In this thesis the NIN artificial neural network structure created by Min Lin et al in 2014 is investigated. This is done by varying stage numbers and layer depth. By doing this ten different networks including the original NIN were created. Testing is carried out on a preprocessed version of the CIFAR10 dataset for these ten networks for a maximum of 150’000 iterations. The results show that the number of stages generally affect NIN performance more than layer depth does. The network with three stages and a layer depth of two performs best at a top accuracy of 87.44%. This is below Min Lin et al’s results. However, this is likely due to overfitting and lack of specifics on their training methods. The thesis concludes that studies of different types of micro-networks in the NIN are required. Studies are also required on deeper NINs with larger datasets to prevent the overfitting observed in the results. Larger datasets could be obtained by data augmentation. Furthermore, the results suggests that less complicated (by less complicated it is meant that the network stages have less depth) NIN implementations are more accurate than deeper ones. / I detta kandidatexamensarbete unders¨oks den artificiella neuronn¨at-strukturen NIN som skapades av Min Lin et al 2014. Detta g¨ors genom att variera antalet moduler och lagerdjupet i dessa. Detta ger tio olika n¨atverk inklusive det originala NIN-n¨atet. Testerna utf¨ors p˚a en f¨orprocesserad version av CIFAR10-databasen f¨or alla tio n¨atverk i maximalt 150’000 iterationer. Resultaten visar p˚a att modulantalet generellt p˚averkar NINs prestanda mer ¨an lagerdjupet. N¨atverket med tre moduler och ett lagerdjup p˚a tv˚a har h¨ogst prestanda med 87.44% r¨att. Detta ¨ar under Min Lin et als resultat. Detta beror dock troligen p˚a overfitting och en brist p˚a n¨armare detaljer r¨orande tr¨aningsmetoderna de anv¨ant. Slutsatsen av arbetet ¨ar att studier p˚a olika typer av mikron¨atverk i NIN beh¨ovs. Studier p˚a djupare NINn¨atverk med st¨orre dataset f¨or att f¨orhindra den overfitting som syns i resultaten beh¨ovs ¨aven. St¨orre dataset skulle kunna erh˚allas genom data-augmentering. Resultaten verkar ¨aven antyda att mindre komplicerade (med mindre komplicerade menas att modulerna har mindre djup) NIN-implementationer har h¨ogre prestada ¨an djupa s˚adana.

This work researches named entity recognition (NER) with respect to images of documents with a domain-specific layout, by means of Convolutional Neural Networks (CNNs). Examples of such documents are receipts, invoices, forms and scientific papers, the latter of which are used in this work. An NER task is first performed statically, where a static number of entity classes is extracted per document. Networks based on the deep VGG-16 network are used for this task. Here, experimental evaluation shows that framing the task as a classification task, where the network classifies each bounding box coordinate separately, leads to the best network performance. Also, a multi-headed architecture is introduced, where the network has an independent fully-connected classification head per entity. VGG-16 achieves better performance with the multi-headed architecture than with its default, single-headed architecture. Additionally, it is shown that transfer learning does not improve performance of these networks. Analysis suggests that the networks trained for the static NER task learn to recognise document templates, rather than the entities themselves, and therefore do not generalize well to new, unseen templates. For a dynamic NER task, where the type and number of entity classes vary per document, experimental evaluation shows that, on large entities in the document, the Faster R-CNN object detection framework achieves comparable performance to the networks trained on the static task. Analysis suggests that Faster R-CNN generalizes better to new templates than the networks trained for the static task, as Faster R-CNN is trained on local features rather than the full document template. Finally, analysis shows that Faster R-CNN performs poorly on small entities in the image and suggestions are made to improve its performance.

Machine learning methods provide an opportunity to improve the classification of skin lesions and the early diagnosis of melanoma by providing decision support for general practitioners. So far most studies have been looking at the creation of features that best indicate melanoma. Representation learning methods such as neural networks have outperformed hand-crafted features in many areas. This work aims to evaluate the performance of convolutional neural networks in relation to earlier machine learning algorithms and expert diagnosis. In this work, convolutional neural networks were trained on datasets of dermoscopy images using weights initialized from a random distribution, a network trained on the ImageNet dataset and a network trained on Dermnet, a skin disease atlas.  The ensemble sum prediction of the networks achieved an accuracy of 89.3% with a sensitivity of 77.1% and a specificity of 93.0% when based on the weights learned from the ImageNet dataset and the Dermnet skin disease atlas and trained on non-polarized light dermoscopy images.  The results from the different networks trained on little or no prior data confirms the idea that certain features are transferable between different data. Similar classification accuracies to that of the highest scoring network are achieved by expert dermatologists and slightly higher results are achieved by referenced hand-crafted classifiers.  The trained networks are found to be comparable to practicing dermatologists and state-of-the-art machine learning methods in binary classification accuracy, benign – melanoma, with only little pre-processing and tuning.

Bokeh is defined as a soft out of focus blur. An image with bokeh has a subject in focus and an artistically blurry background. To capture images with real bokeh, specific camera parameter choices need to be made. One essential choice is to use a big lens with a wide aperture. Because of smartphone cameras’ small size, it becomes impossible to achieve real bokeh. Commonly, new models of smartphones have artificial bokeh implemented when taking pictures, but it is uncommon to be able to capture videos with artificial bokeh. Video segmentation is more complicated than image segmentation because it puts a higher demand on performance. The result should also be temporally consistent. In this project, the aim is to create a method that can apply real-time video bokeh on a smartphone.   The project consists of two parts. The first part is to segment the subject of the video. This process is performed with convolutional neural networks. Three image segmentation networks were implemented for video, trained, and evaluated. The model that illustrated the most potential was the SINet model and was chosen as the most suitable architecture for the task.  The second part of the project is to manipulate the background to be aesthetically pleasing while at the same time mimicking real optics to some degree. This is achieved by creating a depth and contrast map. With the depth map, the background can be blurred based on depth. The shape of the bokeh shapes also varies with the depth. The contrast map is used to locate bokeh points. The main part of the project is the segmentation part.  The result for this project is a method that achieves an accurate segmentation and creates an artistic background. The different architectures illustrated similar results in terms of accuracy but different in terms of inference time. Situations existed where the segmentation failed and included too much of the background. This could potentially be counteracted with a bigger and more varied dataset. The method is performed in real-time on a computer but no conclusions could be made if it works in real-time on a smartphone.

This paper investigates the use of deep Convolutional Neural Networks for modeling the intronic regulation of Alternative Splicing on the basis of DNA sequence. By training the CNN on massively parallel synthetic DNA libraries of Alternative 5'-splicing and Alternatively Skipped exon events, the model is capable of predicting the relative abundance of alternatively spliced mRNA isoforms on held-out library data to a very high accuracy (R2 = 0.77 for Alt. 5'-splicing). Furthermore, the CNN is shown to generalize alternative splicing across cell lines efficiently. The Convolutional Neural Net is tested against a Logistic regression model and the results show that while prediction accuracy on the synthetic library is notably higher compared to the LR model, the CNN is worse at generalizing to new intronic contexts. Tests on non-synthetic human SNP genes suggest the CNN is dependent on the relative position of the intronic region it was trained for, a problem which is alleviated with LR. The increased library prediction accuracy of the CNN compared to Logistic regression is concluded to come from the non-linearity introduced by the deep layer architecture. It adds the capacity to model complex regulatory interactions and combinatorial RBP effects which studies have shown largely affect alternative splicing. However, the architecture makes interpreting the CNN hard, as the regulatory interactions are encoded deep within the layers. Nevertheless, high-performance modeling of alternative splicing using CNNs may still prove useful in numerous Synthetic biology applications, for example to model differentially spliced genes as is done in this paper. / Den här uppsatsen undersöker hur djupa neurala nätverk baserade på faltning ("Convolutions") kan användas för att modellera den introniska regleringen av Alternativ Splicing med endast DNA-sekvensen som indata. Nätverket tränas på ett massivt parallelt bibliotek av syntetiskt DNA innehållandes Alternativa Splicing-event där delar av de introniska regionerna har randomiserats. Uppsatsen visar att nätverksarkitekturen kan förutspå den relativa mängden alternativt splicat RNA till en mycket hög noggrannhet inom det syntetiska biblioteket. Modellen generaliserar även alternativ splicing mellan mänskliga celltyper väl. Hursomhelst, tester på icke-syntetiska mänskliga gener med SNP-mutationer visar att nätverkets prestanda försämras när den introniska region som används som indata flyttas i jämförelse till den relativa position som modellen tränats på. Uppsatsen jämför modellen med Logistic regression och drar slutsatsen att nätverkets förbättrade prestanda grundar sig i dess förmåga att modellera icke-linjära beroenden i datan. Detta medför dock svårigheter i att tolka vad modellen faktiskt lärt sig, eftersom interaktionen mellan reglerande element är inbäddat i nätverkslagren. Trots det kan högpresterande modellering av alternativ splicing med hjälp av neurala nät vara användbart, exempelvis inom Syntetisk biologi där modellen kan användas för att kontrollera regleringen av splicing när man konstruerar syntetiska gener.

Manual segmentation of brain tumours is a time consuming process, results often show high variability, and there is a call for automation in clinical practice. In this thesis the use of deep convolutional networks for automatic glioma segmentation in MRI is investigated. The implemented networks are evaluated on data used in the brain tumor segmentation challenge (BraTS). It is found that 3D convolutional networks generally outperform 2D convolutional networks, and that the best networks can produce segmentations that closely resemble human segmentations. Convolutional networks are also evaluated as feature extractors with linear SVM classifiers on top, and although the sensitivity is improved considerably, the segmentations are heavily oversegmented. The importance of the amount of data available is investigated as well by comparing results from networks trained on both 2013 and the greatly extended 2014 data set, but it is found that the method of producing ground-truth was also a contributing factor. The networks does not beat the previous high-scores on the BraTS data, but several simple improvement areas are identified to take the networks further. / Manuell segmentering av hjärntumörer är en tidskrävande process, segmenteringarna är ofta varierade mellan experter, och automatisk segmentering skulle vara användbart för kliniskt bruk. Den här rapporten undersöker användningen av deep convolutional networks (ConvNets) för automatisk segmentering av gliom i MR-bilder. De implementerade nätverken utvärderas med hjälp av data från brain tumor segmentation challenge (BraTS). Studien finner att 3D-nätverk har generellt bättre resultat än 2D-nätverk, och att de bästa nätverken har förmågan att ge segmenteringar som liknar mänskliga segmenteringar. ConvNets utvärderas också som feature extractors, med linjära SVM som klassificerare. Den här metoden ger segmenteringar med hög känslighet, men är också till hög grad översegmenterade. Vikten av att ha mer träningsdata undersöks också genom att träna på två olika stora dataset, men metoden för att få fram de riktiga segmenteringarna har troligen också stor påverkan på resultatet. Nätverken slår inte de tidigare rekorden på BraTS, men flera viktiga men enkla förbättringsområden är identifierade som potentiellt skulle förbättra resultaten.

We investigated the advantage of combining RGB images with depth data to get more robust object classifications and detections using pre-trained deep convolutional neural networks. We relied upon the raw images from publicly available datasets captured using Microsoft Kinect cameras. The raw images varied in size, and therefore required resizing to fit our network. We designed a resizing method called "bleeding edge" to avoid distorting the objects in the images. We present a novel method of interpolating the missing depth pixel values by comparing to similar RGB values. This method proved superior to the other methods tested. We showed that a simple colormap transformation of the depth image can provide close to state-of-art performance. Using our methods, we can present state-of-art performance on the Washington Object dataset and we provide some results on the Washington Scenes (V1) dataset. Specifically, for the detection, we used contours at different thresholds to find the likely object locations in the images. For the classification task we can report state-of-art results using only RGB and RGB-D images, depth data alone gave close to state-of-art results. For the detection task we found the RGB only detector to be superior to the other detectors.

This thesis examines how convolutional neural networks can applied to the problem of hand detection and hand pose estimation. Two families of convolutional neural networks are trained, aimed at performing the task of classification or regression. The networks are trained on specialized data generated from publicly available datasets. The algorithms used to generate the specialized data are also disclosed. The main focus has been to investigate the different structural properties of convolutional neural networks, not building optimized hand detection, or hand pose estimation, systems. Experiments revealed, that classifier networks featuring a relatively high number of convolutions offers the highest performance on external validation data. Additionally, shallow classifier networks featuring a relatively low number of convolutions, yields a high classification accuracy on training and testing data, but a very low accuracy on the validation set. This effect uncovers one of the fundamental difficulties in building a hand detection system: The asymmetric classification problem. In further investigation, it is also remarked, that relatively shallow classifier networks probably becomes color sensitive. Furthermore, regressor networks featuring multiscale inputs typically yielded the lowest error, when tasked with computing key-point locations directly from data. It is also revealed, that color data implicitly contain more information, making it easier to compute key-point locations, especially in the image space. However, to be able to derive the color invariant features, deeper regressor networks are required. / I detta examensarbete undersöks hur faltande neuronnät kan användas för detektering av, samt skattning av pose hos, händer. Två familjer av neuronnät tränas, med syftet att utföra klassificering eller regression. Neuronnäten tränas med specialiserad data genererad ur publikt tillgängliga dataset. Algoritmerna för att generera den specialiserade datan presenteras även i sin helhet. Huvudsyftet med arbetet, har varit att undersöka neuronnätens strukturella egenskaper, samt relatera dessa till prestanda, och inte bygga ett färdigt system för handdetektering eller skattning av handpose. Experimenten visade, att neuronnät för klassificering med ett relativt stor antal faltningar ger högst prestanda på valideringsdata. Vidare, så verkar neuronnät för klassificering med relativt litet antal faltningar ge en god prestanda på träning- och testdata, men mycket dålig prestand på valideringsdata. Detta sambandet avslöjar en fundamental svårighet med att träna ett neuronnät för klassificering av händer, nämligen det kraftigt asymmetriska klassificeringsproblemet. I vidare undersökningar visar det sig också, att neuronnät för klassificering med ett relativt litet antal faltningar troligtvis enbart blir färgkänsliga. Experimenten visade också, att neuronnät för regression som använde sig av data i flera skalor gav lägst fel när de skulle beräkna positioner av handmarkörer direkt ur data. Slutligen framkom det, att färgdata, i konstrast till djupdata, implicit innehåller mer information, vilket gör det relativt sett lättare att beräkna markörer, framför allt i det tvådimensionella bildrummet. Dock, för att kunna få fram den implicita informationen, så krävs relativt djupa neuronnät.

Research on brain-computer interfaces (BCIs) has been around for decades and recently the inner speech paradigm was picked up in the area. The realization of a functioning BCI could improve the life quality of many people, especially persons affected by Locked-In-Syndrome or similar illnesses. Although implementing a working BCI is too large of a commitment for a master's thesis, this thesis will focus on investigating machine learning methods to decode inner speech using data collected from the non-invasive and portable method electroencephalography (EEG). Among the methods investigated are three CNN architectures and transfer learning. The results show that the EEGNet architecture consistently reaches high classification accuracies, with the best model achieving an accuracy of 29.05%.

Machine learning techniques are being applied to many data-intensive problems because they can accurately provide classification of complex data using appropriate training. Often, the performance of machine learning can exceed the performance of traditional techniques because machine learning can take advantage of higher dimensionality than traditional algorithms. In this work, acoustic data sets taken using a rapid scanning technique on concrete bridge decks provided an opportunity to both apply machine learning algorithms to improve detection performance and also to investigate the ways that training of neural networks can be aided by data augmentation approaches. Early detection and repair can enhance safety and performance as well as reduce long-term maintenance costs of concrete bridges. In order to inspect for non-visible internal cracking (called delaminations) of concrete bridges, a rapid inspection method is needed. A six-channel acoustic impact-echo sounding apparatus is used to generate large acoustic data sets on concrete bridge decks at high speeds. A machine learning data processing architecture is described to accurately detect and map delaminations based on the acoustic responses. The machine learning approach achieves accurate results at speeds between 25 and 45 km/h across a bridge deck and successfully demonstrates the use of neural networks to analyze this type of acoustic data. In order to obtain excellent performance, model training generally requires large data sets. However, in many potentially interesting cases, such as bridge deck defect detection, acquiring enough data for training can be difficult. Data augmentation can be used to increase the effective size of the training data set. Acoustic signal data augmentation is demonstrated in conjunction with a machine learning model for acoustic defect detection on bridge decks. Four different augmentation methods are applied to data using two different augmentation strategies. This work demonstrates that a "goldilocks" data augmentation approach can be used to increase machine learning performance when only a limited data set is available. The major technical contributions of this work include application of machine learning to acoustic data sets relevant to bridge deck inspection, solving an important problem in the field of nondestructive evaluation, and a more generalized approach to data augmentation of limited acoustic data sets to expand the classes of acoustic problems that machine learning can successfully address.