To see the other types of publications on this topic, follow the link: DEEP LEARNING MODEL.
Dissertations / Theses on the topic 'DEEP LEARNING MODEL'
Create a spot-on reference in APA, MLA, Chicago, Harvard, and other styles
Consult the top 50 dissertations / theses for your research on the topic 'DEEP LEARNING MODEL.'
Next to every source in the list of references, there is an 'Add to bibliography' button. Press on it, and we will generate automatically the bibliographic reference to the chosen work in the citation style you need: APA, MLA, Harvard, Chicago, Vancouver, etc.
You can also download the full text of the academic publication as pdf and read online its abstract whenever available in the metadata.
Browse dissertations / theses on a wide variety of disciplines and organise your bibliography correctly.
1
Meng, Zhaoxin. "A deep learning model for scene recognition." Thesis, Mittuniversitetet, Institutionen för informationssystem och –teknologi, 2019. http://urn.kb.se/resolve?urn=urn:nbn:se:miun:diva-36491.
Full textAbstract:
Scene recognition is a hot research topic in the field of image recognition. It is necessary that we focus on the research on scene recognition, because it is helpful to the scene understanding topic, and can provide important contextual information for object recognition. The traditional approaches for scene recognition still have a lot of shortcomings. In these years, the deep learning method, which uses convolutional neural network, has got state-of-the-art results in this area. This thesis constructs a model based on multi-layer feature extraction of CNN and transfer learning for scene recognition tasks. Because scene images often contain multiple objects, there may be more useful local semantic information in the convolutional layers of the network, which may be lost in the full connected layers. Therefore, this paper improved the traditional architecture of CNN, adopted the existing improvement which enhanced the convolution layer information, and extracted it using Fisher Vector. Then this thesis introduced the idea of transfer learning, and tried to introduce the knowledge of two different fields, which are scene and object. We combined the output of these two networks to achieve better results. Finally, this thesis implemented the method using Python and PyTorch. This thesis applied the method to two famous scene datasets. the UIUC-Sports and Scene-15 datasets. Compared with traditional CNN AlexNet architecture, we improve the result from 81% to 93% in UIUC-Sports, and from 79% to 91% in Scene- 15. It shows that our method has good performance on scene recognition tasks.
2
Zeledon, Lostalo Emilia Maria. "FMRI IMAGE REGISTRATION USING DEEP LEARNING." OpenSIUC, 2019. https://opensiuc.lib.siu.edu/theses/2641.
Full textAbstract:
fMRI imaging is considered key on the understanding of the brain and the mind, for this reason has been the subject of tremendous research connecting different disciplines. The intrinsic complexity of this 4-D type of data processing and analysis has been approached with every single computational perspective, lately increasing the trend to include artificial intelligence. One step critical on the fMRI pipeline is image registration. A model of Deep Networks based on Fully Convolutional Neural Networks, spatial transformation neural networks with a self-learning strategy was proposed for the implementation of a Fully deformable model image registration algorithm. Publicly available fMRI datasets with images from real-life subjects were used for training, testing and validating the model. The model performance was measured in comparison with ANTs deformable registration method with good results suggesting that Deep Learning can be used successfully for the development of the field using the basic strategy of studying the brain using the brain-self strategies.
3
Giovanelli, Francesco. "Model Agnostic solution of CSPs with Deep Learning." Master's thesis, Alma Mater Studiorum - Università di Bologna, 2019. http://amslaurea.unibo.it/18633/.
Full textAbstract:
Negli ultimi anni, le tecniche di Deep Learning sono state notevolmente migliorate, permettendo di affrontare con successo numerosi problemi. Il Deep Learning ha un approccio sub-simbolico ai problemi, perciò non si rende necessario descrivere esplicitamente informazioni sulla struttura del problema per fare sì che questo possa essere affrontato con successo; l'idea è quindi di utilizzare reti neurali di Deep Learning per affrontare problemi con vincoli (CSPs), senza dover fare affidamento su conoscenza esplicita riguardo ai vincoli dei problemi. Chiamiamo questo approccio Model Agnostic; esso può rivelarsi molto utile se usato sui CSP, dal momento che è spesso difficile esprimerne tutti i dettagli: potrebbero esistere vincoli, o preferenze, che non sono menzionati esplicitamente, e che sono intuibili solamente dall'analisi di soluzioni precedenti del problema. In questi casi, un modello di Deep Learning in grado di apprendere la struttura del CSP potrebbe avere applicazioni pratiche rilevanti.
In particolar modo, in questa tesi si è indagato sul fatto che una Deep Neural Network possa essere capace di risolvere il rompicapo delle 8 regine. Sono state create due diverse reti neurali, una rete Generatore e una rete Discriminatore, che hanno dovuto apprendere differenti caratteristiche del problema. La rete Generatore è stata addestrata per produrre un singolo assegnamento, in modo che questo sia globalmente consistente; la rete Discriminatore è stata invece addestrata a distinguere tra soluzioni ammissibili e non ammissibili, con l'idea che possa essere utilizzata come controllore dell'euristica.
Infine, sono state combinate le due reti in un unico modello, chiamato Generative Adversarial Network (GAN), in modo che esse possano scambiarsi conoscenza riguardo al problema, con l'obiettivo di migliorare le prestazioni di entrambe.
4
Matsoukas, Christos. "Model Distillation for Deep-Learning-Based Gaze Estimation." Thesis, KTH, Skolan för elektroteknik och datavetenskap (EECS), 2019. http://urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-261412.
Full textAbstract:
With the recent advances in deep learning, the gaze estimation models reached new levels, in terms of predictive accuracy, that could not be achieved with older techniques. Nevertheless, deep learning consists of computationally and memory expensive algorithms that do not allow their integration for embedded systems. This work aims to tackle this problem by boosting the predictive power of small networks using a model compression method called "distillation". Under the concept of distillation, we introduce an additional term to the compressed model’s total loss which is a bounding term between the compressed model (the student) and a powerful one (the teacher). We show that the distillation method introduces to the compressed model something more than noise. That is, the teacher’s inductive bias which helps the student to reach a better optimum due to the adaptive error deduction. Furthermore, we show that the MobileNet family exhibits unstable training phases and we report that the distilled MobileNet25 slightly outperformed the MobileNet50. Moreover, we try newly proposed training schemes to increase the predictive power of small and thin networks and we infer that extremely thin architectures are hard to train. Finally, we propose a new training scheme based on the hintlearning method and we show that this technique helps the thin MobileNets to gain stability and predictive power.Den senaste utvecklingen inom djupinlärning har hjälp till att förbättra precisionen hos gaze estimation-modeller till nivåer som inte tidigare varit möjliga. Dock kräver djupinlärningsmetoder oftast både stora mängder beräkningar och minne som därmed begränsar dess användning i inbyggda system med små minnes- och beräkningsresurser. Det här arbetet syftar till att kringgå detta problem genom att öka prediktiv kraft i små nätverk som kan användas i inbyggda system, med hjälp av en modellkomprimeringsmetod som kallas distillation". Under begreppet destillation introducerar vi ytterligare en term till den komprimerade modellens totala optimeringsfunktion som är en avgränsande term mellan en komprimerad modell och en kraftfull modell. Vi visar att destillationsmetoden inför mer än bara brus i den komprimerade modellen. Det vill säga lärarens induktiva bias som hjälper studenten att nå ett bättre optimum tack vare adaptive error deduction. Utöver detta visar vi att MobileNet-familjen uppvisar instabila träningsfaser och vi rapporterar att den destillerade MobileNet25 överträffade sin lärare MobileNet50 något. Dessutom undersöker vi nyligen föreslagna träningsmetoder för att förbättra prediktionen hos små och tunna nätverk och vi konstaterar att extremt tunna arkitekturer är svåra att träna. Slutligen föreslår vi en ny träningsmetod baserad på hint-learning och visar att denna teknik hjälper de tunna MobileNets att stabiliseras under träning och ökar dess prediktiva effektivitet.
5
Lim, Steven. "Recommending TEE-based Functions Using a Deep Learning Model." Thesis, Virginia Tech, 2021. http://hdl.handle.net/10919/104999.
Full textAbstract:
Trusted execution environments (TEEs) are an emerging technology that provides a protected hardware environment for processing and storing sensitive information. By using
TEEs, developers can bolster the security of software systems. However, incorporating
TEE into existing software systems can be a costly and labor-intensive endeavor. Software
maintenance—changing software after its initial release—is known to contribute the majority of the cost in the software development lifecycle. The first step of making use of a TEE
requires that developers accurately identify which pieces of code would benefit from being
protected in a TEE. For large code bases, this identification process can be quite tedious and
time-consuming. To help reduce the software maintenance costs associated with introducing
a TEE into existing software, this thesis introduces ML-TEE, a recommendation tool that
uses a deep learning model to classify whether an input function handles sensitive information or sensitive code. By applying ML-TEE, developers can reduce the burden of manual
code inspection and analysis. ML-TEE's model was trained and tested on functions from
GitHub repositories that use Intel SGX and on an imbalanced dataset. The accuracy of
the final model used in the recommendation system has an accuracy of 98.86% and an F1
score of 80.00%. In addition, we conducted a pilot study, in which participants were asked
to identify functions that needed to be placed inside a TEE in a third-party project. The
study found that on average, participants who had access to the recommendation system's
output had a 4% higher accuracy and completed the task 21% faster.Master of Science
Improving the security of software systems has become critically important. A trusted execution environment (TEE) is an emerging technology that can help secure software that uses or stores confidential information. To make use of this technology, developers need to identify which pieces of code handle confidential information and should thus be placed in a TEE. However, this process is costly and laborious because it requires the developers to understand the code well enough to make the appropriate changes in order to incorporate a TEE. This process can become challenging for large software that contains millions of lines of code. To help reduce the cost incurred in the process of identifying which pieces of code should be placed within a TEE, this thesis presents ML-TEE, a recommendation system that uses a deep learning model to help reduce the number of lines of code a developer needs to inspect. Our results show that the recommendation system achieves high accuracy as well as a good balance between precision and recall. In addition, we conducted a pilot study and found that participants from the intervention group who used the output from the recommendation system managed to achieve a higher average accuracy and perform the assigned task faster than the participants in the control group.
6
Del, Vecchio Matteo. "Improving Deep Question Answering: The ALBERT Model." Master's thesis, Alma Mater Studiorum - Università di Bologna, 2020. http://amslaurea.unibo.it/20414/.
Full textAbstract:
Natural Language Processing is a field of Artificial Intelligence referring to the ability of computers to understand human speech and language, often in a written form, mainly by using Machine Learning and Deep Learning methods to extract patterns.
Languages are challenging by definition, because of their differences, their abstractions and their ambiguities; consequently, their processing is often very demanding, in terms of modelling the problem and resources. Retrieving all sentences in a given text is something that can be easily accomplished with just few lines of code, but what about checking whether a given sentence conveys a message with sarcasm or not? This is something difficult for humans too and therefore, it requires complex modelling mechanisms to be addressed. This kind of information, in fact, poses the problem of its encoding and representation in a meaningful way.
The majority of research involves finding and understanding all characteristics of text, in order to develop sophisticated models to address tasks such as Machine Translation, Text Summarization and Question Answering.
This work will focus on ALBERT, from Google Research, which is one of the recently released state-of-the-art models and investigate its performance on the Question Answering task. In addition, some ideas will be developed and experimented in order to improve model's performance on the Stanford Question Answering Dataset (SQuAD), after exploring breakthrough changes that made training and fine-tuning of huge language models possible.
7
Wu, Xinheng. "A Deep Unsupervised Anomaly Detection Model for Automated Tumor Segmentation." Thesis, The University of Sydney, 2020. https://hdl.handle.net/2123/22502.
Full textAbstract:
Many researches have been investigated to provide the computer aided diagnosis (CAD) automated tumor segmentation in various medical images, e.g., magnetic resonance (MR), computed tomography (CT) and positron-emission tomography (PET). The recent advances in automated tumor segmentation have been achieved by supervised deep learning (DL) methods trained on large labelled data to cover tumor variations. However, there is a scarcity in such training data due to the cost of labeling process. Thus, with insufficient training data, supervised DL methods have difficulty in generating effective feature representations for tumor segmentation. This thesis aims to develop an unsupervised DL method to exploit large unlabeled data generated during clinical process. Our assumption is unsupervised anomaly detection (UAD) that, normal data have constrained anatomy and variations, while anomalies, i.e., tumors, usually differ from the normality with high diversity. We demonstrate our method for automated tumor segmentation on two different image modalities. Firstly, given that bilateral symmetry in normal human brains and unsymmetry in brain tumors, we propose a symmetric-driven deep UAD model using GAN model to model the normal symmetric variations thus segmenting tumors by their being unsymmetrical. We evaluated our method on two benchmarked datasets. Our results show that our method outperformed the state-of-the-art unsupervised brain tumor segmentation methods and achieved competitive performance to the supervised segmentation methods. Secondly, we propose a multi-modal deep UAD model for PET-CT tumor segmentation. We model a manifold of normal variations shared across normal CT and PET pairs; this manifold representing the normal pairing that can be used to segment the anomalies. We evaluated our method on two PET-CT datasets and the results show that we outperformed the state-of-the-art unsupervised methods, supervised methods and baseline fusion techniques.
8
Kayesh, Humayun. "Deep Learning for Causal Discovery in Texts." Thesis, Griffith University, 2022. http://hdl.handle.net/10072/415822.
Full textAbstract:
Causality detection in text data is a challenging natural language processing task. This is a trivial task for human beings as they acquire vast background knowledge throughout their lifetime. For example, a human knows from their experience that heavy rain may cause flood or plane accidents may cause death. However, it is challenging to automatically detect such causal relationships in texts due to the availability of limited contextual information and the unstructured nature of texts. The task is even more challenging for social media short texts such as Tweets as often they are informal, short, and grammatically incorrect. Generating hand-crafted linguistic rules is an option but is not always effective to detect causal relationships in text because they are rigid and require grammatically correct sentences. Also, the rules are often domain-specific and not always portable to another domain. Therefore, supervised learning techniques are more appropriate in the above scenario. Traditional machine learning-based model also suffers from the high dimensional features of texts. This is why deep learning-based approaches are becoming increasingly popular for natural language processing tasks such as causality detection. However, deep learning models often require large datasets with high-quality features to perform well. Extracting deeply-learnable causal features and applying them to a carefully designed deep learning model is important. Also, preparing a large human-labeled training dataset is expensive and time-consuming. Even if a large training dataset is available, it is computationally expensive to train a deep learning model due to the complex structure of neural networks. We focus on addressing the following challenges: (i) extracting highquality causal features, (ii) designing an effective deep learning model to learn from the causal features, and (iii) reducing the dependency on large training datasets. Our main goals in this thesis are as follows: (i) we aim to study the different aspects of causality and causal discovery in text in depth. (ii) We aim to develop strategies to model causality in text, (iii) and finally, we aim to develop frameworks to design effective and efficient deep neural network structures to discover causality in texts.Thesis (PhD Doctorate)
Doctor of Philosophy (PhD)
School of Info & Comm Tech
Science, Environment, Engineering and Technology
Full Text
9
Зайяд, Абдаллах Мухаммед. "Ecrypted Network Classification With Deep Learning." Master's thesis, КПІ ім. Ігоря Сікорського, 2020. https://ela.kpi.ua/handle/123456789/34069.
Full textAbstract:
Дисертація складається з 84 сторінок, 59 Цифри та 29 джерел у
довідковому списку.
Проблема: Оскільки світ стає більш безпечним, для забезпечення
належної передачі даних між сторонами, що спілкуються, було
використано більше протоколів шифрування. Класифікація мережі стала
більше клопоту з використанням деяких прийомів, оскільки перевірка
зашифрованого трафіку в деяких країнах може бути незаконною. Це
заважає інженерам мережі мати можливість класифікувати трафік, щоб
відрізняти зашифрований від незашифрованого трафіку.
Мета роботи: Ця стаття спрямована на проблему, спричинену попередніми
методами, використовуваними в шифрованій мережевій класифікації.
Деякі з них обмежені розміром даних та обчислювальною потужністю. У
даній роботі використовується рішення алгоритму глибокого навчання для
вирішення цієї проблеми.
Основні завдання дослідження:
1. Порівняйте попередні традиційні методи та порівняйте їх
переваги та недоліки
2. Вивчити попередні супутні роботи у сучасній галузі
досліджень.
3. Запропонуйте більш сучасний та ефективний метод та алгоритм
для зашифрованої класифікації мережевого трафіку
Об'єкт дослідження: Простий алгоритм штучної нейронної
мережі для точної та надійної класифікації мережевого трафіку, що не
залежить від розміру даних та обчислювальної потужності.
Предмет дослідження: На основі даних, зібраних із приватного
потоку трафіку у нашому власному інструменті моделювання мережі. За
4
допомогою запропонованого нами методу визначаємо відмінності
корисних навантажень мережевого трафіку та класифікуємо мережевий
трафік. Це допомогло відокремити або класифікувати зашифровані від
незашифрованого трафіку.
Методи дослідження: Експериментальний метод.
Ми провели наш експеримент із моделюванням мережі та збиранням
трафіку різних незашифрованих протоколів та зашифрованих протоколів.
Використовуючи мову програмування python та бібліотеку Keras, ми
розробили згорнуту нейронну мережу, яка змогла прийняти корисне
навантаження зібраного трафіку, навчити модель та класифікувати трафік
у нашому тестовому наборі з високою точністю без вимоги високої
обчислювальної потужності.This dissertation consists of 84 pages, 59 Figures and 29 sources in the reference list. Problem: As the world becomes more security conscious, more encryption protocols have been employed in ensuring suecure data transmission between communicating parties. Network classification has become more of a hassle with the use of some techniques as inspecting encrypted traffic can pose to be illegal in some countries. This has hindered network engineers to be able to classify traffic to differentiate encrypted from unencrypted traffic. Purpose of work: This paper aims at the problem caused by previous techniques used in encrypted network classification. Some of which are limited to data size and computational power. This paper employs the use of deep learning algorithm to solve this problem. The main tasks of the research: 1. Compare previous traditional techniques and compare their advantages and disadvantages 2. Study previous related works in the current field of research. 3. Propose a more modern and efficient method and algorithm for encrypted network traffic classification The object of research: Simple artificial neural network algorithm for accurate and reliable network traffic classification that is independent of data size and computational power. The subject of research: Based on data collected from private traffic flow in our own network simulation tool. We use our proposed method to identify the differences in network traffic payloads and classify network traffic. It helped to separate or classify encrypted from unencrypted traffic. 6 Research methods: Experimental method. We have carried out our experiment with network simulation and gathering traffic of different unencrypted protocols and encrypted protocols. Using python programming language and the Keras library we developed a convolutional neural network that was able to take in the payload of the traffic gathered, train the model and classify the traffic in our test set with high accuracy without the requirement of high computational power.
10
Zhao, Yajing. "Chaotic Model Prediction with Machine Learning." BYU ScholarsArchive, 2020. https://scholarsarchive.byu.edu/etd/8419.
Full textAbstract:
Chaos theory is a branch of modern mathematics concerning the non-linear dynamic systems that are highly sensitive to their initial states. It has extensive real-world applications, such as weather forecasting and stock market prediction. The Lorenz system, defined by three ordinary differential equations (ODEs), is one of the simplest and most popular chaotic models. Historically research has focused on understanding the Lorenz system's mathematical characteristics and dynamical evolution including the inherent chaotic features it possesses. In this thesis, we take a data-driven approach and propose the task of predicting future states of the chaotic system from limited observations. We explore two directions, answering two distinct fundamental questions of the system based on how informed we are about the underlying model. When we know the data is generated by the Lorenz System with unknown parameters, our task becomes parameter estimation (a white-box problem), or the ``inverse'' problem. When we know nothing about the underlying model (a black-box problem), our task becomes sequence prediction. We propose two algorithms for the white-box problem: Markov-Chain-Monte-Carlo (MCMC) and a Multi-Layer-Perceptron (MLP). Specially, we propose to use the Metropolis-Hastings (MH) algorithm with an additional random walk to avoid the sampler being trapped into local energy wells. The MH algorithm achieves moderate success in predicting the $\rho$ value from the data, but fails at the other two parameters. Our simple MLP model is able to attain high accuracy in terms of the $l_2$ distance between the prediction and ground truth for $\rho$ as well, but also fails to converge satisfactorily for the remaining parameters. We use a Recurrent Neural Network (RNN) to tackle the black-box problem. We implement and experiment with several RNN architectures including Elman RNN, LSTM, and GRU and demonstrate the relative strengths and weaknesses of each of these methods. Our results demonstrate the promising role of machine learning and modern statistical data science methods in the study of chaotic dynamic systems. The code for all of our experiments can be found on \url{https://github.com/Yajing-Zhao/}
11
Saitas-Zarkias, Konstantinos. "Insights into Model-Agnostic Meta-Learning on Reinforcement Learning Tasks." Thesis, KTH, Skolan för elektroteknik och datavetenskap (EECS), 2021. http://urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-290903.
Full textAbstract:
Meta-learning has been gaining traction in the Deep Learning field as an approach to build models that are able to efficiently adapt to new tasks after deployment. Contrary to conventional Machine Learning approaches, which are trained on a specific task (e.g image classification on a set of labels), meta-learning methods are meta-trained across multiple tasks (e.g image classification across multiple sets of labels). Their end objective is to learn how to solve unseen tasks with just a few samples. One of the most renowned methods of the field is Model-Agnostic Meta-Learning (MAML). The objective of this thesis is to supplement the latest relevant research with novel observations regarding the capabilities, limitations and network dynamics of MAML. For this end, experiments were performed on the meta-reinforcement learning benchmark Meta-World. Additionally, a comparison with a recent variation of MAML, called Almost No Inner Loop (ANIL) was conducted, providing insights on the changes of the network’s representation during adaptation (meta-testing). The results of this study indicate that MAML is able to outperform the baselines on the challenging Meta-World benchmark but shows little signs actual ”rapid learning” during meta-testing thus supporting the hypothesis that it reuses features learnt during meta-training.Meta-Learning har fått dragkraft inom Deep Learning fältet som ett tillvägagångssätt för att bygga modeller som effektivt kan anpassa sig till nya uppgifter efter distribution. I motsats till konventionella maskininlärnings metoder som är tränade för en specifik uppgift (t.ex. bild klassificering på en uppsättning klasser), så metatränas meta-learning metoder över flera uppgifter (t.ex. bild klassificering över flera uppsättningar av klasser). Deras slutmål är att lära sig att lösa osedda uppgifter med bara några få prover. En av de mest kända metoderna inom området är Model-Agnostic Meta-Learning (MAML). Syftet med denna avhandling är att komplettera den senaste relevanta forskningen med nya observationer avseende MAML: s kapacitet, begränsningar och nätverksdynamik. För detta ändamål utfördes experiment på metaförstärkningslärande riktmärke Meta-World. Dessutom gjordes en jämförelse med en ny variant av MAML, kallad Almost No Inner Loop (ANIL), som gav insikter om förändringarna i nätverkets representation under anpassning (metatestning). Resultaten av denna studie indikerar att MAML kan överträffa baslinjerna för det utmanande Meta-Worldriktmärket men visar små tecken på faktisk ”snabb inlärning” under metatestning, vilket stödjer hypotesen att den återanvänder funktioner som den lärt sig under metaträning.
12
Lindespång, Victor. "Bildklassificering av bilar med hjälp av deep learning." Thesis, Örebro universitet, Institutionen för naturvetenskap och teknik, 2017. http://urn.kb.se/resolve?urn=urn:nbn:se:oru:diva-58361.
Full textAbstract:
Den här rapporten beskriver hur en bildklassificerare skapades med förmågan att via en given bild på en bil avgöra vilken bilmodell bilen är av. Klassificeringsmodellen utvecklades med hjälp av bilder som företaget CAB sparat i samband med försäkringsärenden som behandlats via deras nuvarande produkter. Inledningsvis i rapporten så beskrivs teori för maskininlärning och djupinlärning på engrundläggande nivå för att leda in läsaren på ämnesområdet som rör rapporten, och fortsätter sedan med problemspecifika metoder som var till nytta för det aktuella problemet. Rapporten tar upp metoder för hur datan bearbetats i förväg, hur träningsprocessen gick till med de valda verktygen samt diskussion kring resultatet och vad som påverkade det – med kommentarer om vad som kan göras i framtiden för att förbättra slutprodukten.This report describes how an image classifier was created with the ability to identify car makeand model from a given picture of a car. The classifier was developed using pictures that the company CAB had saved from insurance errands that was managed through their current products. First of all the report begins with a brief theoretical introduction to machine learning and deep learning to guide the reader in to the subject of the report, and then continues with problemspecific methods that were of good use for the project. The report brings up methods for how the data was processed before training took place, how the training process went with the chosen tools for this project and also discussion about the result and what effected it – with comments about what can be done in the future to improve the end product.
13
Hellström, Terese. "Deep-learning based prediction model for dose distributions in lung cancer patients." Thesis, Stockholms universitet, Fysikum, 2021. http://urn.kb.se/resolve?urn=urn:nbn:se:su:diva-196891.
Full textAbstract:
Background To combat one of the leading causes of death worldwide, lung cancer treatment techniques and modalities are advancing, and the treatment options are becoming increasingly individualized. Modern cancer treatment includes the option for the patient to be treated with proton therapy, which can in some cases spare healthy tissue from excessive dose better than conventional photon radiotherapy. However, to assess the benefit of proton therapy compared to photon therapy, it is necessary to make both treatment plans to get information about the Tumour Control Probability (TCP) and the Normal Tissue Complication Probability (NTCP). This requires excessive treatment planning time and increases the workload for planners. Aim This project aims to investigate the possibility for automated prediction of the treatment dose distribution using a deep learning network for lung cancer patients treated with photon radiotherapy. This is an initial step towards decreasing the overall planning time and would allow for efficient estimation of the NTCP for each treatment plan and lower the workload of treatment planning technicians. The purpose of the current work was also to understand which features of the input data and training specifics were essential for producing accurate predictions. Methods Three different deep learning networks were developed to assess the difference in performance based on the complexity of the input for the network. The deep learning models were applied for predictions of the dose distribution of lung cancer treatment and used data from 95 patient treatments. The networks were trained with a U-net architecture using input data from the planning Computed Tomography (CT) and volume contours to produce an output of the dose distribution of the same image size. The network performance was evaluated based on the error of the predicted mean dose to Organs At Risk (OAR) as well as the shape of the predicted Dose-Volume Histogram (DVH) and individual dose distributions. Results The optimal input combination was the CT scan and lung, mediastinum envelope and Planning Target Volume (PTV) contours. The model predictions showed a homogenous dose distribution over the PTV with a steep fall-off seen in the DVH. However, the dose distributions had a blurred appearance and the predictions of the doses to the OARs were therefore not as accurate as of the doses to the PTV compared to the manual treatment plans. The performance of the network trained with the Houndsfield Unit input of the CT scan had similar performance as the network trained without it. Conclusions As one of the novel attempts to assess the potential for a deep learning-based prediction model for the dose distribution based on minimal input, this study shows promising results. To develop this kind of model further a larger data set would be needed and the training method could be expanded as a generative adversarial network or as a more developed U-net network.
14
Viebke, André. "Accelerated Deep Learning using Intel Xeon Phi." Thesis, Linnéuniversitetet, Institutionen för datavetenskap (DV), 2015. http://urn.kb.se/resolve?urn=urn:nbn:se:lnu:diva-45491.
Full textAbstract:
Deep learning, a sub-topic of machine learning inspired by biology, have achieved wide attention in the industry and research community recently. State-of-the-art applications in the area of computer vision and speech recognition (among others) are built using deep learning algorithms. In contrast to traditional algorithms, where the developer fully instructs the application what to do, deep learning algorithms instead learn from experience when performing a task. However, for the algorithm to learn require training, which is a high computational challenge. High Performance Computing can help ease the burden through parallelization, thereby reducing the training time; this is essential to fully utilize the algorithms in practice. Numerous work targeting GPUs have investigated ways to speed up the training, less attention have been paid to the Intel Xeon Phi coprocessor. In this thesis we present a parallelized implementation of a Convolutional Neural Network (CNN), a deep learning architecture, and our proposed parallelization scheme, CHAOS. Additionally a theoretical analysis and a performance model discuss the algorithm in detail and allow for predictions if even more threads are available in the future. The algorithm is evaluated on an Intel Xeon Phi 7120p, Xeon E5-2695v2 2.4 GHz and Core i5 661 3.33 GHz using various architectures and thread counts on the MNIST dataset. Findings show a 103.5x, 99.9x, 100.4x speed up for the large, medium, and small architecture respectively for 244 threads compared to 1 thread on the coprocessor. Moreover, a 10.9x - 14.1x (large to small) speed up compared to the sequential version running on Xeon E5. We managed to decrease training time from 7 days on the Core i5 and 31 hours on the Xeon E5, to 3 hours on the Intel Xeon Phi when training our large network for 15 epochs
15
Iannello, Michele. "Deep Learning and Constrained Optimization for Epidemic Control." Master's thesis, Alma Mater Studiorum - Università di Bologna, 2022. http://amslaurea.unibo.it/25815/.
Full textAbstract:
The SARS-CoV-2 pandemic has galvanized the interest of the scientific community. Particular interest has been posed on methodologies apt at predicting the trend of the epidemiological curve, i.e. the daily number of infected individuals in the population.
In this work, we argue for a model capable of producing intervention plans focused on counteracting the negative effects of an outbreak, with real applications on the ongoing pandemic.
To do so, we relied on the use of Machine Learning models and Combinatorial Optimization approaches.
The project entails the development of a new predictive model capable of forecasting the number of infected individuals depending on non-pharmaceutical interventions. The development of the model required the use of state-of-the-art techniques, which relied on prior knowledge injection to guide the training process. The model is then used to boost a combinatorial process effectively producing the intervention plan. The ultimate result is a working prototype of a Decision Support System capable of assisting policy-makers during a virus outbreak.
16
Wang, Junpeng. "Interpreting and Diagnosing Deep Learning Models: A Visual Analytics Approach." The Ohio State University, 2019. http://rave.ohiolink.edu/etdc/view?acc_num=osu1555499299957829.
Full text
17
Wang, Wei. "Image Segmentation Using Deep Learning Regulated by Shape Context." Thesis, KTH, Skolan för elektroteknik och datavetenskap (EECS), 2018. http://urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-227261.
Full textAbstract:
In recent years, image segmentation by using deep neural networks has made great progress. However, reaching a good result by training with a small amount of data remains to be a challenge. To find a good way to improve the accuracy of segmentation with limited datasets, we implemented a new automatic chest radiographs segmentation experiment based on preliminary works by Chunliang using deep learning neural network combined with shape context information. When the process was conducted, the datasets were put into origin U-net at first. After the preliminary process, the segmented images were then repaired through a new network with shape context information. In this experiment, we created a new network structure by rebuilding the U-net into a 2-input structure and refined the processing pipeline step. In this proposed pipeline, the datasets and shape context were trained together through the new network model by iteration. The proposed method was evaluated on 247 posterior-anterior chest radiographs of public datasets and n-folds cross-validation was also used. The outcome shows that compared to origin U-net, the proposed pipeline reaches higher accuracy when trained with limited datasets. Here the "limited" datasets refer to 1-20 images in the medical image field. A better outcome with higher accuracy can be reached if the second structure is further refined and shape context generator's parameter is fine-tuned in the future.Under de senaste åren har bildsegmentering med hjälp av djupa neurala nätverk gjort stora framsteg. Att nå ett bra resultat med träning med en liten mängd data kvarstår emellertid som en utmaning. För att hitta ett bra sätt att förbättra noggrannheten i segmenteringen med begränsade datamängder så implementerade vi en ny segmentering för automatiska röntgenbilder av bröstkorgsdiagram baserat på tidigare forskning av Chunliang. Detta tillvägagångssätt använder djupt lärande neurala nätverk kombinerat med "shape context" information. I detta experiment skapade vi en ny nätverkstruktur genom omkonfiguration av U-nätverket till en 2-inputstruktur och förfinade pipeline processeringssteget där bilden och "shape contexten" var tränade tillsammans genom den nya nätverksmodellen genom iteration.Den föreslagna metoden utvärderades på dataset med 247 bröströntgenfotografier, och n-faldig korsvalidering användes för utvärdering. Resultatet visar att den föreslagna pipelinen jämfört med ursprungs U-nätverket når högre noggrannhet när de tränas med begränsade datamängder. De "begränsade" dataseten här hänvisar till 1-20 bilder inom det medicinska fältet. Ett bättre resultat med högre noggrannhet kan nås om den andra strukturen förfinas ytterligare och "shape context-generatorns" parameter finjusteras.
18
Di, Giacomo Emanuele. "A Deep Learning approach for predicting COSMO-Model's execution time." Master's thesis, Alma Mater Studiorum - Università di Bologna, 2021.
Find full textAbstract:
I modelli di previsione meteorologica sono programmi che permettono di simulare il tempo meteorologico futuro e formularne dunque una previsione. Il tempo di esecuzione di questi modelli è un aspetto critico, in quanto la loro utilità è basata sulle tempistiche con cui vengono prodotti i risultati. La previsione del tempo di esecuzione di un modello numerico di previsione meteorologica permette di ottimizzare sia la pianificazione dell'esecuzione del modello stesso che l'allocazione delle risorse a disposizione, nonché di individuare eventuali anomalie che si possono presentare e a fronte delle quali possono essere adottate delle contromisure. Nel presente lavoro si mostra come grazie all'uso di modelli deep learning si riescano ad ottenere risultati molto precisi per la previsione dei tempi di esecuzione su sistema HPC GALILEO del modello numerico di previsione meteorologica COSMO-Model in uso presso la Struttura Idro-Meteo-Clima di Arpae Emilia-Romagna. Il lavoro si compone di due parti: nella prima, sono stati definiti i parametri che influenzano i tempi di esecuzione del modello e sono state generate due tipologie di dataset, selezionando ed eseguendo su GALILEO numerose configurazioni di COSMO-Model. Nella seconda parte, i dataset sono stati usati per addestrare e valutare i modelli deep learning in grado di prevedere il tempo di esecuzione. Dalla loro valutazione è emerso come questi modelli deep learning permettano di ottenere risultati accurati per la previsione del tempo di esecuzione di COSMO-Model.
19
Beaudoin, Jean-Michel. "Growing deep roots : learning from the Essipit's culturally adapted model of Aboriginal forestry." Thesis, University of British Columbia, 2014. http://hdl.handle.net/2429/46590.
Full textAbstract:
Aboriginal peoples are seeking sustainable ways to steward and develop forests. Sustainable forestry is central to Aboriginal life and culture. Research indicates that the industrial forestry model has failed to address their socio-economic needs. To date, Aboriginal involvement in forestry is characterized by a limited economic role in forest development, limited influence over forest management, and an inability to integrate Aboriginal culture and values.
The case study of Essipit (Quebec, Canada) provides new insight on how Aboriginal communities can contribute to sustainable forestry. Growing deep roots means using a culturally adapted model of forestry that is consistent with Aboriginal culture and values, which is therefore more likely to support long-term social change and economic growth. To ensure reliability and validity, this research employed four data gathering techniques: observation, documentation, interviews and focus-groups. Results identify the entrepreneurship framework that led to the success of Aboriginal forest enterprises in Essipit, the level of authority held by Essipit over forest governance, and Essipit objectives for forest-based development. Therefore, this thesis provides a framework that aims to support Aboriginal forest development in theory and practice.
Despite constraints, such as timber access, capacity and institutions, Essipit was successful in engaging in forestry. Acquiring exclusive commercial rights to harvest wildlife became a key strategy that allowed Essipit to address social needs and create leverage for future forest-based activities. Essipit innovated in forest governance: they created a partnership with the forest company Boisaco and, thus, gained authority over forest management decisions at the operational level. Results indicate that the profitability motives of the forest industry are
iii
insufficient, because Essipit has other objectives and priorities. The forest industry looks primarily at the tree, while Essipit looks at everything that surrounds and supports it.
This research emphasizes the importance of developing a model that will outlast changes in government or industry. A forestry model that has deep roots is integrated into the community and the culture. It can sustain these types of changes and keep growing. Without this understanding of Aboriginal experiences, knowledge and objectives, local initiatives and government policies will remain uninformed and, potentially, fail.
20
Li, Mengtong. "An intelligent flood evacuation model based on deep learning of various flood scenarios." Doctoral thesis, Kyoto University, 2021. http://hdl.handle.net/2433/263634.
Full text
21
Zarrinkoub, Sahand. "Transfer Learning in Deep Structured Semantic Models for Information Retrieval." Thesis, KTH, Skolan för elektroteknik och datavetenskap (EECS), 2020. http://urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-286310.
Full textAbstract:
Recent approaches to IR include neural networks that generate query and document vector representations. The representations are used as the basis for document retrieval and are able to encode semantic features if trained on large datasets, an ability that sets them apart from classical IR approaches such as TF-IDF. However, the datasets necessary to train these networks are not available to the owners of most search services used today, since they are not used by enough users. Thus, methods for enabling the use of neural IR models in data-poor environments are of interest. In this work, a bag-of-trigrams neural IR architecture is used in a transfer learning procedure in an attempt to increase performance on a target dataset by pre-training on external datasets. The target dataset used is WikiQA, and the external datasets are Quora’s Question Pairs, Reuters’ RCV1 and SQuAD. When considering individual model performance, pre-training on Question Pairs and fine-tuning on WikiQA gives us the best individual models. However, when considering average performance, pre-training on the chosen external dataset result in lower performance on the target dataset, both when all datasets are used together and when they are used individually, with different average performance depending on the external dataset used. On average, pre-training on RCV1 and Question Pairs gives the lowest and highest average performance respectively, when considering only the pre-trained networks. Surprisingly, the performance of an untrained, randomly generated network is high, and beats the performance of all pre-trained networks on average. The best performing model on average is a neural IR model trained on the target dataset without prior pre-training.Nya modeller inom informationssökning inkluderar neurala nät som genererar vektorrepresentationer för sökfrågor och dokument. Dessa vektorrepresentationer används tillsammans med ett likhetsmått för att avgöra relevansen för ett givet dokument med avseende på en sökfråga. Semantiska särdrag i sökfrågor och dokument kan kodas in i vektorrepresentationerna. Detta möjliggör informationssökning baserat på semantiska enheter, vilket ej är möjligt genom de klassiska metoderna inom informationssökning, som istället förlitar sig på den ömsesidiga förekomsten av nyckelord i sökfrågor och dokument. För att träna neurala sökmodeller krävs stora datamängder. De flesta av dagens söktjänster används i för liten utsträckning för att möjliggöra framställande av datamängder som är stora nog att träna en neural sökmodell. Därför är det önskvärt att hitta metoder som möjliggör användadet av neurala sökmodeller i domäner med små tillgängliga datamängder. I detta examensarbete har en neural sökmodell implementerats och använts i en metod avsedd att förbättra dess prestanda på en måldatamängd genom att förträna den på externa datamängder. Måldatamängden som används är WikiQA, och de externa datamängderna är Quoras Question Pairs, Reuters RCV1 samt SquAD. I experimenten erhålls de bästa enskilda modellerna genom att föträna på Question Pairs och finjustera på WikiQA. Den genomsnittliga prestandan över ett flertal tränade modeller påverkas negativt av vår metod. Detta äller både när samtliga externa datamänder används tillsammans, samt när de används enskilt, med varierande prestanda beroende på vilken datamängd som används. Att förträna på RCV1 och Question Pairs ger den största respektive minsta negativa påverkan på den genomsnittliga prestandan. Prestandan hos en slumpmässigt genererad, otränad modell är förvånansvärt hög, i genomsnitt högre än samtliga förtränade modeller, och i nivå med BM25. Den bästa genomsnittliga prestandan erhålls genom att träna på måldatamängden WikiQA utan tidigare förträning.
22
Miao, Yishu. "Deep generative models for natural language processing." Thesis, University of Oxford, 2017. http://ora.ox.ac.uk/objects/uuid:e4e1f1f9-e507-4754-a0ab-0246f1e1e258.
Full textAbstract:
Deep generative models are essential to Natural Language Processing (NLP) due to their outstanding ability to use unlabelled data, to incorporate abundant linguistic features, and to learn interpretable dependencies among data. As the structure becomes deeper and more complex, having an effective and efficient inference method becomes increasingly important. In this thesis, neural variational inference is applied to carry out inference for deep generative models. While traditional variational methods derive an analytic approximation for the intractable distributions over latent variables, here we construct an inference network conditioned on the discrete text input to provide the variational distribution. The powerful neural networks are able to approximate complicated non-linear distributions and grant the possibilities for more interesting and complicated generative models. Therefore, we develop the potential of neural variational inference and apply it to a variety of models for NLP with continuous or discrete latent variables. This thesis is divided into three parts. Part I introduces a generic variational inference framework for generative and conditional models of text. For continuous or discrete latent variables, we apply a continuous reparameterisation trick or the REINFORCE algorithm to build low-variance gradient estimators. To further explore Bayesian non-parametrics in deep neural networks, we propose a family of neural networks that parameterise categorical distributions with continuous latent variables. Using the stick-breaking construction, an unbounded categorical distribution is incorporated into our deep generative models which can be optimised by stochastic gradient back-propagation with a continuous reparameterisation. Part II explores continuous latent variable models for NLP. Chapter 3 discusses the Neural Variational Document Model (NVDM): an unsupervised generative model of text which aims to extract a continuous semantic latent variable for each document. In Chapter 4, the neural topic models modify the neural document models by parameterising categorical distributions with continuous latent variables, where the topics are explicitly modelled by discrete latent variables. The models are further extended to neural unbounded topic models with the help of stick-breaking construction, and a truncation-free variational inference method is proposed based on a Recurrent Stick-breaking construction (RSB). Chapter 5 describes the Neural Answer Selection Model (NASM) for learning a latent stochastic attention mechanism to model the semantics of question-answer pairs and predict their relatedness. Part III discusses discrete latent variable models. Chapter 6 introduces latent sentence compression models. The Auto-encoding Sentence Compression Model (ASC), as a discrete variational auto-encoder, generates a sentence by a sequence of discrete latent variables representing explicit words. The Forced Attention Sentence Compression Model (FSC) incorporates a combined pointer network biased towards the usage of words from source sentence, which significantly improves the performance when jointly trained with the ASC model in a semi-supervised learning fashion. Chapter 7 describes the Latent Intention Dialogue Models (LIDM) that employ a discrete latent variable to learn underlying dialogue intentions. Additionally, the latent intentions can be interpreted as actions guiding the generation of machine responses, which could be further refined autonomously by reinforcement learning. Finally, Chapter 8 summarizes our findings and directions for future work.
23
Miserocchi, Andrea. "The Fokker-Planck equation as model for the stochastic gradient descent in deep learning." Bachelor's thesis, Alma Mater Studiorum - Università di Bologna, 2019. http://amslaurea.unibo.it/18290/.
Full textAbstract:
La discesa stocastica del gradiente (SGD) è alla base degli algoritmi di
ottimizzazione di reti di Deep Learning più usati in AI, dal riconoscimento
delle immagini all’elaborazione del linguaggio naturale. Questa tesi si propone di descrivere un modello basato sull’equazione di Fokker-Planck della dinamica del SGD. Si introduce la teoria dei processi stocastici, con particolare enfasi sulle equazioni di Langevin e sull’equazione di Fokker-Planck. Si mostra come il SGD minimizzi un funzionale sulla densità di probabilità dei pesi, non dipendente direttamente dalla funzione di costo. Infine si discutono le implicazioni di questa inferenza variazionale ottenuta dal SGD.
24
Sievert, Rolf. "Instance Segmentation of Multiclass Litter and Imbalanced Dataset Handling : A Deep Learning Model Comparison." Thesis, Linköpings universitet, Datorseende, 2021. http://urn.kb.se/resolve?urn=urn:nbn:se:liu:diva-175173.
Full textAbstract:
Instance segmentation has a great potential for improving the current state of littering by autonomously detecting and segmenting different categories of litter. With this information, litter could, for example, be geotagged to aid litter pickers or to give precise locational information to unmanned vehicles for autonomous litter collection. Land-based litter instance segmentation is a relatively unexplored field, and this study aims to give a comparison of the instance segmentation models Mask R-CNN and DetectoRS using the multiclass litter dataset called Trash Annotations in Context (TACO) in conjunction with the Common Objects in Context precision and recall scores. TACO is an imbalanced dataset, and therefore imbalanced data-handling is addressed, exercising a second-order relation iterative stratified split, and additionally oversampling when training Mask R-CNN. Mask R-CNN without oversampling resulted in a segmentation of 0.127 mAP, and with oversampling 0.163 mAP. DetectoRS achieved 0.167 segmentation mAP, and improves the segmentation mAP of small objects most noticeably, with a factor of at least 2, which is important within the litter domain since small objects such as cigarettes are overrepresented. In contrast, oversampling with Mask R-CNN does not seem to improve the general precision of small and medium objects, but only improves the detection of large objects. It is concluded that DetectoRS improves results compared to Mask R-CNN, as well does oversampling. However, using a dataset that cannot have an all-class representation for train, validation, and test splits, together with an iterative stratification that does not guarantee all-class representations, makes it hard for future works to do exact comparisons to this study. Results are therefore approximate considering using all categories since 12 categories are missing from the test set, where 4 of those were impossible to split into train, validation, and test set. Further image collection and annotation to mitigate the imbalance would most noticeably improve results since results depend on class-averaged values. Doing oversampling with DetectoRS would also help improve results. There is also the option to combine the two datasets TACO and MJU-Waste to enforce training of more categories.
25
Chen, Kuang-Yu, and 陳廣瑜. "Deep Learning Model Compression with Information Guide." Thesis, 2017. http://ndltd.ncl.edu.tw/handle/w26d3z.
Full text
26
WU, GUAN-WEI, and 吳冠瑋. "Applying Deep Learning Model to Medicine Discrimination." Thesis, 2019. http://ndltd.ncl.edu.tw/handle/sz7t8v.
Full textAbstract:
碩士逢甲大學
資訊工程學系
107
Medicines dispensing verification means that whether the medicines in the patient’s medicine bag are the same as the prescription prescribed by the doctor. The correct of medicines dispensing can be said to be the most basic safe drug condition for the patient. However, there are various kinds medicines but they may have a similar appearance and packaging. Therefore, the possibility of human error is greatly increased. Although many medical units have set up an online medicines discriminator system and provide people with using text or photos to get information about medicines, they are not often used. In order to implement the patients medication safety, this study proposes a new medicines discrimination system based on computer vision. When the pharmacist is dispensing medicines, we use the image recognition model with a high recognition ability to monitor the dispensing table to assist the pharmacist in dispensing the medicine, thereby ensuring the contents of the patients medicine bag is correct. This system is an object detection model based on YOLOv3. It has three major features: First, the recognition speed is fast, it can react to the situation of the dispensing table in real time. Second, it is easy to use and easy to deploy to the current dispensing units. Third, this system is proactive and requires no frequent operation, does not impose an additional burden on the pharmacist. In this study, a deep learning-based image recognition technology was used, and 380 medicines images were used as training data to propose a medicine discrimination system based on deep learning.
27
Hsu, Chun-Wei, and 徐莙惟. "Deep Learning Enabled Process Independent Lithographic Model." Thesis, 2019. http://ndltd.ncl.edu.tw/handle/utrw8z.
Full text
28
Liu, Zheng-Wei, and 劉政威. "Waterfall Model for Deep Reinforcement Learning Based Scheduling." Thesis, 2019. http://ndltd.ncl.edu.tw/handle/a3yn5q.
Full textAbstract:
碩士國立中央大學
通訊工程學系在職專班
107
The fourth generation of communication systems has been able to meet the multimedia application needs of mobile devices. Through the scheduling service provided by the base station, the user equipment can obtain the data packets required by the downlink of the communication system to meet and obtain better application services, so the channel resources are allocated and the calculation of the user group scheduling service is provided. The law is quite critical. This paper implements a mobile communication scheduling learning platform, and proposes a Deep Deterministic Policy Gradient model. The waterfall model concept is used to analyze the scheduling algorithm flow into three stages: sorting selection, resource evaluation and channel allocation. A waterfall scheduling method that enables more data throughput per unit time and meets more user needs in the current communication environment. The mobile communication scheduling learning platform is composed of six modular components: base station and channel resources, enhanced learning neural network, user equipment attributes, application service types, environmental information and reward functions, and phase micro-algorithms and dependency injection. . Using inversion control and dependency injection to reduce platform software coupling, it is quite easy to maintain the stage micro-algorithm and the six module components.
29
Yen, Yi-Tung, and 顏逸東. "Enhanced Car-Following Model with Deep Reinforcement Learning." Thesis, 2019. http://ndltd.ncl.edu.tw/handle/499erb.
Full textAbstract:
碩士國立臺灣大學
資訊工程學研究所
107
With the rapid evolution of artificial intelligence and technology, autonomous vehicle is regarded as the future of transportation. One of the important functions that autonomous vehicle should be equipped is a well-designed car-following model. With a well-designed car-following model, autonomous vehicle can drive in a safe, comfortable and efficient manner. This will increase driving safety, passenger comfort and improve road efficiency. This thesis designs and implements an enhanced car-following model. According to the laws, regulations and standards, we modeled the safety, comfort and efficiency into quantified reward functions. Using reinforcement learning, the network agent learns the best policy to achieve the maximum reward by repeated the learning process. The evaluation results show that our model not only reduces the number of inefficient and dangerous headways but also eliminates the jerk to achieve more efficient and comfortable driving than human drivers. The model outperformed 79% human drivers in public dataset. The achieved efficiency is 98% of the optimal bound. Furthermore, compared to the SUMO’s ACC model, given the same number of departed vehicles, our model enables more arrived vehicles and higher average speed to improve the overall road capacity.
30
Shukla, Aditya. "Model Extraction and Active Learning." Thesis, 2020. https://etd.iisc.ac.in/handle/2005/4420.
Full textAbstract:
Machine learning models are increasingly being offered as a service by big companies such as Google, Microsoft and Amazon. They use Machine Learning as a Service (MLaaS) to expose these machine learning models to the end-users through cloud-based Application Programming Interface (API). Such APIs allow users to query ML models with data samples in a black-box fashion, returning only the corresponding output predictions. MLaaS models are generally monetized by billing the user for each query made. Prior work has shown that it is possible to extract these models. They developed model extraction attacks that extract an approximation of the MLaaS model by making black-box queries to it. However, none of them satisfy all the four criteria essential for practical model extraction: (i) the ability to extract deep learning models, (ii) non-requirement of domain knowledge, (iii) the ability to work with a limited query budget and (iv) non-requirement of annotations. In collaboration with Pal et al., we propose a novel model extraction attack that makes use of active learning techniques and unannotated public data to satisfy all the aforementioned criteria. However, as we show in the experiments, no one active learning technique is well-suited for different datasets and under different query budget constraints. Given the plethora of active learning techniques at the adversary’s disposal and the black-box nature of the model under attack, the choice of the technique to be used is difficult but integral: the chosen technique is a strong determinant of the quality of the extracted model. In this work, we wish to devise an active learning technique that combines the benefits of existing active learning techniques, as applicable to different budgets and different datasets, yielding on average extracted models that exhibit a high-test agreement with the MLaaS model. In particular, we show that a combination of the DFAL technique of Ducoffe et al. and the Coreset technique of Sener et al. is able to leverage the benefits of both the base techniques, outperforming both DFAL and Coreset in a majority of our experiments. The model extraction attack using this technique achieves, on average, a performance of 4.70× over uniform noise baseline by using only 30% (30,000 data samples) of the unannotated public data. Moreover, the attack using this technique remains undetected by PRADA, a state-of-the-art model extraction detection method.
31
CHEN, CHIA-HSI, and 陳家羲. "Develop Forecasting Model for Financial Crisis with Deep Learning." Thesis, 2018. http://ndltd.ncl.edu.tw/handle/3kp4yb.
Full textAbstract:
碩士東吳大學
會計學系
106
Financial crisis forecasting of a company is extremely important for both investigators and company manager. For investigators, they are able to take actions before the crisis happens and therefore prevent assets loss. For company managers, they can adjust the management policy or direction according to the forecasting results so that the crisis would not happen.In this paper, we develop a deep neural network and train it using TEJ Financial database to obtain a forecasting model for financial crisis. Our model outperforms traditional shallow neural network by 10% in terms of both test and prediction accuracy. The prediction accuracy of our model is up to 70%. On the other hand, we also find that the data and model can be further improved to train more complex model and deal with time-series data, which may result in more accurate model.
32
Liu, Wen-Cheng, and 劉文誠. "A Web Service for Automatic Deep Learning Model Generation." Thesis, 2019. http://ndltd.ncl.edu.tw/handle/utec82.
Full textAbstract:
碩士國立中央大學
資訊工程學系
107
As technology advances, deep learning has changed the way many industries produce, such as detecting defects, identifying objects, and so on. The core network model is the core of the algorithm and the essence of the training after big data. However, for most operators, how to train a usable model from scratch is a major difficulty in introducing artificial intelligence on the production line. How to quickly and easily complete a usable deep learning model becomes an issue that most non-employed workers want to know. Usually, training a highly accurate deep learning model requires a complex network architecture in addition to a large amount of data, and can be completed after numerous fine-tuning. The acquisition of data is relatively easy for the production line operators, and the network architecture needs to take time to understand the details. It is not completed in a moment and a half, and the threshold for entry is relatively improved, which is not conducive to industrial upgrading of various factories. This study combines the deep learning model suite Keras with the web language of the client and server side to provide a web-based artificial intelligence system that can quickly train deep learning models. The system allows users to set parameters and upload training data through the graphical interface, so that users in non-employed fields can quickly train the required models without spending too much time on the details of deep learning.
33
Pradhan, Sipun Kumar, and 施庫瑪. "A Rapid Deep Learning Model for Goal-Oriented Dialog." Thesis, 2016. http://ndltd.ncl.edu.tw/handle/6y6zpd.
Full textAbstract:
碩士國立中央大學
資訊工程學系
104
Open-domain Question Answering (QA) systems aim at providing the exact answer(s) to questions formulated in natural language, without restriction of domain. My research goal in this thesis is to develop learning models that can automatically induce new facts without having to be re-trained, in particular its structure and meaning in order to solve multiple Open-domain QA tasks. The main advantage of this framework is that it requires little feature engineering and domain specificity whilst matching or surpassing state-of-the-art results. Furthermore, it can easily be trained to be used with any kind of Open-domain QA. I investigate a new class of learning models called memory neural networks. Memory neural networks reason with inference components combined with a long-term memory component; they learn how to use these jointly. The long-term memory can be read and written to, with the goal of using it for prediction. I investigate these models in the context of question answering (QA) where the long-term memory effectively acts as a (dynamic) knowledge base, and the output is a textual response. Finally, I show that an end-to-end dialog system based on memory neural networks can reach promising and learn to perform non-trivial operations. I confirm those results by comparing my system to various well-crafted baseline Datasets and future work is discussed.
34
Tsao, Yeh-Wen, and 曹爗文. "A Fast Deep Learning Model for Time Series Prediction." Thesis, 2019. http://ndltd.ncl.edu.tw/handle/94h345.
Full textAbstract:
碩士國立臺灣大學
資訊工程學研究所
107
Time series forecasting is an important research area across many domains, such as predictions of financial market, weather, electricity consumption, and traffic jam situation. However, most of recent works are usually time-consuming and complex. In this paper, we propose a deep learning model to tackle this issue, and deliver efficient performance. Our model uses purely Convolutional Neural Network (CNN) structure to capture both long-term and short-term features. Thorough empirical studies based upon the total seven different dataset demonstrate that the our model can outperform state-of-the-art methods over training time with comparable performance.
35
Chang, Chao-Mei, and 張昭美. "Taiwanese speech commands recognition model based on deep learning." Thesis, 2019. http://ndltd.ncl.edu.tw/handle/knb7ws.
Full textAbstract:
碩士國立交通大學
資訊學院資訊學程
107
Most of the recent machine learning papers are aimed at images and videos, such as face recognition, large image database identification, unmanned autopilot, object recognition, AlphaGo, object movement trajectory prediction, changing image style and creating virtual portrait style-based GAN. However, due to the development trend of voice assistants, it is necessary to closely cooperate with local language materials and culture habits. Therefore, the focus is on local language audio processing and machine learning . Benefiting from the prosperous deep learning progress, diverse languages are no longer an obstacle to communication, but rather a manifestation of diverse cultures. It is time to pay attention to regional languages such as Taiwanese. The paper uses a variety of audio pre-processing and CNN, LSTM, GRU and other attempts to use the depth model. 1. Taiwanese speech command recognition. 2. Specific Taiwanese speech keyword trigger. 3. Identify the audio segment between Mandarin and Taiwanese. 4. Use AI to write Taiwanese local drama. Finally, it is applied to the Android mobile app, so that the user can use the voice "góabeh" (I want to) "khuànn-siòng-phìnn" (See photo), "thian-im-ga̍k" (Listen to music), "khà-tiān-uē" (Make a phone call), "hip-siàng" (Photograph) evokes the corresponding application.
36
Lin, Yi-Hsiu, and 林怡秀. "Question Generation from Knowledge Base Using Deep Learning Model." Thesis, 2018. http://ndltd.ncl.edu.tw/handle/mkj9vx.
Full textAbstract:
碩士國立交通大學
資訊科學與工程研究所
106
With the advancement of data-driven approach, the lack of corpora has become the main obstacle of the natural language processing research. Compared with English corpora, publicly available Mandarin corpora is even more lacking. Our paper purposes to solve this problem by using existing question answering dataset and knowledge base to create a new Mandarin question answering dataset. In this study, we first collect the data from CN-DBpedia and question answering dataset from WebQA and web crawler, and propose a method to combine them in the form of pairs as our training data, and then using sequence-to-sequence model to generate questions from knowledge base. The generated questions then incorporate with entities in knowledge base as the answers to create a new Mandarin question answering dataset. In our experiment, we develop a template-based question generation baseline in order to evaluate our model by human evaluation. Our model achieves an acceptable performance compare to the template-based baseline.
37
BAINS, INDERPREET SINGH. "WEB SECURITY IN IoT NETWORKS USING DEEP LEARNING MODEL." Thesis, 2020. http://dspace.dtu.ac.in:8080/jspui/handle/repository/18061.
Full textAbstract:
The vision of IoT is to interface day by day utilized items (which have the capacity of detecting
and activation) to the Internet. This may or might possibly include human. IoT field is as yet
developing and has many open issues. We develop on the digital security issues. The Web of
things (IoT) is still in its beginning phases and has pulled in much enthusiasm for some
mechanical parts including clinical fields, coordination’s following, savvy urban communities
and autos. Anyway, as a paradigm, it is defenseless to a scope of significant intrusion threats. In
IoT whenever there is a web attack then we need to remove the attack by installing software so
by using these models we can remove the attack from the system. It presents a threat
investigation of the IoT and uses an Artificial Neural Network (ANN) to battle these threats. In
this, profound learning method for digital security and prevention of attacks is used in which a
convolution 1d with multiple convolutions is used to increase the accuracy of the user. We have
proposed profound models of learning and assessed those utilizing most recent CICIDS2017
datasets for DDoS assault recognition which has given most noteworthy precision as 99.38%. It
is essential to create an effective intrusion discovery framework which uses deep learning
mechanism to overcome attack issues in IOT framework. In this, a CNN i.e convolutional neural
system is developed with various convolution layers and accuracy of attack detection is
increased.
38
"Efficient and Online Deep Learning through Model Plasticity and Stability." Doctoral diss., 2020. http://hdl.handle.net/2286/R.I.62959.
Full textAbstract:
abstract: The rapid advancement of Deep Neural Networks (DNNs), computing, and sensing technology has enabled many new applications, such as the self-driving vehicle, the surveillance drone, and the robotic system. Compared to conventional edge devices (e.g. cell phone or smart home devices), these emerging devices are required to deal with much more complicated and dynamic situations in real-time with bounded computation resources. However, there are several challenges, including but not limited to efficiency, real-time adaptation, model stability, and automation of architecture design.
To tackle the challenges mentioned above, model plasticity and stability are leveraged to achieve efficient and online deep learning, especially in the scenario of learning streaming data at the edge:
First, a dynamic training scheme named Continuous Growth and Pruning (CGaP) is proposed to compress the DNNs through growing important parameters and pruning unimportant ones, achieving up to 98.1% reduction in the number of parameters.
Second, this dissertation presents Progressive Segmented Training (PST), which targets catastrophic forgetting problems in continual learning through importance sampling, model segmentation, and memory-assisted balancing. PST achieves state-of-the-art accuracy with 1.5X FLOPs reduction in the complete inference path.
Third, to facilitate online learning in real applications, acquisitive learning (AL) is further proposed to emphasize both knowledge inheritance and acquisition: the majority of the knowledge is first pre-trained in the inherited model and then adapted to acquire new knowledge. The inherited model's stability is monitored by noise injection and the landscape of the loss function, while the acquisition is realized by importance sampling and model segmentation. Compared to a conventional scheme, AL reduces accuracy drop by >10X on CIFAR-100 dataset, with 5X reduction in latency per training image and 150X reduction in training FLOPs.
Finally, this dissertation presents evolutionary neural architecture search in light of model stability (ENAS-S). ENAS-S uses a novel fitness score, which addresses not only the accuracy but also the model stability, to search for an optimal inherited model for the application of continual learning. ENAS-S outperforms hand-designed DNNs when learning from a data stream at the edge.
In summary, in this dissertation, several algorithms exploiting model plasticity and model stability are presented to improve the efficiency and accuracy of deep neural networks, especially for the scenario of continual learning.Dissertation/Thesis
Doctoral Dissertation Electrical Engineering 2020
39
Frazão, Xavier Marques. "Deep learning model combination and regularization using convolutional neural networks." Master's thesis, 2014. http://hdl.handle.net/10400.6/5605.
Full textAbstract:
Convolutional neural networks (CNNs) were inspired by biology. They are hierarchical neural
networks whose convolutional layers alternate with subsampling layers, reminiscent of simple
and complex cells in the primary visual cortex [Fuk86a]. In the last years, CNNs have emerged
as a powerful machine learning model and achieved the best results in many object recognition
benchmarks [ZF13, HSK+12, LCY14, CMMS12].
In this dissertation, we introduce two new proposals for convolutional neural networks. The
first, is a method to combine the output probabilities of CNNs which we call Weighted
Convolutional Neural Network Ensemble. Each network has an associated weight that makes
networks with better performance have a greater influence at the time to classify a pattern
when compared to networks that performed worse. This new approach produces better results
than the common method that combines the networks doing just the average of the output
probabilities to make the predictions. The second, which we call DropAll, is a generalization
of two well-known methods for regularization of fully-connected layers within convolutional
neural networks, DropOut [HSK+12] and DropConnect [WZZ+13]. Applying these methods
amounts to sub-sampling a neural network by dropping units. When training with DropOut, a
randomly selected subset of the output layer’s activations are dropped, when training with
DropConnect we drop a randomly subsets of weights. With DropAll we can perform both methods
simultaneously. We show the validity of our proposals by improving the classification error on a
common image classification benchmark.
40
HUANG, PO-YU, and 黃柏毓. "Predicting Social Insurance Payment Behavior Based on Deep Learning Model." Thesis, 2017. http://ndltd.ncl.edu.tw/handle/04210692742079168717.
Full textAbstract:
碩士逢甲大學
資訊工程學系
105
The social insurance is an important part of the social security system. In Taiwan, the social insurance system is classified by occupational groups and managed by different government agencies. According to the Executive Yuan of Taiwan, this pension system includes five separate social insurance programs covering public servants and teachers, laborers, military personnel, farmers, and a national pension insurance program for those not covered by the above four employment-based categories. Ministry of Health and Welfare in Taiwan is responsible for many types of social insurances like National Pension Insurance, National Health Insurance and Long-term Care Services Program. In addition, Ministry of Health and Welfare provides subsidized health insurance coverage for the underprivileged and ensures that senior citizens with no employment-based retirement benefits will still have the basic economic necessities in their elderly life. Unfortunately, most social insurances are impacted by various problems and have been facing the crisis of pension bankruptcy. Althougth the traditional actuarial methods use many hypotheses to analyze cash flow, they mostly focus on trend analysis with a macro view of the participants. Due to a large number of the insured, it is very hard to predict the payment behavior of each individual. To make better prediction, we propose to build payment behavior models based on machine learning technology to predict personal payment behavior accurately. Using the number of the participants for each personal payment behavior and corresponding insurance premiums, we can make better cash flow prediction in order to help the social insurance operations become sustainable. This research uses the seven year's data from Taiwan's National Pension Insurance as the source of experimental data. With the implementation of deep learning model, we could analyze and predict the future payment behaviors of the insured.
41
Lee, Yi-Nan, and 李奕男. "Deep Visual Semantic Transform Model Learning from Multi-Label Images." Thesis, 2017. http://ndltd.ncl.edu.tw/handle/48kv54.
Full textAbstract:
碩士國立臺灣師範大學
資訊工程學系
105
Learning the relation between images and text semantics has been an important problem in the field of machine learning and computer vision. This paper addresses this problem. We observe that there is a semantic relation between texts, for example, “sky” and “cloud” have a close semantic relation, and “sky” and “car” have a weak semantic relation. We suppose the semantic relation between texts can be different depending on images. For example, an image contains both sky and car. The word “sky” and “car” are initially semantically irrelevant, but may have a connection because of the image containing these concepts. Therefore, we propose a Convolutional Neural Network based model to link the semantic relation between an image and its text labels. The main difference between our work and existing visual semantic embedding models is that the output of our model is a linear transformation function. In other words, each input image is treated as a function to determine the relation between each word and the image, and to predict the possible labels for the image. Finally, this model is validated on the NUS-WIDE dataset and the experimental results show that the model has a great performance on predicting labels for images.
42
LEE, JHONG-TING, and 李仲庭. "Apply TensorFlow deep learning model for time series forecasting problem." Thesis, 2018. http://ndltd.ncl.edu.tw/handle/y6cses.
Full textAbstract:
碩士開南大學
資訊學院碩士在職專班
106
This study takes the TensorFlow as a backend engine for deep learning. The Multi-Layer Perceptron (MLP) is built to solve the time series forecasting problems. The case study are daily stock closing prices in Taiwan, i.e. the Taiwan Semiconductor Manufacturing Company Limited (TSMC), Uni-President Enterprises Corporation (Uni-President), and Largan Precision Company Limited (LARGAN Precision). We collect 120 daily records from 2017/01/03 to 2017/07/04. Around 20 input features we used are: the Trade Volume, the Trade Value, the Opening Price, the Highest Price, the Lowest Price, the Closing Price, the Delta Price, the Transaction amount, and other Technical Analysis Indicators. Then, the Stepwise Regression Analysis is adopted as a filter for screening out some input features really correlative to the Label (the predict closing price). The numerical results are summarized as follows: the Mean Absolute Percentage Error (MAPE) and Standard Deviation (SD) in training and predicting stages for the TSMC are (0.17%, 0.06) and (0.33%, 0.05); (0.15%, 0.06) and (0.20%, 0.04) for the Uni-President; (0.35%, 0.10) and (0.37%, 0.05) for the LARGAN Precision. Keywords: Deep learning, Time series, TensorFlow, Stock price prediction
43
LAI, HUNG-JU, and 賴泓儒. "STT-MRAM Co-design Deep Learning Model for IoT Applications." Thesis, 2019. http://ndltd.ncl.edu.tw/handle/c2a8z5.
Full textAbstract:
碩士逢甲大學
通訊工程學系
107
At present, there are fewer STT-MRAM applications. The mian memories on the current market are SRAM, DRAM, and Flash Memory. However, these memories consume more power than STT-MRAM, which are not suitable for resource-limited IoT devices. The memories equipped with IoT devices must be low energy consumption, rapidly operatrion , access endurace, and samll area. STT-MARM just meets these requirements. In particular, STT-MARM is non-volatile. After the power turns off, the data are still reserved. Therefore, it is an emergent memory for IoT applications. In this thesis, we propose an application architecture of STT-MRAM as a deep learning model, which is implemeted on FPGA, and then use Qsys (SOPC) architecture to implement a CNN neural network. Finally, we use the MNIST data set to evaluate the performance and accuracy of STT-MRAM co-design deep learning model for IoT applications, compared with other memories SRAM, SDRAM, and MRAM.
44
Chung, Hao-Ting, and 鐘皓廷. "Building Student Course Performance Prediction Model Based on Deep Learning." Thesis, 2018. http://ndltd.ncl.edu.tw/handle/m2z8n3.
Full textAbstract:
碩士國立臺北科技大學
資訊工程系
106
The deferral of graduation rate in Taiwan’s universities is estimated 16%, which will affect the scheduling of school resources. Therefore, if we can expect to take notice of students’ academic performance and provide guidance to students who cannot pass the threshold as expected, we can effectively reduce the waste of school resources. In this research, we use recent years’ student data attributes and course results as training data to construct student performance prediction model. The K-Means algorithm was used to classify all courses from the freshman to the senior. The related courses will be grouped in the same cluster, which will more likely to find similar features and improve the accuracy of the prediction. Then, this research constructs independent neural networks for each course according to the different academic year. Each model will be pre-trained by using De-noising Autoencoder. After pre-training, the corresponding structure and weights are taken as the initial value of the neural network model. Each neural network is treated as a base predictor. All predictors will be integrated into an Ensemble predictor according to different years’ weights to predict the current student’s course performance. As the students finish the course at the end of each semester, the prediction model will continue track and update to enhance model accuracy through online learning.
45
Teng, Yu-Han, and 鄧鈺翰. "Using a multimodal architecture Research on Deep Learning Model Analysis." Thesis, 2019. http://ndltd.ncl.edu.tw/handle/7het95.
Full textAbstract:
碩士國立中央大學
資訊管理學系
107
With the popularity of social networks and e-commerce sites, users have switched from passively receiving messages to actively disseminating messages. The value of comments and online messages is also becoming more and more important. Analysis and research over the past few years. Trying to analyze trends about specific product products, topics, reviews, and tweets. Play an important role in all aspects. This study uses different vectorization processes to verify the multimodal analysis model and confirm that the model can effectively improve the accuracy. This study proposes a combination of two models. This feature is combined with deep learning neural network construction to build a multimodal analysis model. Model 1 is a deep learning model based on Glove vector, attention mechanism and GRU neural network architecture. Model 2 is a deep learning model based on Word2Vec vector, attention mechanism and CNN neural network architecture. Multimodal analysis model is validated by K-fold cross validation and F1 measurement method. The experimental results prove that the multimodal analysis model proposed in this study has higher accuracy than related research. Using the high-level multi-modal combination method, the features of multiple models are extracted and combined to form a combined feature, and this feature is trained in neural network. The feature set can be mutually assisted, and the accuracy can be 91.56% through the two vectors and the optimal neural network architecture combined with the multi-modal method. And the model verification shows 93% verification value, which proves that the multimodal analysis model proposed in this study is used in the field of comment texts, which can effectively improve the accuracy of model prediction and improve its accuracy.
46
Huang, Chu-Chih, and 黃炬智. "Classification of Chinese Articulation Disorder based on Deep Learning Model." Thesis, 2019. http://ndltd.ncl.edu.tw/handle/3v7km6.
Full textAbstract:
碩士國立臺灣科技大學
電子工程系
107
Articulation disorder means having difficulties during pronunciations, leading to incorrect articulations and unclear sentences. Articulation disorder has been a common child language issue. Currently, there is no any unified sayings for articulation disorder's classification in the Taiwan's medical field. Thus, a speech therapist is required for analysis and treatment in hospitals. After a series of pronunciations, a speech therapist will make an analysis based on children's pronunciations. Children will return to the hospitals for months continuously to improve their conditions. Nevertheless, the treatment can only benefit children with articulation disorder by receiving treatments in hospitals, slowing down the treatment cycle. The purpose of this work is to automate the diagnosis for articulation disorder by combining the latest AI's convolutional neural network (CNN). Results show that LeNet-5 which achieved 94.56 Top-1 accuracy and 0.995 avg F1-score with the smallest model size is more suitable to apply articulations disorder application on mobile devices.
47
LIN, HAN-LONG, and 林翰隆. "Building Graduate Salary Grading Prediction Model Based on Deep Learning." Thesis, 2019. http://ndltd.ncl.edu.tw/handle/z4hkqx.
Full textAbstract:
碩士國立臺北科技大學
資訊工程系
107
This paper used deep learning to build a salary grading prediction model. Due to the order relationship between each grading of salary grading, this paper regards this kind of problem as an ordinal regression problem. This paper used multiple output deep neural network to solve the ordinal regression problem so that the network learns the correlation between these salary grading during training. This model is pre-trained using Stacked De-noising Autoencoder. After pre-training, the corresponding weights are taken as the initial weights of neural network. During training, this paper used the Dropout and Bootstrap Aggregating to improve model performance. This model used the graduates’ personal information, grades, and family data as input feature, and predict salary grading of graduating or graduated students. This result will be provided to the school researchers to grasp the salary trend.
48
Lin, Yung-Chien, and 林詠謙. "Predicting the Computation Time of Deep Learning Model on Accelerators." Thesis, 2019. http://ndltd.ncl.edu.tw/handle/w2wncd.
Full textAbstract:
碩士國立臺灣大學
資訊工程學研究所
107
With the rapid development of deep learning, in order to improve the efficiency of implementation, the hardware for deep learning is becoming more and more important, but the platform with higher performance is often accompanied by high prices. Therefore, the goal of this research is to let users can quickly calculate the performance of a system, and even can easily analyze its performance before getting the target hardware. There are a lot of related researches at present, but most of them use formulas to make performance predictions, and this method often uses linear methods to do calculations, so many details are ignored. The method used in this study is to collect enough relevant data and use deep learning to learn the computation time under different configurations. This study also predicts different neural networks, even for unavailable hardware.
49
Gutta, Sreedevi. "Improving photoacoustic imaging with model compensating and deep learning methods." Thesis, 2018. https://etd.iisc.ac.in/handle/2005/4390.
Full textAbstract:
Photoacoustic imaging is a hybrid biomedical imaging technique combining optical ab-
sorption contrast with ultrasonic resolution. It is a non-invasive technique that is scalable
to reveal structural, functional, and molecular information of the tissue under investiga-
tion. The important step in photoacoustic tomography is image reconstruction, which
enables quanti cation of tissue functional properties. The photoacoustic image recon-
struction problem is typically ill-posed and requires an utilization of regularization to
provide meaningful results. The aim of this thesis work is to develop methods that
can improve photoacoustic image reconstruction, especially in realistic imaging scenar-
ios, where the utility of standard image reconstruction methods is limited in terms of
providing good quality photoacoustic images.
The photoacoustic image reconstruction problem is typically solved using either
weighted or ordinary least squares (LS), with regularization term being added for stabil-
ity, which account only for data imperfections (noise). Numerical modeling of acoustic
wave propagation requires discretization of imaging region and is typically developed
based on many assumptions, such as speed of sound being constant in the tissue, making
it imperfect. Two variants of total least squares (TLS) were proposed, namely ordinary TLS and Sparse TLS, which account for model imperfections. The ordinary TLS
is implemented in the Lanczos bidiagonalization framework to make it computationally
efficient. The Sparse TLS utilizes the total variation penalty to promote recovery of
high frequency components in the re- constructed image. The Lanczos truncated TLS
(Lanczos T-TLS) and Sparse TLS methods were compared with the recently established
state-of-the-art methods, such as Lanczos Tikhonov and Exponential Filtering. The TLS methods exhibited better performance for experimental data as well as in cases where
modeling errors were present, such as few acoustic detectors malfunctioning and speed
of sound variations. Also, the TLS methods do not require any prior information about
the errors present in the model or data, making it attractive for real-time scenarios.
The model-based reconstruction methods, such as Tikhonov regularization scheme,
require an appropriate selection of explicit regularization parameter, which is a com-
putationally expensive procedure. The Tikhonov scheme promotes the smooth features
in the reconstructed image due to the smooth regularizer, thus leading to loss of sharp
features. A simple and computationally efficient extrapolation method was developed,
which provides the solution at zero regularization, by assuming that the solution is a
function of regularization. The reconstructed results using this method were shown in
three variants (Lanczos, Traditional, and Exponential) of Tikhonov ltering on numer-
ical and experimental phantom data. The proposed extrapolation method performance
was shown to be superior than the standard error estimate technique with an added
advantage of being atleast four times faster in terms of computation, and providing an
improvement as high as 2.6 times in terms of standard gures of merit.
Photoacoustic signals collected at the boundary of tissue are always band-limited.
A deep neural network (DNN) with ve fully connected layers (similar to the decoder
network) was proposed to enhance the bandwidth of the detected photoacoustic signal,
thereby improving the quantitative accuracy of the reconstructed photoacoustic images.
A least square based deconvolution method that utilizes the Tikhonov regularization
framework was used for comparison with the proposed network. The DNN-based method
was evaluated using both numerical and experimental data. The results show that the
DNN-based method was capable of enhancing the bandwidth of the detected photoa-
coustic signal, which in turn improves the contrast recovery and quality of reconstructed
photoacoustic images without adding any signi cant computational burden.
Analytical photoacoustic image reconstruction methods such as back-projection re-
quire large amount of data for accurate reconstruction of initial pressure distribution.
Model-based iterative algorithms are proven to provide quantitatively accurate recon-
structions compared to analytical methods in limited data cases. These methods start
from an initial guess of the solution (obtained through analytical methods) and itera-
tively improve the solution via applying regularization. These are challenging to deploy
in real-time due to their high computational complexity and also difficulty in choosing
optimal reconstruction parameters. A deep convolutional neural network, with archi-
tecture similar to SRGAN, a generative adversarial network (GAN) to obtain images
of super resolution (SR), was utilized in the photoacoustic image reconstruction pro-
cess to provide desired image characteristics obtainable by model-based algorithms with
computation effciency equal to analytical methods. The network was trained with back-
projected reconstruction as input and output being ground truth image. The proposed
method was evaluated using both numerical and experimental phantoms and was shown
to be superior compared to the state-of-the-art model-based methods. Moreover, the
proposed method takes approximately one second on the GPU, making the approach
attractive in real-time.
50
Pereira, Carlos Manuel Silva. "Deep learning techniques for grapevine variety classification using natural images." Doctoral thesis, 2020. http://hdl.handle.net/10348/9969.
Full textAbstract:
Doctoral Thesis in InformaticsThis thesis proposes a computer vision system for automatic classification of six endogenous grapevine varieties of the Portuguese Douro Demarcated Region from natural images. For a full understanding about the applied methodologies and the developed experiments in this research work, we structured this document into six sections. The first ones are reserved for the revision of the literature about image processing in agriculture, such as, image processing techniques (enhancement and color model conversion) and image segmentation methods that inspired us to develop the proposed leaf segmentation algorithm. The theoretical background about the machine learning process, namely, deep learning and convolutional neural networks were presented for an easier understanding of the methodologies applied on our research proposal. The remaining ones are reserved for the presentation of the materials and methods, the major conclusions and possible future work developments. Our proposed system is hard to develop because it presents many constraints. First, the presence in natural vineyard images of savage foliage, weed, multiple leaves with overlapping, occlusion, and obstruction by objects due to the shadows, dust, insects and other adverse climatic conditions that occur in natural environment at the moment of image capturing; second, high similarity of the images among different grapevine varieties; third, leaf senescence and significant changes on the grapevine leaf and bunches images in the harvest seasons, namely, due to adverse climatic conditions, diseases and presence of pesticides; fourth, the low volume of images available. In addition, the vineyards of the Douro Region are also characterized for having more than one grapevine variety per parcel and even for row. Knowing the susceptibility of a particular variety to a specific disease, its identification using this automatic system, will help, for example, in a more specific and targeted treatment. Besides that, many wine producers are entitled to this large number of grapevine varieties to produce their most expensive wines. As the title of this thesis highlights, the deep learning techniques were used to solve the presented constratints. With this advanced neural technologies, the performance of transfer learning schemes based on AlexNet architecture was evaluated for classification of grapevine varieties using diverse pre-processed datasets. Thus, two natural vineyard image datasets were constructed from which different pre-processed datasets are generated with the application of some image processing methods, including a proposed four-corners-in-one image warping algorithm for deep training purposes. After detailing some network schemes, we present and discuss some of the experimental results obtained by the proposed approach, which we judge promising and encouraging to help Douro wine growers in the automatic grapevine varieties classification for future implementation of a robotic grape harvest.
Esta tese propõe um sistema de visão computacional para classificação automática de seis variedades de videira endógenas da Região Demarcada do Douro a partir de imagens naturais. Para um completo entendimento sobre as metodologias aplicadas e as experièncias desenvolvidas neste trabalho de investigação, estruturamos este documento em seis capítulos. Os primeiros são reservados à revisão da literatura sobre processamento de imagens na agricultura, como técnicas de processamento de imagens (realce da imagem e conversão de modelos de cores) e métodos de segmentação de imagens que nos inspiraram a desenvolver um algoritmo de segmentação de folhas. Os fundamentos teóricos sobre o processo de aprendizagem de máquina, a saber, aprendizagem profunda e redes neuronais convolucionais, são apresentados para facilitar a compreensão das metodologias aplicadas na nossa proposta de trabalho. Os demais capítulos ficam reservados para a apresentação dos materiais e métodos, as principais conclusões e possíveis desenvolvimentos futuros do trabalho. Torna-se difícil desenvolver o sistema que se propõe porque apresenta muitos constrangimentos. Primeiro, a presença em imagens naturais de vinhas de folhagem selvagem, erva daninha, várias folhas com sobreposição, oclusão e obstrução por objetos devido às sombras, poeira, insetos e outras condições climáticas adversas que ocorrem no ambiente natural no momento da captação de imagem; segundo, a alta similaridade das imagens entre diferentes variedades de videira; terceiro, senescência foliar e mudanças significativas nas imagens de folhas e cachos de videira nas safras, devido a condições climáticas adversas, doenças e presença de pesticidas; quarto, o baixo volume de imagens disponíveis. Além disso, as vinhas da região do Douro também se caracterizam por possuir mais de uma variedade de videira por parcela e até por bardo. Conhecendo a suscetibilidade de uma variedade específica a uma doença específica, usando este sistema automático, a sua identificação ajudará, por exemplo, num tratamento mais específico e direcionado. Além disso, muitos produtores de vinho têm utilizado um grande número de variedades de videira para produzir os seus vinhos de referência e, portanto, mais caros. Como o título desta tese destaca, as técnicas de aprendizagem profunda foram usadas para resolver os constrangimentos apresentados. Com estas tecnologias neuronais avançadas, o desempenho dos esquemas de aprendizagem de transferência baseados na arquitetura AlexNet foi avaliado na classificação de variedades de videira usando diversos conjuntos de dados pré-processados. Assim, foram construídos dois conjuntos de dados de imagem de vinhas naturais a partir dos quais foram gerados diferentes conjuntos de dados pré-processados com a aplicação de alguns métodos de processamento de imagem, incluindo um algoritmo de distorção de imagem chamado four-corners-in-one para fins de treino. Depois de detalharmos alguns esquemas de rede, apresentamos e discutimos alguns dos resultados experimentais obtidos pela abordagem proposta, que julgamos promissores e encorajadores para ajudar os viticultores do Douro na classificação automática das variedades de videira para futura implementação de um robot para colheita de uvas.