I en framtid där förutsägelser och beslut som tas av maskininlärningssystem överträffar människors förmåga behöver systemen att vara tolkbara för att vi skall kunna lita på och förstå dem. Vår studie utforskar världen av tolkbar maskininlärning genom att designa och undersöka artefakter. Vi genomför experiment för att utforska förklarbarhet, tolkbarhet samt tekniska utmaningar att skapa maskininlärningsmodeller för att identifiera liknande men unika objekt. Slutligen genomför vi ett användartest för att utvärdera toppmoderna förklaringsverktyg i ett direkt mänskligt sammanhang. Med insikter från dessa experiment diskuterar vi den potentiella framtiden för detta fält / In a future where predictions and decisions made by machine learning systems outperform humans we need the systems to be interpretable in order for us to trust and understand them. Our study explore the realm of interpretable machine learning through designing artifacts. We conduct experiments to explore explainability, interpretability as well as technical challenges of creating machine learning models to identify objects that appear similar to humans. Lastly, we conduct a user test to evaluate current state-of-the-art visual explanatory tools in a human setting. From these insights, we discuss the potential future of this field.

Deepfakes technology is a face swap technique that enables anyone to replace faces in a video, with highly realistic results. Despite its usefulness, if used maliciously, this technique can have a significant impact on society, for instance, through the spreading of fake news or cyberbullying. This makes the ability of deepfakes detection a problem of utmost importance. In this paper, I tackle the problem of deepfakes detection by identifying deepfakes forgeries in video sequences. Inspired by the state-of-the-art, I study the ensembling of different machine learning solutions built on convolutional neural networks (CNNs) and use these models as objects for comparison between ensemble and single model performances. Existing work in the research field of deepfakes detection suggests that escalated challenges posed by modern deepfake videos make it increasingly difficult for detection methods. I evaluate that claim by testing the detection performance of four single CNN models as well as six stacked ensembles on three modern deepfakes datasets. I compare various ensemble approaches to combine single models and in what way their predictions should be incorporated into the ensemble output. The results I found was that the best approach for deepfakes detection is to create an ensemble, though, the ensemble approach plays a crucial role in the detection performance. The final proposed solution is an ensemble of all available single models which use the concept of soft (weighted) voting to combine its base-learners’ predictions. Results show that this proposed solution significantly improved deepfakes detection performance and substantially outperformed all single models.

No description available.

No description available.

No description available.

Segmentation of magnetic resonance images is an important part of planning radiotherapy treat-ments for patients with brain tumours but due to the number of images contained within a scan and the level of detail required, manual segmentation is a time consuming task. Convolutional neural networks have been proposed as tools for automated segmentation and shown promising results. However, the data sets used for training these deep learning models are often imbalanced and contain data that does not contribute to the performance of the model. By carefully selecting which data to train on, there is potential to both speed up the training and increase the network’s ability to detect tumours. This thesis implements the method of importance sampling for training a convolutional neural network for patch-based segmentation of three dimensional multimodal magnetic resonance images of the brain and compares it with the standard way of sampling in terms of network performance and training time. Training is done for two different patch sizes. Features of the most frequently sampled volumes are also analysed. Importance sampling is found to speed up training in terms of number of epochs and also yield models with improved performance. Analysis of the sampling trends indicate that when patches are large, small tumours are somewhat frequently trained on, however more investigation is needed to confirm what features may influence the sampling frequency of a patch. / Segmentering av magnetröntgenbilder är en viktig del i planeringen av strålbehandling av patienter med hjärntumörer. Det höga antalet bilder och den nödvändiga precisionsnivån gör dock manuellsegmentering till en tidskrävande uppgift. Faltningsnätverk har därför föreslagits som ett verktyg förautomatiserad segmentering och visat lovande resultat. Datamängderna som används för att träna dessa djupinlärningsmodeller är ofta obalanserade och innehåller data som inte bidrar till modellensprestanda. Det finns därför potential att både skynda på träningen och förbättra nätverkets förmåga att segmentera tumörer genom att noggrant välja vilken data som används för träning. Denna uppsats implementerar importance sampling för att träna ett faltningsnätverk för patch-baserad segmentering av tredimensionella multimodala magnetröntgenbilder av hjärnan. Modellensträningstid och prestanda jämförs mot ett nätverk tränat med standardmetoden. Detta görs förtvå olika storlekar på patches. Egenskaperna hos de mest valda volymerna analyseras också. Importance sampling uppvisar en snabbare träningsprocess med avseende på antal epoker och resulterar också i modeller med högre prestanda. Analys av de oftast valda volymerna indikerar att under träning med stora patches förekommer små tumörer i en något högre utsträckning. Vidareundersökningar är dock nödvändiga för att bekräfta vilka aspekter som påverkar hur ofta en volym används.

Medical imaging is a key tool used in healthcare to diagnose and prognose patients by aiding the detection of a variety of diseases and conditions. In practice, medical image screening must be performed by clinical practitioners who rely primarily on their expertise and experience for disease diagnosis. The ability of convolutional neural networks (CNNs) to extract hierarchical features and determine classifications directly from raw image data makes CNNs a potentially useful adjunct to the medical image analysis process. A common challenge in successfully implementing CNNs is optimizing hyperparameters for training. In this study, we propose a method which utilizes scheduled hyperparameters and Bayesian optimization to classify cancerous and noncancerous tissues (i.e., segmentation) from head and neck computed tomography (CT) and positron emission tomography (PET) scans. The results of this method are compared using CT imaging with and without PET imaging for 2D and 3D image segmentation models.

When two human beings engage in a conversation, feedback is generally present since it helps in modulating and guiding the conversation for the involved parties. When a robotic agent engages in a conversation with a human, the robot is not capable of understanding the feedback given by the human as other humans would. In this thesis, we model human feedback as a Multivariate Time Series to be classified as positive, negative or neutral. We explore state-of-the-art Deep Learning architectures such as InceptionTime, a Convolutional Neural Network approach, and the Time Series Encoder, a Transformer approach. We demonstrate state-of-the art performance in accuracy, loss and f1-score of such models and improved performance in all metrics when compared to best performing approaches in previous studies such as the Random Forest Classifier. While InceptionTime and the Time Series Encoder reach an accuracy of 85.09% and 84.06% respectively, the Random Forest Classifier stays back with an accuracy of 81.99%. Moreover, InceptionTime reaches an f1-score of 85.07%, the Time Series Encoder of 83.27% and the Random Forest Classifier of 77.61%. In addition to this, we study the data classified by both Deep Learning approaches to outline relevant, redundant and trivial human feedback signals over the whole dataset as well as for the positive, negative and neutral cases. / När två människor konverserar, är feedback (återmatning) en del av samtalet eftersom det hjälper till att styra och leda samtalet för de samtalande parterna. När en robot-agent samtalar med en människa, kan den inte förstå denna feedback på samma sätt som en människa skulle kunna. I den här avhandlingen modelleras människans feedback som en flervariabeltidsserie (Multivariate Time Series) som klassificeras som positiv, negativ eller neutral. Vi utforskar toppmoderna djupinlärningsarkitekturer som InceptionTime, en CNN-metod och Time Series Encoder, som är en Transformer-metod. Vi uppnår hög noggrannhet, F1 och lägre värden på förlustfunktionen jämfört med tidigare högst presterande metoder, som Random Forest-metoder. InceptionTime och Time Series Encoder uppnår en noggrannhet på 85,09% respektive 84,06%, men Random Forest-klassificeraren uppnår endast 81,99%. Dessutom uppnår InceptionTime ett F1 på 85,07%, Time Series Encoder 83,27%, och Random Forest-klassificeraren 77,61. Utöver detta studerar vi data som har klassificerats av båda djupinlärningsmetoderna för att hitta relevanta, redundanta och enklare mänskliga feedback-signaler över hela datamängden, samt för positiva, negativa och neutrala datapunkter.

No description available.

Bolagsverket, a Swedish government agency receives cases both in paper form via mail, document form via e-mail and also digital forms. These cases may be about registering people in a company, changing the share capital, etc. However, handling and confirming all these papers can be time consuming, and it would be beneficial for Bolagsverket if this process could be automated with as little human input as possible. This thesis investigates if it is possible to identify whether a paper contains a signature or not by using artificial intelligence (AI) and convolutional neural networks (CNN), and also if it is possible to determine how many signatures a given paper has. If these problems prove to be solvable, it could potentially lead to a great benefit for Bolagsverket. In this paper, a residual neural network (ResNet) was implemented which later was trained on sample data provided by Bolagsverket. The results demonstrate that it is possible to determine whether a paper has a signature or not with a 99% accuracy, which was tested on 1000 images where the model was trained on 8787 images. A second ResNet architecture was implemented to identify the number of signatures, and the result shows that this was possible with an accuracy score of 94.6%.