This thesis tackles the problem of understanding deep neural network with piece- wise linear activation functions. We leverage tropical geometry, a relatively new field in algebraic geometry to characterize the decision boundaries of a single hidden layer neural network. This characterization is leveraged to understand, and reformulate three interesting applications related to deep neural network. First, we give a geo- metrical demonstration of the behaviour of the lottery ticket hypothesis. Moreover, we deploy the geometrical characterization of the decision boundaries to reformulate the network pruning problem. This new formulation aims to prune network pa- rameters that are not contributing to the geometrical representation of the decision boundaries. In addition, we propose a dual view of adversarial attack that tackles both designing perturbations to the input image, and the equivalent perturbation to the decision boundaries.

No description available.

Recent advancements in deep learning have demonstrated its incredible ability to learn generalizable patterns and relationships automatically from data in a number of mainstream applications. However, the generalization power of deep learning methods largely comes at the costs of working with very large datasets and using highly compute-intensive models. Many applications cannot afford these costs needed to ensure generalizability of deep learning models. For instance, obtaining labeled data can be costly in scientific applications, and using large models may not be feasible in resource-constrained environments involving portable devices. This dissertation aims to improve efficiency in machine learning by exploring different ways to learn generalizable neural networks that require less labeled data and computational resources. We demonstrate that using physics supervision in scientific problems can reduce the need for labeled data, thereby improving data efficiency without compromising model generalizability. Additionally, we investigate the potential of transfer learning powered by transformers in scientific applications as a promising direction for further improving data efficiency. On the computational efficiency side, we present two efforts for increasing parameter efficiency of neural networks through novel architectures and structured network pruning. / Doctor of Philosophy / Deep learning is a powerful technique that can help us solve complex problems, but it often requires a lot of data and resources. This research aims to make deep learning more efficient, so it can be applied in more situations. We propose ways to make the deep learning models require less data and less computer power. For example, we leverage the physics rules as additional information for training the neural network to learn from less labeled data and we use a technique called transfer learning to leverage knowledge from data that is from other distribution. Transfer learning may allow us to further reduce the need for labeled data in scientific applications. We also look at ways to make the deep learning models use less computational resources, by effectively reducing their sizes via novel architectures or pruning out redundant structures.

In order to comprehend the environments from different aspects, a large variety of computer vision methods are developed to detect objects, classify objects or even segment them semantically. Semantic segmentation is growing in significance due to its broad applications in fields such as robotics, environmental understanding for virtual or augmented reality, and autonomous driving. The development of convolutional neural networks, as a powerful tool, has contributed to solving classification or object detection tasks with the trend of larger and deeper models. It is hard to compare the models from the perspective of depth since they are of different structure. At the same time, semantic segmentation is computationally demanding for the reason that it requires classifying each pixel to certain classes. Running these complicated processes on resource-constrained embedded systems may cause performance degradation in terms of inference time and accuracy. Network pruning, a model compression technique, targeting to eliminate the redundant parameters in the models based on a certain evaluation rule, is one solution. Most traditional network pruning methods, structural or nonstructural, apply zero masks to cover the original parameters rather than literally eliminate the connections. A new pruning method, Torch-Pruning, has a general-purpose library for structural pruning. This method is based on the dependency between parameters and it can remove groups of less important parameters and reconstruct the new model. A cutting-edge research work towards solving several computer vision tasks, Yolov8 has proposed several pre-trained models from nano, small, medium to large and xlarge with similar structure but different parameters for different applications. This thesis applies Torch-Pruning to Yolov8 semantic segmentation models to compare the performance of pruning based on existing models with similar structures, thus it is meaningful to compare the depth of the model as a factor. Several configurations of the pruning have been explored. The results show that greater depth does not always lead to better performance. Besides, pruning can bring about more generalization ability for Gaussian noise at medium level, from 20% to 40% compared with the original models. / För att förstå miljöer från olika perspektiv har en mängd olika datorseendemetoder utvecklats för att upptäcka objekt, klassificera objekt eller till och med segmentera dem semantiskt. Semantisk segmentering växer i betydelse på grund av dess breda tillämpningar inom områden som robotik, miljöförståelse för virtuell eller förstärkt verklighet och autonom körning. Utvecklingen av konvolutionella neurala nätverk, som är ett kraftfullt verktyg, har bidragit till att lösa klassificerings- eller objektdetektionsuppgifter med en trend mot större och djupare modeller. Det är svårt att jämföra modeller från djupets perspektiv eftersom de har olika struktur. Samtidigt är semantisk segmentering beräkningsintensiv eftersom den kräver att varje pixel klassificeras till vissa klasser. Att köra dessa komplicerade processer på resursbegränsade inbäddade system kan orsaka prestandanedgång när det gäller inferenstid och noggrannhet. Nätverksbeskärning, en modellkomprimeringsteknik som syftar till att eliminera överflödiga parametrar i modellerna baserat på en viss utvärderingsregel, är en lösning. De flesta traditionella nätverksbeskärningsmetoder, både strukturella och icke-strukturella, tillämpar nollmasker för att täcka de ursprungliga parametrarna istället för att bokstavligen eliminera anslutningarna. En ny beskärningsmetod, Torch-Pruning, har en allmän användningsområde för strukturell beskärning. Denna metod är baserad på beroendet mellan parametrar och den kan ta bort grupper av mindre viktiga parametrar och återskapa den nya modellen. Ett banbrytande forskningsarbete för att lösa flera datorseenduppgifter, Yolov8, har föreslagit flera förtränade modeller från nano, liten, medium till stor och xstor med liknande struktur men olika parametrar för olika tillämpningar. Denna avhandling tillämpar Torch-Pruning på Yolov8 semantiska segmenteringsmodeller för att jämföra prestandan för beskärning baserad på befintliga modeller med liknande strukturer, vilket gör det meningsfullt att jämföra djupet som en faktor. Flera konfigurationer av beskärningen har utforskats. Resultaten visar att större djup inte alltid leder till bättre prestanda. Dessutom kan beskärning medföra en större generaliseringsförmåga för gaussiskt brus på medelnivå, från 20% till 40%, jämfört med de ursprungliga modellerna.

With the fast development of embedded deep-learning computing systems, applications powered by deep learning are moving from the cloud to the edge. When deploying NN onto the devices under complex environments, there are various types of possible faults: soft errors caused by cosmic radiation and radioactive impurities, voltage instability, aging, temperature variations, etc. Thus, more attention is drawn on the reliability of the NN embedded system. In this project, we build a virtual simulation system in Gazebo to simulate and test the working of an embedded NN system in the virtual environment in indoor navigation. The system can detect objects in the virtual environment with the help of the virtual camera(the image sensor) and the object detection module, which is based on YOLO v3, and make corresponding control decisions. We also designed and simulated the corresponding error injection module according to the working principle of the image sensor, and tested the functionality, and fault tolerance of the YOLO network. At the same time, network pruning algorithm is also introduced to study the relationship between different degrees of network pruning and network fault tolerance to sensor faults. / Med den snabba utvecklingen av inbyggda datorsystem för djupinlärning flyttas applikationer som drivs av djupinlärning från molnet till kanten. När man distribuerar NN på enheterna under komplexa miljöer finns det olika typer av möjliga fel: mjuka fel orsakade av kosmisk strålning och radioaktiva föroreningar, spänningsinstabilitet, åldrande, temperaturvariationer, illvilliga angripare, etc. Därför är mer uppmärksamhet ritade om tillförlitligheten hos det inbyggda NN-systemet. I det här projektet bygger vi ett virtuellt simuleringssystem för att simulera och testa hur ett inbäddat NN-system fungerar i den virtuella miljö vi ställer upp. Systemet kan upptäcka objekt i den virtuella miljön enligt den virtuella kameran och objektdetekteringsmodulen, som är baserad på YOLO v3, och göra motsvarande kontrollstrategier. Vi designade och simulerade också motsvarande felinsprutningsmodul enligt bildsensorns arbetsprincip och testade funktionalitet, tillförlitlighet och feltolerans hos YOLO-nätverket. Samtidigt nätverk beskärningsalgoritm introduceras också för att studera sambandet mellan olika grader av nätverksbeskärning och nätverksfeltolerans.