As much as the diverse and rich offer on e-commerce websites helps the users find what they need at one market place, the online catalogs are sometimes too overwhelming. Recommender systems play an important role in e-commerce websites as they improve the customer journey by helping the users find what they want at the right moment. These recommendations can be based on users’ characteristics, demographics, purchase or session history.In this thesis we focus on identifying complementary relationship between products in the case of the largest e-commerce company in the Netherlands. Complementary products are products that go well together, products that might be a necessity to the chosen product or simply a nice addition to it. At the company, there is big potential as complementary products increase the average purchase value and they exist for less than 20% of the whole catalog.We propose a content-based recommender system for detecting complemen- tary products, using a supervised deep learning approach that relies on Siamese Neural Network (SNN).The purpose of this thesis is three-fold; Firstly, the main goal is to create a SNN model that will be able to predict complementary products for any given product based on the content. For this purpose, we implement and compare two different models: Siamese Convolutional Neu- ral Network and Siamese Long Short-Term Memory (LSTM) Recurrent Neural Network. We feed these neural networks with pairs of products taken from the company, which are either complementary or non-complementary. Secondly, the basic assumption of our approach is that most of the important features for a product are included in its title, but we conduct experiments including the product description and brand as well. Lastly, we propose an extension of the SNN approach to handle millions of products in a matter of seconds.∼As a result from the experiments, we conclude that Siamese LSTM can predict complementary products with highest accuracy of 85%. Our assumption that the title is the most valuable attribute was confirmed. In addition, trans- forming our solution to a K-nearest-neighbour problem in order to optimize it for millions of products gave promising results. / Så mycket som det mångfaldiga och rika utbudet på e-handelswebbplatser hjälper användarna att hitta det de behöver på en marknadsplats, är online- katalogerna ibland för överväldigande. Rekommendationssystem en viktig roll på e-handelswebbplatser eftersom de förbättrar kundupplevelsen genom att hjälpa användarna att hitta vad de vill ha i rätt ögonblick. Dessa rekommen- dationer kan baseras på användarens egenskaper, demografi, inköps- eller ses- sionshistorik.I denna avhandling fokuserar vi på att identifiera komplementära förhållanden mellan produkter för det största e-handelsföretaget i Nederländerna. Komplet- terande produkter är produkter passar väl ihop, produkter som kan vara en nödvändighet för den valda produkten eller helt enkelt ett trevligt tillskott till den. På företaget finns det stor potential eftersom kompletterande produkter ökar det genomsnittliga inköpsvärdet och de tillhandahålls för mindre än 20% av hela katalogen.Vi föreslår ett innehållsbaserat rekommendationssystem för att upptäcka kom- pletterande produkter, med en övervakad strategi för inlärning som bygger på Siamese Neural Network (SNN). Syftet med denna avhandling är i tre steg; För det första är huvudmålet att skapa en SNN-modell som kan förutsäga komplet- terande produkter för en given produkt baserat på innehållet. För detta ändamål implementerar och jämför vi två olika modeller: Siamese Convolutional Neu- ral Network och Siamese Long Short-Term Memory (LSTM) Recurrent Neural Network. Vi matar in data i dessa neurala nätverk med par produkter hämta- de från företaget, som antingen är komplementära eller icke-komplementära. Det andra grundläggande antagandet av vår metod att de flesta av de viktiga funktionerna för en produkt ingår i dess titel, men vi genomför också expe- riment inklusive produktbeskrivningen och varumärket. Slutligen föreslår vi en utvidgning av SNN-metoden för att hantera miljoner produkter på några sekunder.∼Som ett resultat av eperimenten drar vi slutsatsen att Siamese LSTM kan för- utsäga komplementära produkter med högsta noggrannhet på 85%. Vårt antagande att titeln är det mest värdefulla attributet bekräftades. Därtill är om- vandling av vår lösning till ett K-närmaste grannproblem för att optimera den för miljontals produkter gav lovande resultat.

Submitted by PPG Ci?ncia da Computa??o (ppgcc@pucrs.br) on 2017-11-21T17:02:34Z No. of bitstreams: 1 Alan_Diego_dos_Santos_dis.pdf: 1881856 bytes, checksum: cf0113b0b67e0771e4b2920440d41e2b (MD5) / Rejected by Caroline Xavier (caroline.xavier@pucrs.br), reason: Devolvido devido ? falta da folha de rosto (p?gina com as principais informa??es) no arquivo PDF, passando direto da capa para a ficha catalogr?fica. on 2017-11-29T19:03:08Z (GMT) / Submitted by PPG Ci?ncia da Computa??o (ppgcc@pucrs.br) on 2017-11-30T15:50:58Z No. of bitstreams: 1 Alan_Diego_dos_Santos_dis.pdf: 1884320 bytes, checksum: 6e508a972289e66527fd4b76cbae3586 (MD5) / Approved for entry into archive by Caroline Xavier (caroline.xavier@pucrs.br) on 2017-12-04T16:14:52Z (GMT) No. of bitstreams: 1 Alan_Diego_dos_Santos_dis.pdf: 1884320 bytes, checksum: 6e508a972289e66527fd4b76cbae3586 (MD5) / Made available in DSpace on 2017-12-04T16:18:35Z (GMT). No. of bitstreams: 1 Alan_Diego_dos_Santos_dis.pdf: 1884320 bytes, checksum: 6e508a972289e66527fd4b76cbae3586 (MD5) Previous issue date: 2017-03-29 / Triagem virtual de bancos de dados de ligantes ? amplamente utilizada nos est?gios iniciais do processo de descoberta de f?rmacos. Abordagens computacionais ?docam? uma pequena mol?cula dentro do s?tio ativo de um estrutura biol?gica alvo e avaliam a afinidade das intera??es entre a mol?cula e a estrutura. Todavia, os custos envolvidos ao aplicar algoritmos de docagem molecular em grandes bancos de ligantes s?o proibitivos, dado a quantidade de recursos computacionais necess?rios para essa execu??o. Nesse contexto, estrat?gias de aprendizagem de m?quina podem ser aplicadas para ranquear ligantes baseadas na afinidade com determinada estrutura biol?gica e, dessa forma, reduzir o n?mero de compostos qu?micos a serem testados. Nesse trabalho, propomos um modelo para ranquear ligantes baseados na arquitetura de redes neurais siamesas. Esse modelo calcula a compatibilidade entre receptor e ligante usando grades de propriedades bioqu?micas. N?s tamb?m mostramos que esse modelo pode aprender a identificar intera??es moleculares importantes entre ligante e receptor. A compatibilidade ? calculada baseada em rela??o ? conforma??o do ligante, independente de sua posi??o e orienta??o em rela??o ao receptor. O modelo proposto foi treinado usando ligantes ativos previamente conhecidos e mol?culas chamarizes (decoys) em um modelo de receptor totalmente flex?vel (Fully Flexible Receptor - FFR) do complexo InhA-NADH da Mycobacterium tuberculosis, encontrando ?timos resultados. / Structure-based virtual screening (SBVS) on compounds databases has been widely applied in early stage of the drug discovery on drug target with known 3D structure. In SBVS, computational approaches usually ?dock? small molecules into binding site of drug target and ?score? their binding affinity. However, the costs involved in applying docking algorithms into huge compounds databases are prohibitive, due to the computational resources required by this operation. In this context,different types of machine learning strategies can be applied to rank ligands, based on binding affinity,and to reduce the number of compounds to be tested. In this work, we propose a deep learning energy-based model using siamese neural networks to rank ligands. This model takes as inputs grids of biochemical properties of ligands and receptors and calculates their compatibility. We show that the model can learn to identify important biochemical interactions between ligands and receptors. Besides, we demonstrate that the compatibility score is computed based only on conformation of small molecule, independent of its position and orientation in relation to the receptor. The proposed model was trained using known ligands and decoys in a Fully Flexible Receptor model of InhA-NADH complex (PDB ID: 1ENY), having achieved outstanding results.

Research on aversive control of behaviour has dramatically declined over the past decades. This trend is primarily a consequence of an over-reliance on shock-based procedures, which have been increasingly criticized on ethical, practical, and ecological validity grounds. The continued study of aversive regulation thus requires the development of viable alternatives. Six preliminary experiments, triggered by serendipitous observations of Betta splendens’ reactions to unintended water disturbances, allowed for (a) developing a water flows (WFs) experimental paradigm; (b) confirming the aversive function of WFs; and (c) demonstrating the feasibility of the WFs paradigm as an alternative to the use of electric shock, as it does not involve painful stimulation and carries a higher level of inherent ecological validity. Based on the relevance of free-operant avoidance phenomena (Sidman, 1953a) for the study of aversive control, the fact that these have only been demonstrated in one fish species (goldfish) using shocks, and that the only attempt to show another form of avoidance in Betta splendens produced inconclusive results (Otis & Cerf, 1963), the WFs paradigm was implemented in two experiments aimed at addressing these issues. These studies were aligned with a research program on spatiotemporal analysis of behaviour that has demonstrated, over the course of several decades, that a comprehensive understanding of behavioural processes requires an approach that includes, but is not limited to, the study of rates of discrete responses (e.g., key pecks of a pigeon). Accordingly, a more holistic interpretation of experimental data than is typical for behavioural studies was attained through a combined analysis of the frequency and temporal distribution of a target response (crossings in a shuttle-tank), patterns of swimming trajectories, instances and durations of the aversive stimulus, and the occurrence of behaviour related to different features of the experimental tank. In Experiment 1, Betta splendens exposed to a free-operant avoidance procedure reliably escaped WFs but did not develop avoidance behaviour even though escape improved with practice. Moreover, adding a warning stimulus (curtains of air bubbles - CABs) to the free-operant procedure did not produce increments in avoidance behaviour, as has been demonstrated in other species. Considering these findings, Experiment 2 maintained the same free-operant avoidance contingencies, but escape responses were now scheduled to produce the WFs (punishment and extinction of escape). The result of this manipulation was not a substantial decrease of escape, but an initial large increase of this response, followed by a progressive decrease to approximately pre-punishment levels. In addition, punishment did not result in increased avoidance responding as an alternative response. The explanations for these unexpected findings relate to the duration of the CABs; sign- and goal-tracking effects; uncontrolled stimulation produced by water pump activation/operation; unintended reinforcement (mirror reflections and delay between the pump activation and WFs reaching full strength); and the development of responses that allowed the fish to reduce their exposure to high-intensity WFs (i.e., alternative behaviour). The need for investigating the effects of adjusting the WF procedures to the ecology and biology of Betta splendens is also discussed, particularly in regard to their territoriality and predominant defensive response (immobility) in relation to the experimental apparatuses and the target response (changing compartments).

Submitted by Luciana Ferreira (lucgeral@gmail.com) on 2018-11-21T11:05:44Z No. of bitstreams: 2 Dissertação - Vinícius Araújo Santos - 2018.pdf: 14601054 bytes, checksum: e02a8bcd3cdc93bf2bf202c3933b3f27 (MD5) license_rdf: 0 bytes, checksum: d41d8cd98f00b204e9800998ecf8427e (MD5) / Approved for entry into archive by Luciana Ferreira (lucgeral@gmail.com) on 2018-11-21T11:06:26Z (GMT) No. of bitstreams: 2 Dissertação - Vinícius Araújo Santos - 2018.pdf: 14601054 bytes, checksum: e02a8bcd3cdc93bf2bf202c3933b3f27 (MD5) license_rdf: 0 bytes, checksum: d41d8cd98f00b204e9800998ecf8427e (MD5) / Made available in DSpace on 2018-11-21T11:06:26Z (GMT). No. of bitstreams: 2 Dissertação - Vinícius Araújo Santos - 2018.pdf: 14601054 bytes, checksum: e02a8bcd3cdc93bf2bf202c3933b3f27 (MD5) license_rdf: 0 bytes, checksum: d41d8cd98f00b204e9800998ecf8427e (MD5) Previous issue date: 2018-10-11 / Coordenação de Aperfeiçoamento de Pessoal de Nível Superior - CAPES / Visual Odometry is an important process in image based navigation of robots. The standard methods of this field rely on the good feature matching between frames where feature detection on images stands as a well adressed problem within Computer Vision. Such techniques are subject to illumination problems, noise and poor feature localization accuracy. Thus, 3D information on a scene may mitigate the uncertainty of the features on images. Deep Learning techniques show great results when dealing with common difficulties of VO such as low illumination conditions and bad feature selection. While Visual Odometry and Deep Learning have been connected previously, no techniques applying Siamese Convolutional Networks on depth infomation given by disparity maps have been acknowledged as far as this work’s researches went. This work aims to fill this gap by applying Deep Learning to estimate egomotion through disparity maps on an Siamese architeture. The SiameseVO-Depth architeture is compared to state of the art techniques on OV by using the KITTI Vision Benchmark Suite. The results reveal that the chosen methodology succeeded on the estimation of Visual Odometry although it doesn’t outperform the state-of-the-art techniques. This work presents fewer steps in relation to standard VO techniques for it consists of an end-to-end solution and demonstrates a new approach of Deep Learning applied to Visual Odometry. / Odometria Visual é um importante processo na navegação de robôs baseada em imagens. Os métodos clássicos deste tema dependem de boas correspondências de características feitas entre imagens sendo que a detecção de características em imagens é um tema amplamente discutido no campo de Visão Computacional. Estas técnicas estão sujeitas a problemas de iluminação, presença de ruído e baixa de acurácia de localização. Nesse contexto, a informação tridimensional de uma cena pode ser uma forma de mitigar as incertezas sobre as características em imagens. Técnicas de Deep Learning têm demonstrado bons resultados lidando com problemas comuns em técnicas de OV como insuficiente iluminação e erros na seleção de características. Ainda que já existam trabalhos que relacionam Odometria Visual e Deep Learning, não foram encontradas técnicas que utilizem Redes Convolucionais Siamesas com sucesso utilizando informações de profundidade de mapas de disparidade durante esta pesquisa. Este trabalho visa preencher esta lacuna aplicando Deep Learning na estimativa do movimento por de mapas de disparidade em uma arquitetura Siamesa. A arquitetura SiameseVO-Depth proposta neste trabalho é comparada à técnicas do estado da arte em OV utilizando a base de dados KITTI Vision Benchmark Suite. Os resultados demonstram que através da metodologia proposta é possível a estimativa dos valores de uma Odometria Visual ainda que o desempenho não supere técnicas consideradas estado da arte. O trabalho proposto possui menos etapas em comparação com técnicas clássicas de OV por apresentar-se como uma solução fim-a-fim e apresenta nova abordagem no campo de Deep Learning aplicado à Odometria Visual.

Current systems for counting cars on parking lots usually use specialized equipment, such as barriers at the parking lot entrance. Usage of such equipment is not suitable for free or residential parking areas. However, even in these car parks, it can help keep track of their occupancy and other data. The system designed in this thesis uses the YOLOv4 model for visual detection of cars in photos. It then calculates an embedding vector for each vehicle, which is used to describe cars and compare whether the car has changed over time at the same parking spot. This information is stored in the database and used to calculate various statistical values like total cars count, average occupancy, or average stay time. These values can be retrieved using REST API or be viewed in the web application.

<div><div><div><p>Face redaction is needed by law enforcement and mass media outlets to guarantee privacy. In this thesis, a performance analysis of several face redaction/obscuration methods, such as blurring and pixelation is presented. The analysis is based on various threat models and obscuration attackers to achieve a comprehensive evaluation. We show that the traditional blurring and pixelation methods cannot guarantee privacy. To provide a more secured privacy protection, we propose two novel obscuration methods that are based on the generative adversarial networks. The proposed methods not only remove the identifiable information, but also preserve the non-identifiable facial information (as known as the utility information), such as expression, age, skin tone and gender.</p><p>We also propose methods for change detection in satellite imagery. In this thesis, we consider two types of building changes: 2D appearance change and 3D height change. We first present a model with an attention mechanism to detect the building appearance changes that are caused by natural disasters. Furthermore, to detect the changes of building height, we present a height estimation model that is based on building shadows and solar angles without relying on height annotation. Both change detection methods require good building segmentation performance, which might be hard to achieve for the low-quality images, such as off-nadir images. To solve this issue, we use uncertainty modeling and satellite imagery metadata to achieve accurate building segmentation for the noisy images that are taken from large off-nadir angles.</p></div></div></div>

One of the things augmented reality depends on is object tracking, which is a problem classically found in cinematography and security. However, the algorithms designed for the classical application are often too expensive computationally or too complex to run on simpler mobile hardware. One of the methods to do object tracking is with a trained neural network, this has already led to great results but is unfortunately still running into some of the same problems as the classical algorithms. For this reason a neural network designed specifically for object tracking on mobile hardware needs to be developed. This thesis will propose two di erent neural networks designed for object tracking on mobile hardware. Both are based on a siamese network structure and methods to improve their accuracy using filtering are also introduced. The first network is a modified version of “CNN architecture for geometric matching” that utilizes an a ne regression to perform object tracking. This network was shown to underperform in the MOT benchmark as-well as the VOT benchmark and therefore not further developed. The second network is an object detector based on “SqueezeDet” in a siamese network structure utilizing the performance optimized layers of “MobileNets”. The accuracy of the object detector network is shown to be competitive in the VOT benchmark, placing at the 16th place compared to trackers from the 2016 challenge. It was also shown to run in real-time on mobile hardware. Thus the one shot object detection network used for a tracking application can improve the experience of augmented reality applications on mobile hardware.

Change detection using satellite imagery has applications in urban development, disaster response and precision agriculture. Current deep learning models show promising results. However, on-board computers are typically highly constrained which poses a challenge for deployment. On-board processing is desirable for saving bandwidth by downlinking only novel and valuable data. The goal of this work is to determine what change detection models are most technically feasible for on-board use in satellites. The novel patch based model MobileGoNogo is evaluated along current state-of-the-art models. Technical feasibility was determined by observing accuracy, inference time, storage buildup, memory usage and resolution on a satellite computer tasked with detecting changes in buildings from the SpaceNet 7 dataset. Three high level approaches were taken; direct classification, post classification and patch-based change detection. None of the models compared in the study fulfilled all requirements for general technical feasibility. Direct classification models were highly resource intensive and slow. Post classification model had critically low accuracy but desirable storage characteristics. Patch based MobileGoNogo performed better by all metrics except in resolution where it is significantly lower than any other model. We conclude that the novel model offers a feasible solution for low resolution, noncritical applications. / Upptäckt av förändringar med hjälp av satellitbilder har tillämpningar inom bl.a. stadsutveckling, katastrofinsatser och precisionsjordbruk. De nuvarande modellerna för djupinlärning visar lovande resultat. Datorerna ombord satelliter är dock vanligtvis mycket begränsade, vilket innebär en utmaning för användningen av dessa modeller. Databehandling ombord är önskvärd för att spara bandbredd genom att endast skicka ner nya och värdefulla data. Målet med detta arbete är att fastställa vilka modeller för upptäckt av förändringar som är mest tekniskt genomförbara för användning ombord på satelliter. Den nya bildfältbaserade modellen MobileGoNogo utvärderas tillsammans med de senaste modellerna. Den tekniska genomförbarheten fastställdes genom att observera träffsäkerhet, inferenstid, lagring, minnesanvändning och upplösning på en satellitdator med uppgift att upptäcka förändringar i byggnader från SpaceNet 7dataset. Tre tillvägagångssätt på hög nivå användes: direkt klassificering, postklassificering och fältbaserad klassificering. Ingen av de modeller som jämfördes i studien uppfyllde alla krav på allmän teknisk genomförbarhet. Direkta klassificeringsmodeller var mycket resurskrävande och långsamma. Postklassificeringsmodellen hade kritiskt låg träffsäkerhet men önskvärda lagringsegenskaper. Den bildfältbaserade MobileGoNogo-modellen var bättre i alla mätvärden utom i upplösningen, där den var betydligt lägre än någon annan modell. Vi drar slutsatsen att den nya modellen erbjuder en genomförbar lösning för icke-kritiska tillämpningar med låg upplösning.

Laser powder bed fusion is an additive manufacturing technique that is capable of building metallic parts by spreading many layers of metal powder over a build surface and using a laser to melt specific sections of the surface. The part is built by melting consecutive layers on top of each other until the design is completed. However, during this process defects can occur. These defects have impacts on the part’s physical properties, and it is important to detect them for quality assurance. A single part takes several hundred or thousands of layers to build. While each layer is built, cameras and sensors are used to create images of each layer. These images are used for identification and classification of defects that could have a negative impact on a printed part’s physical properties, such as tensile strength. Classification of defects would reduce manual inspection of the printed part. Thus, the classification of defects in each layer must be automated, as it would be infeasible to manually classify each layer. Recently, machine learning have proven to be an effective method for automating defect classification in laser powder bed fusion. However, machine learning and especially deep-learning approaches generally require a large amount of labeled training data, which is typically not available for laser powder bed fusion printed parts. Labeling of images requires manual labor and domain knowledge. One of the greatest obstacles in defect classification, is how machine learning can be applied despite this absence of labeled data. A machine learning approach that show potential for being trained with less data, is the siamese neural network approach. In this thesis, a novel approach for automating defect classification is developed, using layer images from a laser powder bed fusion printing process. In order to cope with the limited access to labeled data, the classifiers are based on the siamese neural network structure. Two siamese neural network structures are developed, a one-shot classifier, which directly classifies the instance, and a hierarchical classifier with a hierarchical classification process according to the hierarchy of the defect classes. The classifiers are evaluated by inferring a test set of images collected from the laser powder bed fusion process. The one-shot classifier is able to classify the images with an accuracy of 70%and the hierarchical classifier with an accuracy of 86%. For the hierarchical classifier area of the ROC curves were calculated to be, 0.96 and 0.95 for the normal vs defect and overheating vs spattering stages respectively. Unlabeled images were added to the training set of a new instance of the hierarchical classifier, which could infer the test set without any major changes to test set accuracy.

Cette étude vise à développer un système d’aide au diagnostic (CAD) pour la détection de lésions épileptogènes, reposant sur l’analyse de données de neuroimagerie, notamment, l’IRM T1 et FLAIR. L’approche adoptée, introduite précédemment par Azami et al., 2016, consiste à placer la tâche de détection dans le cadre de la détection de changement à l'échelle du voxel, basée sur l’apprentissage d’un modèle one-class SVM pour chaque voxel dans le cerveau. L'objectif principal de ce travail est de développer des mécanismes d’apprentissage de représentations, qui capturent les informations les plus discriminantes à partir de l’imagerie multimodale. Les caractéristiques manuelles ne sont pas forcément les plus pertinentes pour la tâche visée. Notre première contribution porte sur l'intégration de différents réseaux profonds non-supervisés, pour extraire des caractéristiques dans le cadre du problème de détection de changement. Nous introduisons une nouvelle configuration des réseaux siamois, mieux adaptée à ce contexte. Le système CAD proposé a été évalué sur l’ensemble d’images IRM T1 des patients atteints d'épilepsie. Afin d'améliorer la performance obtenue, nous avons proposé d'étendre le système pour intégrer des données multimodales qui possèdent des informations complémentaires sur la pathologie. Notre deuxième contribution consiste donc à proposer des stratégies de combinaison des différentes modalités d’imagerie dans un système pour la détection de changement. Ce système multimodal a montré une amélioration importante sur la tâche de détection de lésions épileptogènes sur les IRM T1 et FLAIR. Notre dernière contribution se focalise sur l'intégration des données TEP dans le système proposé. Etant donné le nombre limité des images TEP, nous envisageons de synthétiser les données manquantes à partir des images IRM disponibles. Nous démontrons que le système entraîné sur les données réelles et synthétiques présente une amélioration importante par rapport au système entraîné sur les images réelles uniquement. / This work represents one attempt to develop a computer aided diagnosis system for epilepsy lesion detection based on neuroimaging data, in particular T1-weighted and FLAIR MR sequences. Given the complexity of the task and the lack of a representative voxel-level labeled data set, the adopted approach, first introduced in Azami et al., 2016, consists in casting the lesion detection task as a per-voxel outlier detection problem. The system is based on training a one-class SVM model for each voxel in the brain on a set of healthy controls, so as to model the normality of the voxel. The main focus of this work is to design representation learning mechanisms, capturing the most discriminant information from multimodality imaging. Manual features, designed to mimic the characteristics of certain epilepsy lesions, such as focal cortical dysplasia (FCD), on neuroimaging data, are tailored to individual pathologies and cannot discriminate a large range of epilepsy lesions. Such features reflect the known characteristics of lesion appearance; however, they might not be the most optimal ones for the task at hand. Our first contribution consists in proposing various unsupervised neural architectures as potential feature extracting mechanisms and, eventually, introducing a novel configuration of siamese networks, to be plugged into the outlier detection context. The proposed system, evaluated on a set of T1-weighted MRIs of epilepsy patients, showed a promising performance but a room for improvement as well. To this end, we considered extending the CAD system so as to accommodate multimodality data which offers complementary information on the problem at hand. Our second contribution, therefore, consists in proposing strategies to combine representations of different imaging modalities into a single framework for anomaly detection. The extended system showed a significant improvement on the task of epilepsy lesion detection on T1-weighted and FLAIR MR images. Our last contribution focuses on the integration of PET data into the system. Given the small number of available PET images, we make an attempt to synthesize PET data from the corresponding MRI acquisitions. Eventually we show an improved performance of the system when trained on the mixture of synthesized and real images.