explanations on quali- tative and quantitative analysis demonstrated that counterfactual explanations made a better impression on participants in most respects. Discussion and Conclusion. From the results, we concluded that explanations do increase the degree of SA of the task to a certain extent, but they do not affect the experi- enced control. We also found that the computer gaming experience may provide higher cooperation engagement and cohesion, resulting in increased experienced control. As for the sub-explanations, the counterfactual ones were overall bet- ter than the consequential ones by providing more information and making the participants feel the robot’s intelligence.

In the field of computer vision, the difficulty of processing the illumination in an arbitrary scene creates complications for a machine's understanding of the scene. Computer vision algorithms commonly use input images acquired by a camera, where the only measured quantity is the amount of light the camera's sensor receives without any additional information about the shadows, occlusion, highlights, light directionality and interreflections, to help with scene processing and understanding. This thesis applies polarization state measurements of the camera's incoming light to implicitly extract information on these illumination phenomena through a polarization-based intrinsic image decomposition. By defining and measuring the polarization state parameters of an image in the form of polarization cosines, the diffuse and specular intrinsic images can be estimated through an iterative local optimization algorithm to give additional information on the scene illumination. The decomposition algorithm is qualitatively evaluated on three image sample sets, which show that the algorithm performs well as long as the measurements are aligned. The diffuse and specular intrinsic images can then be used to improve existing scene post-processing operations without requiring additional illumination information. One of the three image sample sets is used to evaluate the effects of the decomposition on texture modification, glossiness remapping and tonemapping. The results indicate that the three operations can be either simplified or can have their quality improved with the additional information supplied by the diffuse and specular intrinsic images.

Surgical workflow analysis has gained more importance in operating rooms, which could take responsibility for the working condition, the safety of both patients and surgical personnel, as well as the working efficiency. Focusing on the optimization of the workflow, a set of cameras is installed in the Catheterization Laboratory in Reinier de Graaf Gasthuis for multiple computer vision related researches. However, our cameras are calibrated only once after installation. The orientation and position of the cameras could be changed after days or months, which could lead to a wrong localization.
Compared with the traditional calibration method (calibration patterns or markers), we propose a new image-based camera pose estimation pipeline tested in the Catheterization Laboratory. Our proposed pipeline exploits the object detection model (Scaled-yolov4) to detect fixed objects. The mean average precision with 50\% IoU threshold (mAP@.5) achieves more than 0.99 for all detected objects. Then use a self-supervised key-point detector and descriptor (SuperPoint). With the detected key-points, a feature matching technique based on graph neural networks (SuperGlue) is adopted to match the key-points on the target image with reference points annotated in the image databases (image DBs). The point-correspondences between the image coordinates and the 3D coordinates are applied to solve Perspective-n-Point (PnP) problem to compute the orientation and position of each camera (Camera pose). The final camera pose estimation achieves a 5.79 pixel reprojection error with a 4.97 cm Euclidean distance error.
Compared with other image-based camera pose estimation techniques, our pipeline requires no 3D reconstruction or 3D point cloud in the scene model. Using the video from real procedures, we show that the pipeline can estimate the camera pose with high accuracy.

In inspection and display scenarios, reconstructing and rendering the entire surface of a building is a critical step in presenting the overall condition of the building. In building reconstruction, most works are based on point clouds because of their enhanced availability. In recent years, neural radiance fields (NeRF) have become a common function for implementing novel view synthesis. Compared to other traditional 3D graphic methods, NeRF-based models have a solid ability to produce photorealistic images with rich details that point clouds based methods cannot offer. As a result, we decided to investigate the performance of this technique in architectural scenes and look for ways to improve it for more significant scenes.

This thesis explores the ability to reconstruct large-field scenes with NeRF-based models. NeRF introduced a fully-connected network to predict the volume density and view-dependent emitted radiance at the special location, which will be projected into an image through classic volume rendering techniques. Due to the limitation of near-field ambiguity and parameterization of unbounded scenes, the original NeRF does not perform well on 360° input view, especially when the inputs are sparse. An inverted sphere parameterization that facilitates free view synthesis is introduced to address this limitation so that the foreground and background views can be trained separately. Besides that, we also compare the performance of tracing different light geometries, ray and cone, respectively. Meanwhile, to generate the reconstructed scene precisely, raw RGB images should be pre-processed to estimate the corresponding camera parameters. Finally, customized camera paths should be prepared to generate the final rendered video.

According to our experiments, training foreground and background separately is a promising method to solve practical large-scale scene reconstruction problems. A complete wrap-around view of the target building can be obtained using adjusted camera path parameters. Furthermore, introducing conical frustum casting into the original model also provides an alternative method to implement reconstruction. We named this method mip-NeRF++, which can contribute to the final results to some extent.

Background: In recent years, attention to the genetic architecture of normal-range variation in facial morphology has risen and through GWAS genetic loci associated with facial morphology have been identified. However these give no insight in how the face is shaped by genetics. To investigate the relationship between the genotype and the phenotype, predictive modelling can be used. Predictive modelling is a term used to describe genetic prediction models: tools that aim to predict a phenotype from the genotype.
Objectives: The objective of this project was to investigate the possibility of predictive modelling of facial features from DNA. This method could be used to visualize the variation in facial features caused by the underlying genetics
Methods: In this project the LDAK genetic prediction model was used, this is a software package for phenotype prediction, which uses multilinear regression. Two different datasets were used for this study: Generation R and the Rotterdam Study. Generation R data existed of subjects at the age of nine years old and from the Rotterdam study subjects ≥ 45 years old were included. The datasets were processed individually. From both datasets, phenotype and genotype data was used. The phenotype data consisted of 3D facial meshes that were reduced to 200 endophenotypes with an auto-encoder prior to this study, and genotype data consisted of SNPs acquired using genotyping arrays. The prediction model was trained on 90% of the data, the other 10% was used for testing, where the facial morphology was predicted based on the SNPs. To evaluate the prediction a similarity measure was computed between the predicted faces and the ground truth faces. The similarity measure was computed between each predicted face and all ground truth faces in the test set, thereafter they were ranked in ascending order based on the computed similarity. Next the rank of the true ground truth was determined. Based on the ranking, an accuracy plot was constructed for both datasets and the accuracy ratios (AR) were computed.
Results: For the Generation R dataset the AR found was 0.06 for the Generation R dataset and 0.02 for the Rotterdam Study dataset. The results indicate there is some predictive power, however the AR’s are only slightly above the lower bound for presence of predictive performance. Furthermore, there was a difference in the AR for the Generation R and Rotterdam Study dataset, which could be the result of increased environmental component in facial morphology which reduced the genetic predictability. However, currently the predictive power is minimal. This could be caused by several factors, such as the number of subjects and prediction model that is restricted to only linear relationships.
Conclusion: The objective of this project was to explore the possibility of predictive modelling of facial features from DNA. It was found that there was some predictive power, however this was very limited. Research on predictive modelling of facial features is still in early stages and further research is required to improve the predictive power.