Deep Neural Networks are widely applied to various kinds of fields nowadays. However, hundreds of thousands of neurons in each layer result in intensive memory storage requirement and a massive number of operations, making it difficult to employ deep neural networks on mobile devices where the hardware resources are limited. One common technique to address the memory limitation is to prune and quantize the neural networks. Besides, due to the frequent usage of Rectified Linear Unit (ReLU) function or network pruning, majority of the data in the weight matrices will be zeros, which will not only take up a large amount of memory space but also cause unnecessary computation operations. In this thesis, a new value-based compression method is put forward to represent sparse matrix more efficiently by eliminating these zero elements, and a customized hardware is implemented to realize the decompression and computation operations. The value-based compression method is aimed to replace the nonzero data in each column of the weight matrix with a reference value (arithmetic mean) and the relative differences between each nonzero element and the reference value. Intuitively, the data stored in each column is likely to contain similar values. Therefore, the differences will have a narrow range, and fewer bits rather than the full form will be sufficient to represent all the differences. In this way, the weight matrix can be further compressed to save memory space. The proposed value-based compression method reduces the memory storage requirement for the fully-connected layers of AlexNet to 37%, 41%, 47% and 68% of the compressed model, e.g., the Compressed Sparse Column (CSC) format, when the data size is set to 8 bits and the sparsity is 20%, 40%, 60% and 80% respectively. In the meanwhile, 41%, 53% and 63% compression rates of the fully-connected layers of the compressed AlexNet model with respect to 8-bit, 16-bit and 32-bit data are achieved when the sparsity is 40%. Similar results are obtained for VGG16 experiment. / Djupa neurala nätverk används i stor utsträckning inom olika fält nuförtiden. Emellertid ställer hundratusentals neuroner per lager krav på intensiv minneslagring och ett stort antal operationer, vilket gör det svårt att använda djupa neurala nätverk på mobila enheter där hårdvaruresurserna är begränsade. En vanlig teknik för att hantera minnesbegränsningen är att beskära och kvantifiera de neurala nätverken. På grund av den frekventa användningen av Rectified Linear Unit (ReLU) -funktionen eller nätverksbeskärning kommer majoriteten av datat i viktmatriserna att vara nollor, vilket inte bara tar upp mycket minnesutrymme utan också orsakar onödiga beräkningsoperationer. I denna avhandling presenteras en ny värdebaserad komprimeringsmetod för att representera den glesa matrisen mer effektivt genom att eliminera dessa nollelement, och en anpassad hårdvara implementeras för att realisera dekompressions- och beräkningsoperationerna. Den värdebaserade komprimeringsmetoden syftar till att ersätta icke-nolldata i varje kolumn i viktmatrisen med ett referensvärde (aritmetiskt medelvärde) och de relativa skillnaderna mellan varje icke-nollelement och referensvärdet. Intuitivt kommer data som lagras i varje kolumn sannolikt att innehålla liknande värden. Därför kommer skillnaderna att ha ett smalt intervall, och färre bitar snarare än den fullständiga formen kommer att räcka för att representera alla skillnader. På så sätt kan viktmatrisen komprimeras ytterligare för att spara minnesutrymme. Den föreslagna värdebaserade komprimeringsmetoden minskar minneslagringskravet för de helt anslutna lagren av AlexNet till 37%, 41%, 47% och 68% av den komprimerade modellen, t.ex. Compressed Sparse Column (CSC) format, när datastorleken är inställd på 8 bitar och sparsiteten är 20%, 40%, 60% respektive 80%. Under tiden uppnås 41%, 53% och 63% komprimeringshastigheter för de helt anslutna lagren i den komprimerade AlexNet-modellen med avseende på 8- bitars, 16-bitars och 32-bitars data när sparsiteten är 40%. Liknande resultat erhålls för VGG16-experiment.

Covid-19 pandemin är den mest omfattande pandemin i modern tid. Länder världen över har infört olika typer av krav och rekommendationer för att hämma smittspridningen, däribland rekommendationer om att bära munskydd. Tidigare studier har utvecklat modeller och system för att avgöra huruvida personer bär munskydd eller ej. Syftet med denna studie är att bevisa att det är tekniskt genomförbart att utveckla och implementera en IT-artefakt på en Google Coral Dev Board som i realtid inte bara kan avgöra om en person bär munskydd eller ej, utan även kan klassificera om ett munskydd bärs korrekt samt ge återkoppling om eventuell felanvändning. Ett sådant system skulle kunna användas för att uppmuntra och påminna om korrekt användning av munskydd. Det utvecklade systemet bygger på en två-stegs-arkitektur bestående av två stycken Convolutional Neural Networks (CNN), en förtränad ansiktsdetekteringsmodell samt en egenutvecklad klassificeringsmodell som bygger på en MobileNetV3-arkitektur. Klassificeringsmodellen tränades med hjälp av ett dataset bestående av 184 572 bilder och kategoriserar ansikten i fyra olika kategorier: munskydd används korrekt, munskydd täcker ej näsa, munskydd placerat på haka och munskydd används ej. Vid 5-delad korsvalidering uppnådde klassificeringsmodellen en genomsnittlig accuracy på >0.9994. Resultaten för studien var goda och bevisar den tekniska genomförbarheten. Den utvecklade prototypen kunde med hög tillförlitlighet detektera och korrekt klassificera ansikten i en videoström. Dock kommer studien fram till att denna typ av två-stegs-arkitektur är mindre lämplig i situationer då ett större antal ansikten är synliga i bild samtidigt då tiden som krävs för att behandla en bild ökar när antalet ansikten blir fler, vilket innebär att systemet ej kommer upplevas som responsivt.

In the current deep learning era, convolutional neural networks are commonly used as a backbone of systems that process images or videos. A lot of existing neural network architectures are however needlessly overparameterized and their performance can be closely matched by an alternative that uses much smaller amount of parameters. Our aim is to design a method that is able to find such alternative(s) for a given convolutional architecture. We propose a general scheme for architecture reduction and evaluate three algorithms that search for the op- timal smaller architecture. We do multiple experiments with ResNet and Wide ResNet architectures as the base using CIFAR-10 dataset. The best method is able to reduce the number of parameters by 75-85% without any loss in accuracy even in these already quite efficient architectures. 1

Smart wearables, known as emerging paradigms for vital big data capturing, have been attracting intensive attentions. However, one crucial problem is their power-hungriness, i.e., the continuous data streaming consumes energy dramatically and requires devices to be frequently charged. Targeting this obstacle, we propose to investigate the biodynamic patterns in the data and design a data-driven approach for intelligent data compression. We leverage Deep Learning (DL), more specifically, Convolutional Autoencoder (CAE), to learn a sparse representation of the vital big data. The minimized energy need, even taking into consideration the CAE-induced overhead, is tremendously lower than the original energy need. Further, compared with state-of-the-art wavelet compression-based method, our method can compress the data with a dramatically lower error for a similar energy budget. Our experiments and the validated approach are expected to boost the energy efficiency of wearables, and thus greatly advance ubiquitous big data applications in era of smart health.<br><div>In recent years, there has also been a growing interest in edge intelligence for emerging instantaneous big data inference. However, the inference algorithms, especially deep learning, usually require heavy computation requirements, thereby greatly limiting their deployment on the edge. We take special interest in the smart health wearable big data mining and inference. <br></div><div><br></div><div>Targeting the deep learning’s high computational complexity and large memory and energy requirements, new approaches are urged to make the deep learning algorithms ultra-efficient for wearable big data analysis. We propose to leverage knowledge distillation to achieve an ultra-efficient edge-deployable deep learning model. More specifically, through transferring the knowledge from a teacher model to the on-edge student model, the soft target distribution of the teacher model can be effectively learned by the student model. Besides, we propose to further introduce adversarial robustness to the student model, by stimulating the student model to correctly identify inputs that have adversarial perturbation. Experiments demonstrate that the knowledge distillation student model has comparable performance to the heavy teacher model but owns a substantially smaller model size. With adversarial learning, the student model has effectively preserved its robustness. In such a way, we have demonstrated the framework with knowledge distillation and adversarial learning can, not only advance ultra-efficient edge inference, but also preserve the robustness facing the perturbed input.</div>

Context : Air being the free source is used in different ways commercially. In earlier days windmills generate power, water, and electricity. The excessive establishment of windmills for commercial purposes affected avifauna. Most of the birds lost their lives due to collisions with windmills. Turbines used to generate power near airports are also one of the causes for the extinction of birdlife. According to a survey in 2011 in Canada a total of 23,300 bird deaths were caused by wind turbines and also it is estimated that the number of deaths would increase to 2,33,000 in the following 10-15 years. Objectives : The main objective of this thesis is to find a suitable software solution to detect the birds on a series of grayscale images in real-time and in minimum full HD resolution with at least a 15 FPS rate. User-Driven Design Methodology is used for developing, tools are Python and Open-CV. Methods : In this research, a system is designed to detect the bird in an HD Video. Possible methods that can be considered are convolutional neural networks (CNN), vision based detection with background subtraction, contour detection and confusion matrix classification. These methods detect birds in raw images and with help of a classifier make it possible to see the bird in desired pixels with full resolution. We will investigate a bird classification method consisting of two steps, background subtraction, and then object classification. Background subtraction is a fundamental method to extract moving objects from a fixed background. For classification, we will use the YOLO v3 model version for object classification. Results : The project is expected to result in a system design and prototype for the bird identification on a grayscale video stream in at least full HD resolution in a minimum of 15 FPS. The bird should be distinguished from other moving objects like wind turbine blades, trees, or clouds. The proposed solution should identify up to 5 birds simultaneously. Conclusion : After completing each step and arriving at the classification, the methods we have tried, such as Haar Cascades and mobile-net SSD, were not providing us with the desired results. So we opted to use YOLO v3, which had the best accuracy in classifying different objects. By using the YOLO v3 classifier, we have detected the bird with 95% accuracy, blades with 90% accuracy, clouds with 80% accuracy, trees with 70% accuracy. Moreover, we conclude that there is a need for further empirical validation of the models in full-scale industry trials.

The design and adjustment of convolutional neural network architectures is an opaque and mostly trial and error-driven process. The main reason for this is the lack of proper paradigms beyond general conventions for the development of neural networks architectures and lacking effective insights into the models that can be propagated back to design decision. In order for the task-specific design of deep learning solutions to become more efficient and goal-oriented, novel design strategies need to be developed that are founded on an understanding of convolutional neural network models. This work develops tools for the analysis of the inference process in trained neural network models. Based on these tools, characteristics of convolutional neural network models are identified that can be linked to inefficiencies in predictive and computational performance. Based on these insights, this work presents methods for effectively diagnosing these design faults before and during training with little computational overhead. These findings are empirically tested and demonstrated on architectures with sequential and multi-pathway structures, covering all the common types of convolutional neural network architectures used for classification. Furthermore, this work proposes simple optimization strategies that allow for goal-oriented and informed adjustment of the neural architecture, opening the potential for a less trial-and-error-driven design process.

Microscopy imaging based techniques, such as the Cell Painting assay, could be used to generate images that visualize the Mechanism of Action (MoA) of a drug, which could be of great use in drug development. In order to extract information and predict the MoA of a new compound from these images we need powerful image analysis tools. The purpose with this project is to further develop a Deep Learning model to predict the MoA of different drugs from Cell Painting images using Convolutional Neural Networks (CNNs) and Conformal Prediction. The specific task was to compare the accuracy of different CNN architectures and to compare the efficiency of different nonconformity functions.  During the project the CNN architectures ResNet50, ResNet101 and DenseNet121 were compared as well as the nonconformity functions Inverse Probability, Margin and a combination of them both. No significant difference in accuracy between the CNNs and no difference in efficiency between the nonconformity functions was measured. The results showed that the model could predict the MoA of a compound with high accuracy when all compounds were used both in training, validation and test of the model, which validates the implementations. However, it is desirable for the model to be able to predict the MoA of a new compound if the model has been trained on other compounds with the same MoA. This could not be confirmed through this project and the model needs to be further investigated and tested with another dataset in order to be used for that purpose.

Aversion on construction sites can be everything from missingmaterial, fire hazards, or insufficient cleaning. These aversionsappear very often on construction sites and the construction companyneeds to report and take care of them in order for the site to runcorrectly. The reports consist of an image of the aversion and atext describing the aversion. Report categorization is currentlydone manually which is both time and cost-ineffective. The task for this thesis was to implement and evaluate an automaticmultimodal machine learning classifier for the reported aversionsthat utilized both the image and text data from the reports. Themodel presented is a late-fusion model consisting of a Swedish BERTtext classifier and a VGG16 for image classification. The results showed that an automated classifier is feasible for thistask and could be used in real life to make the classification taskmore time and cost-efficient. The model scored a 66.2% accuracy and89.7% top-5 accuracy on the task and the experiments revealed someareas of improvement on the data and model that could be furtherexplored to potentially improve the performance.

No description available.

No description available.