Videos manipulated by machine learning have rapidly increased online in the past years. So called deepfakes can depict people who never participated in a video recording by transposing their faces onto others in it. This raises the concern of authenticity of media, which demand for higher performing detection methods in forensics. Introduction of AI detectors have been of interest, but is held back today by their lack of interpretability. The objective of this thesis was therefore to examine what the explainable AI method local interpretable model-agnostic explanations (LIME) could contribute to forensic investigations of deepfake video.  An evaluation was conducted where 3 multimedia forensics evaluated the contribution of visual explanations of classifications when investigating deepfake video frames. The estimated contribution was not significant yet answers showed that LIME may be used to indicate areas to start examine. LIME was however not considered to provide sufficient proof to why a frame was classified as `fake', and would if introduced be used as one of several methods in the process. Issues were apparent regarding the interpretability of the explanations, as well as LIME's ability to indicate features of manipulation with superpixels.

There is a lack of data-driven training instructions for sports shooters, as instruction has commonly been based on subjective assessments. Many studies have correlated body posture and balance to shooting performance in rifle shooting tasks, but most of them have focused on single aspects of postural control. This thesis has focused on finding relevant rifle shooting factors by examining the entire body over sequences of time. We performed a data collection with 13 human participants who carried out live rifle shooting scenarios while being recorded with multiple biometric sensors, including several body trackers. An experiment was conducted to identify what aspects of rifle shooting could be predicted and explained using these data. We employed a pre-processing pipeline to produce a novel skeleton sequence representation, and used it to train a transformer model. The predictions from this model could be explained on a per sample basis using the attention mechanism, and visualised in an interactive format for humans to interpret. It was possible to separate the different phases of a shooting scenario from body posture with a high classification accuracy (81%). However, no correlation could be shown between shooting performance and body posture from our data. Future work could focus on novel feature engineering, and on examining alternative machine learning approaches. The dataset and pre-processing pipeline, as well as the techniques for generating explainable predictions presented in this thesis has laid the groundwork for future research in the sports shooting domain.

Neural Networks are powerful tools used in Machine Learning to solve complex problems across many domains, including biological classification, self-driving cars, and automated management of distributed systems. However, practitioners' trust in Neural Network models is limited by their inability to answer important questions about their behavior, such as whether they will perform correctly or if they can be entrusted with private data. One major issue with Neural Networks is their "black-box" nature, which makes it challenging to inspect the trained parameters or to understand the learned function. To address this issue, this thesis proposes several new ways to increase the trustworthiness of Neural Network models. The first approach focuses specifically on Piecewise Linear Neural Networks, a popular flavor of Neural Networks used to tackle many practical problems. The thesis explores several different techniques to extract the weights of trained networks efficiently and use them to verify and understand the behavior of the models. The second approach shows how strengthening the training algorithms can provide guarantees that are theoretically proven to hold even for the black-box model. The first part of the thesis identifies errors that can exist in trained Neural Networks, highlighting the importance of domain knowledge and the pitfalls to avoid with trained models. The second part aims to verify the outputs and decisions of the model by adapting the technique of Mixed Integer Linear Programming to efficiently explore the possible states of the Neural Network and verify properties of its outputs. The third part extends the Linear Programming technique to explain the behavior of a Piecewise Linear Neural Network by breaking it down into its linear components, generating model explanations that are both continuous on the input features and without approximations. Finally, the thesis addresses privacy concerns by using Trusted Execution and Differential Privacy during the training process. The techniques proposed in this thesis provide strong, theoretically provable guarantees about Neural Networks, despite their black-box nature, and enable practitioners to verify, extend, and protect the privacy of expert domain knowledge. By improving the trustworthiness of models, these techniques make Neural Networks more likely to be deployed in real-world applications.

In the past decade, machine learning (ML) models have become farmore powerful, and are increasingly being used in many important contexts. At the same time, ML models have become more complex, and harder to understand on their own, which has necessitated an interesting explainable AI (XAI), a field concerned with ensuring that ML and other AI system can be understood by human users and practitioners. One aspect of XAI is the development of ”explainers”, tools that take a more complex system (here: an ML model) and generate a simpler but sufficiently accurate model of this system — either globally or locally —to yield insight into the behaviour of the original system. As ML models have become more complex and prevalent, concerns that they may embody and perpetuate harmful social biases have also risen, with XAI being one proposed tool for bias detection. This paper investigates the ability of two explainers, LIME and SHAP, which explain the prediction of potentially more complex models by way of locally faithful linear models, to detect harmful social bias (here in the form of the influence of the racial makeup of a neighbourhood on property values), in a simple experiment involving two kinds of ML models, line arregression and an ensemble method, trained on the well-known Boston-housing dataset. The results show that LIME and SHAP appear to be helpful in bias detection, while also revealing an instance where the explanations do not quite reflect the workings of the model, while still yielding accurate insight into the predictions the model makes.

Structural gendered inequality permeates intelligent systems, shaping everyday lives and reinforcing gender oppression. This study investigates how uncertainty, as an inherent characteristic of Machine Learning (ML) models, can be translated as a design material to highlight gender bias in Artificial Intelligence (AI) systems. It follows an HCI feminist methodology with a threefold horizon: the re-conceptualisation of the design space that considers human and non-human perspectives (Giaccardi & Redström, 2020); the exploration of ML uncertainty as design materiality (Benjamin et al., 2020) to underscore imbued gender inequality in intelligent systems; and the disputed relations of ML uncertainty as materiality with unpredictability in Explainable AI systems, more specifically Graspable AI (Ghajargar et al., 2021, 2022). As a critical exploratory process, the knowledge contribution is the development of a set of guidelines for the design of better and more equal ML systems.

This research focuses on the learning of a class of behaviors defined as high-level behaviors. High-level behaviors are defined here as behaviors that can be executed using a sequence of identifiable behaviors. Represented by low-level contexts, these behaviors are known a priori to learning and can be modeled separately by a knowledge engineer. The learning task, which is achieved by observing an expert within simulation, then becomes the identification and representation of the low-level context sequence executed by the expert. To learn this sequence, this research proposes FAMTILE - the Fuzzy ARTMAP / Template-Based Interpretation Learning Engine. This algorithm attempts to achieve this learning task by constructing rules that govern the low-level context transitions made by the expert. By combining these rules with models for these low-level context behaviors, it is hypothesized that an intelligent model for the expert can be created that can adequately model his behavior. To evaluate FAMTILE, four testing scenarios were developed that attempt to achieve three distinct evaluation goals: assessing the learning capabilities of Fuzzy ARTMAP, evaluating the ability of FAMTILE to correctly predict expert actions and context choices given an observation, and creating a model of the expert's behavior that can perform the high-level task at a comparable level of proficiency.

Technological advancements made in recent decades in the fields of artificial intelligence (AI) and machine learning (ML) has lead to further automation of tasks previously performed by humans. Manually reviewing and assessing content uploaded to social media and marketplace platforms is one of said tasks that is both tedious and expensive to perform, and could possibly be automated through ML based systems. When introducing ML model predictions to a human decision making process, interpretability and explainability of models has been proven to be important factors for humans to trust in individual sample predictions. This thesis project aims to explore the performance of interpretable ML models used together with humans in an ad review process for a rental marketplace platform. Utilizing the XGBoost framework and SHAP for interpretable ML, a system was built with the ability to score an individual ad and explain the prediction with human readable sentences based on feature importance. The model reached an ROC AUC score of 0.90 and an Average Precision score of 0.64 on a held out test set. An end user survey was conducted which indicated some trust in the model and an appreciation for the local prediction explanations, but low general impact and helpfulness. While most related work focus on model performance, this thesis contributes with a smaller model usability study which can provide grounds for utilizing interpretable ML software in any manual decision making process.

State-of-the-art machine learning methods such as convolutional neural networks (CNNs) are frequently employed in computer vision. Despite their high performance on unseen data, CNNs are often criticized for lacking transparency — that is, providing very limited if any information about the internal decision-making process. In some applications, especially in healthcare, such transparency of algorithms is crucial for end users, as trust in diagnosis and prognosis is important not only for the satisfaction and potential adherence of patients, but also for their health. Explainable artificial intelligence (XAI) aims to open up this “black box,” often perceived as a cryptic and inconceivable algorithm, to increase understanding of the machines’ reasoning.XAI is an emerging field, and techniques for making machine learning explainable are becoming increasingly available. XAI for computer vision mainly focuses on image classification, whereas interpretability in other tasks remains challenging. Here, I examine explainability in computer vision beyond image classification, namely in semantic segmentation and 3D multitarget image regression. This thesis consists of five chapters. In Chapter 1 (Introduction), the background of artificial intelligence (AI), XAI, computer vision, and optics is presented, and the definitions of the terminology for XAI are proposed. Chapter 2 is focused on explaining the predictions of U-Net, a CNN commonly used for semantic image segmentation, and variations of this architecture. To this end, I propose the gradient-weighted class activation mapping for segmentation (Seg-Grad-CAM) method based on the well-known Grad-CAM method for explainable image classification. In Chapter 3, I present the application of deep learning to estimation of optical aberrations in microscopy biodata by identifying the present Zernike aberration modes and their amplitudes. A CNN-based approach PhaseNet can accurately estimate monochromatic aberrations in images of point light sources. I extend this method to objects of complex shapes. In Chapter 4, an approach for explainable 3D multitarget image regression is reported. First, I visualize how the model differentiates the aberration modes using the local interpretable model-agnostic explanations (LIME) method adapted for 3D image classification. Then I “explain,” using LIME modified for multitarget 3D image regression (Image-Reg-LIME), the outputs of the regression model for estimation of the amplitudes. In Chapter 5, the results are discussed in a broader context. The contribution of this thesis is the development of explainability methods for semantic segmentation and 3D multitarget image regression of optical aberrations. The research opens the door for further enhancement of AI’s transparency.:Title Page i List of Figures xi List of Tables xv 1 Introduction 1 1.1 Essential Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 1.1.1 Artificial intelligence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 1.1.2 Explainable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 1.1.3 Proposed definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 1.2 Explainable Artificial Intelligence . . . . . . . . . . . . . . . . . . . . . . . . . . 6 1.2.1 Aims and applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 1.2.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 1.3 Computer Vision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 1.3.1 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 1.3.2 Image classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 1.3.3 Image regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 1.3.4 Image segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 1.4 Optics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 1.4.1 Aberrations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 1.4.2 Zernike polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 1.5 Thesis Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 1.5.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 1.5.2 Dissertation outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 2 Explainable Image Segmentation 23 2.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 2.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 2.3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 2.3.1 CAM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 2.3.2 Grad-CAM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 2.3.3 U-Net . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 2.3.4 Seg-Grad-CAM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 2.4 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 2.4.1 Circles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 2.4.2 TextureMNIST . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 2.4.3 Cityscapes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 2.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 2.5.1 Circles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 2.5.2 TextureMNIST . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41 2.5.3 Cityscapes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 2.6 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 2.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52 3 Estimation of Aberrations 55 3.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 3.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 3.3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58 3.3.1 PhaseNet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58 3.3.2 PhaseNet data generator . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 3.3.3 Retrieval of noise parameters . . . . . . . . . . . . . . . . . . . . . . . . 62 3.3.4 Data generator with phantoms . . . . . . . . . . . . . . . . . . . . . . . 62 3.3.5 Restoration via deconvolution . . . . . . . . . . . . . . . . . . . . . . . . 63 3.3.6 Convolution with the “zero” synthetic PSF . . . . . . . . . . . . . . . . 63 3.4 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64 3.4.1 Astrocytes (synthetic data) . . . . . . . . . . . . . . . . . . . . . . . . . 65 3.4.2 Fluorescent beads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66 3.4.3 Drosophila embryo (live sample) . . . . . . . . . . . . . . . . . . . . . . 67 3.4.4 Neurons (fixed sample) . . . . . . . . . . . . . . . . . . . . . . . . . . . 68 3.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 3.5.1 Astrocytes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 3.5.2 Conclusions on the results for astrocytes . . . . . . . . . . . . . . . . . . 74 3.5.3 Fluorescent beads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75 3.5.4 Conclusions on the results for fluorescent beads . . . . . . . . . . . . . . 81 3.5.5 Drosophila embryo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83 3.5.6 Conclusions on the results for Drosophila embryo . . . . . . . . . . . . . 87 3.5.7 Neurons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 3.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96 4 Explainable Multitarget Image Regression 99 4.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 4.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100 4.3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102 4.3.1 LIME . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102 4.3.2 Superpixel algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104 4.3.3 LIME for 3D image classification . . . . . . . . . . . . . . . . . . . . . . 104 4.3.4 Image-Reg-LIME: LIME for 3D image regression . . . . . . . . . . . . . 107 4.4 Results: Classification of Aberrations . . . . . . . . . . . . . . . . . . . . . . . . 109 viii TABLE OF CONTENTS 4.4.1 Transforming the regression task into classification . . . . . . . . . . . . 110 4.4.2 Data augmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 4.4.3 Parameter search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112 4.4.4 Clustering of 3D images . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 4.4.5 Explanations of classification . . . . . . . . . . . . . . . . . . . . . . . . 114 4.4.6 Conclusions on the results for classification . . . . . . . . . . . . . . . . 117 4.5 Results: Explainable Regression of Aberrations . . . . . . . . . . . . . . . . . . 118 4.5.1 Explanations with a reference value . . . . . . . . . . . . . . . . . . . . 121 4.5.2 Validation of explanations . . . . . . . . . . . . . . . . . . . . . . . . . . 122 4.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125 5 Conclusions and Outlook 127 References 129

Explainable AI (XAI) has grown as an important field over the years. As more complicated AI systems are utilised in decision-making situations, the necessity for explanations for such systems is also increasing in order to ensure transparency and stakeholder trust. This study focuses on a specific type of explanation method, namely counterfactual explanations. Counterfactual explanations provide feedback that outlines what changes should be made to the input to reach a different outcome. This study expands on a previous dissertation in which a proof-of-concept tool was created for comparing several counterfactual explanation methods. This thesis investigates the properties of counterfactual explanation methods along with some appropriate metrics. The identified metrics are then used to evaluate and compare the desirable properties of the counterfactual approaches. The proof-of-concept tool is extended with a properties-metrics mapping module, and a user preference-based system is developed, allowing users to evaluate different counterfactual approaches depending on their preferences. This addition to the proof-of-concept tool is a critical step in providing field researchers with a standardised benchmarking tool.

<p dir="ltr">In the realm of deep learning and neural networks, there has been substantial advancement, but the persistent DNN vulnerability to adversarial attacks has prompted the search for more efficient defense strategies. Unfortunately, this becomes an arms race. Stronger attacks are being develops, while more sophisticated defense strategies are being proposed, which either require modifying the model's structure or incurring significant computational costs during training. The first part of the work makes a significant progress towards breaking this arms race. Let’s consider natural images, where all the feature values are discrete. Our proposed metrics are able to discover all the vulnerabilities surrounding a given natural image. Given sufficient computation resource, we are able to discover all the adversarial examples given one clean natural image, eliminating the need to develop new attacks. For remedial measures, our approach is to introduce a random factor into DNN classification process. Furthermore, our approach can be combined with existing defense strategy, such as adversarial training, to further improve performance.</p>