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

Neuroscience as a field is collecting more data than at any other time in history. The scale of this data allows us to ask fundamental questions about the mechanisms of brain function, the basis of behavior, and the development of disorders. Our ambitious goals as well as the abundance of data being recorded call for reproducible, reliable, and accessible systems to push the field forward. While we have made great strides in building reproducible and accessible machine learning (ML) systems for neuroscience, reliability remains a major issue. In this dissertation, we show that we can leverage existing data and domain expert knowledge to build more reliable ML systems to study animal behavior. First, we consider animal pose estimation, a crucial component in many scientific investigations. Typical transfer learning ML methods for behavioral tracking treat each video frame and object to be tracked independently. We improve on this by leveraging the rich spatial and temporal structures pervasive in behavioral videos. Our resulting weakly supervised models achieve significantly more robust tracking. Our tools allow us to achieve improved results when we have imperfect, limited data while requiring users to label fewer training frames and speeding up training. We can more accurately process raw video data and learn interpretable units of behavior. In turn, these improvements enhance performance on downstream applications. Next, we consider a ubiquitous approach to (attempt to) improve the reliability of ML methods, namely combining the predictions of multiple models, also known as deep ensembling. Ensembles of classical ML predictors, such as random forests, improve metrics such as accuracy by well-understood mechanisms such as improving diversity. However, in the case of deep ensembles, there is an open methodological question as to whether, given the choice between a deep ensemble and a single neural network with similar accuracy, one model is truly preferable over the other. Via careful experiments across a range of benchmark datasets and deep learning models, we demonstrate limitations to the purported benefits of deep ensembles. Our results challenge common assumptions regarding the effectiveness of deep ensembles and the “diversity” principles underpinning their success, especially with regards to important metrics for reliability, such as out-of-distribution (OOD) performance and effective robustness. We conduct additional studies of the effects of using deep ensembles when certain groups in the dataset are underrepresented (so-called “long tail” data), a setting whose importance in neuroscience applications is revealed by our aforementioned work. Altogether, our results demonstrate the essential importance of both holistic systems work and fundamental methodological work to understand the best ways to apply the benefits of modern machine learning to the unique challenges of neuroscience data analysis pipelines. To conclude the dissertation, we outline challenges and opportunities in building next-generation ML systems.

Deep Learning models have seen great application in demanding tasks such as machine translation and autonomous driving. However, building such models has proved challenging, both from a computational perspective and due to the requirement of a plethora of annotated data. Moreover, when challenged on new situations or data distributions (target domain), those models may perform inadequately. Such examples are transitioning from one city to another, different weather situations, or changes in sunlight. Unsupervised Domain adaptation (UDA) exploits unlabelled data (easy access) to adapt models to new conditions or data distributions. Inspired by the fact that environmental changes happen gradually, we focus on Online UDA. Instead of directly adjusting a model to a demanding condition, we constantly perform minor adaptions to every slight change in the data, creating a soft transition from the current domain to the target one. To perform gradual adaptation, we utilized state-of-the-art semantic segmentation approaches on increasing rain intensities (25, 50, 75, 100, and 200mm of rain). We demonstrate that deep learning models can adapt substantially better to hard domains when exploiting intermediate ones. Moreover, we introduce a model switching mechanism that allows adjusting back to the source domain, after adaptation, without dropping performance. / Deep Learning-modeller har sett stor tillämpning i krävande uppgifter som maskinöversättning och autonom körning. Att bygga sådana modeller har dock visat sig vara utmanande, både ur ett beräkningsperspektiv och på grund av kravet på en uppsjö av kommenterade data. Dessutom, när de utmanas i nya situationer eller datadistributioner (måldomän), kan dessa modeller prestera otillräckligt. Sådana exempel är övergång från en stad till en annan, olika vädersituationer eller förändringar i solljus. Unsupervised Domain adaptation (UDA) utnyttjar omärkt data (enkel åtkomst) för att anpassa modeller till nya förhållanden eller datadistributioner. Inspirerade av att miljöförändringar sker gradvis, fokuserar vi på Online UDA. Istället för att direkt anpassa en modell till ett krävande tillstånd, gör vi ständigt mindre anpassningar till varje liten förändring i data, vilket skapar en mjuk övergång från den aktuella domänen till måldomänen. För att utföra gradvis anpassning använde vi toppmoderna semantiska segmenteringsmetoder för att öka regnintensiteten (25, 50, 75, 100 och 200 mm regn). Vi visar att modeller för djupinlärning kan anpassa sig betydligt bättre till hårda domäner när man utnyttjar mellanliggande. Dessutom introducerar vi en modellväxlingsmekanism som tillåter justering tillbaka till källdomänen, efter anpassning, utan att tappa prestanda.

Disertace se zabývá aplikaci neuronových sítí pro detekci objektů na multimodální data v robotice. Celkem cílí na tři oblasti: tvorbu datasetu, zpracování multimodálních dat a trénování neuronových sítí. Nejdůležitější části práce je návrh metody pro tvorbu rozsáhlých anotovaných datasetů bez časové náročného lidského zásahu. Metoda používá neuronové sítě trénované na RGB obrázcích. Užitím dat z několika snímačů pro vytvoření modelu okolí a mapuje anotace z RGB obrázků na jinou datovou doménu jako jsou termální obrázky, či mračna bodů. Pomoci této metody autor vytvořil dataset několika set tisíc anotovaných obrázků a použil je pro trénink neuronové sítě, která následně překonala modely trénované na menších, lidmi anotovaných datasetech. Dále se autor v práci zabývá robustností detekce objektů v několika datových doménách za různých povětrnostních podmínek. Práce také popisuje kompletní řetězec zpracování multimodálních dat, které autor vytvořil během svého doktorského studia. To Zahrnuje vývoj unikátního senzorického zařízení, které je vybavené řadou snímačů běžně užívaných v robotice. Dále autor popisuje proces tvorby rozsáhlého, veřejně dostupného datasetu Brno Urban Dataset. Na závěr autor popisuje software, který vznikl během jeho studia a jak je tento software užit při zpracování dat v rámci jeho práce (Atlas Fusion a Robotic Template Library).

Vision is an important tool for us humans and significant effort has been put into creating solutions that let us measure how we use it. Most common among the techniques to measure gaze direction is to use specialised hardware such as infrared eye trackers. Recently, several Convolutional Neural Network (CNN) based architectures have been suggested yielding impressive results on single Red Green Blue (RGB) images. However, limited research has been done around whether using several sequential images can lead to improved tracking performance. Expanding this research to include low frequency and low quality RGB images can further open up the possibility to improve tracking performance for models using off-the-shelf hardware such as web cameras or smart phone cameras. GazeCapture is a well known dataset used for training RGB based CNN models but it lacks sequences of images and natural eye movements. In this thesis, a geometric gaze estimation model is introduced and synthetic data is generated using Unity to create sequences of images with both RGB input data as well as ground Point of Gaze (POG). To make these images more natural appearing domain adaptation is done using a CycleGAN. The data is then used to train several different models to evaluate whether temporal information can increase accuracy. Even though the improvement when using a Gated Recurrent Unit (GRU) based temporal model is limited over simple sequence averaging, the network achieves smoother tracking than a single image model while still offering faster updates over a saccade (eye movement) compared to averaging. This indicates that temporal features could improve accuracy. There are several promising future areas of related research that could further improve performance such as using real sequential data or further improving the domain adaptation of synthetic data. / Synen är ett viktigt sinne för oss människor och avsevärd energi har lagts ner på att skapa lösningar som låter oss mäta hur vi använder den. Det vanligaste sättet att göra detta idag är att använda specialiserad hårdvara baserad på infrarött ljus för ögonspårning. På senare tid har maskininlärning och modeller baserade på CNN uppnått imponerande resultat för enskilda RGB-bilder men endast begränsad forskning har gjorts kring huruvida användandet av en sekvens av högupplösta bilder kan öka prestandan för dessa modeller ytterligare. Genom att uttöka denna till bildserier med lägre frekvens och kvalitet kan det finnas möjligheter att förbättra prestandan för sekventiella modeller som kan använda data från standard-hårdvara såsom en webbkamera eller kameran i en vanlig telefon. GazeCapture är ett välkänt dataset som kan användas för att träna RGB-baserade CNN-modeller för enskilda bilder. Dock innehåller det inte bildsekvenser eller bilder som fångar naturliga ögonrörelser. För att hantera detta tränades de sekventiella modellerna i denna uppsats med data som skapats från 3D-modeller i Unity. För att den syntetiska datan skulle vara jämförbar med riktiga bilder anpassades den med hjälp av ett CycleGAN. Även om förbättringen som uppnåddes med sekventiella GRU-baserade modeller var begränsad jämfört med en modell som använde medelvärdet för sekvensen så uppnådde den tränade sekventiella modellen jämnare spårning jämfört med enbildsmodeller samtidigt som den uppdateras snabbare vid en sackad (ögonrörelse) än medelvärdesmodellen. Detta indikerar att den tidsmässiga information kan förbättra ögonspårning även för lågfrekventa bildserier med lägre kvalitet. Det finns ett antal intressanta områden att fortsätta undersöka för att ytterligare öka prestandan i liknande system som till exempel användandet av större mängder riktig sekventiell data eller en förbättrad domänanpassning av syntetisk data.

Automatic analysis of facial actions (AFA) can reveal a person’s emotion, intention, and physical state, and make possible a wide range of applications. To enable reliable, valid, and efficient AFA, this thesis investigates automatic analysis of facial actions through transductive, supervised and unsupervised learning. Supervised learning for AFA is challenging, in part, because of individual differences among persons in face shape and appearance and variation in video acquisition and context. To improve generalizability across persons, we propose a transductive framework, Selective Transfer Machine (STM), which personalizes generic classifiers through joint sample reweighting and classifier learning. By personalizing classifiers, STM offers improved generalization to unknown persons. As an extension, we develop a variant of STM for use when partially labeled data are available. Additional challenges for supervised learning include learning an optimal representation for classification, variation in base rates of action units (AUs), correlation between AUs and temporal consistency. While these challenges could be partly accommodated with an SVM or STM, a more powerful alternative is afforded by an end-to-end supervised framework (i.e., deep learning). We propose a convolutional network with long short-term memory (LSTM) and multi-label sampling strategies. We compared SVM, STM and deep learning approaches with respect to AU occurrence and intensity in and between BP4D+ [282] and GFT [93] databases, which consist of around 0.6 million annotated frames. Annotated video is not always possible or desirable. We introduce an unsupervised Branch-and-Bound framework to discover correlated facial actions in un-annotated video. We term this approach Common Event Discovery (CED). We evaluate CED in video and motion capture data. CED achieved moderate convergence with supervised approaches and enabled discovery of novel patterns occult to supervised approaches.

<div> <div> <div> <p>Medication non-adherence and non-compliance left unaddressed can compound into severe medical problems for patients. Identifying patients that are likely to become non-adherent can help reduce these problems. Despite these benefits, monitoring adherence at scale is cost-prohibitive. Social forums offer an easily accessible, affordable, and timely alternative to the traditional methods based on claims data. This study investigates the potential of medication adherence prediction based on social forum data for diabetes and fibromyalgia therapies by using transfer learning from the Medical Expenditure Panel Survey (MEPS). </p><p><br></p> <p>Predictive adherence models are developed by using both survey and social forums data and different random forest (RF) techniques. The first of these implementations uses binned inputs from k-means clustering. The second technique is based on ternary trees instead of the widely used binary decision trees. These techniques are able to handle missing data, a prevalent characteristic of social forums data. </p><p><br></p> <p>The results of this study show that transfer learning between survey models and social forum models is possible. Using MEPS survey data and the techniques listed above to derive RF models, less than 5% difference in accuracy was observed between the MEPS test dataset and the social forum test dataset. Along with these RF techniques, another RF implementation with imputed means for the missing values was developed and shown to predict adherence for social forum patients with an accuracy >70%. </p> </div> </div> <div> <div> <p><br></p> </div> </div> </div> <div> <div> <div> <p>This thesis shows that a model trained with verified survey data can be used to complement traditional medical adherence models by predicting adherence from unverified, self-reported data in a dynamic and timely manner. Furthermore, this model provides a method for discovering objective insights from subjective social reports. Additional investigation is needed to improve the prediction accuracy of the proposed model and to assess biases that may be inherent to self-reported adherence measures in social health networks. </p> </div> </div> </div>

The organization International Aid Services (IAS) provides people in East Africawith clean water through well drilling. The wells are located in surroundingsfar away for the investors to inspect and therefore IAS wishes to be able to monitortheir wells to get a better overview if different types of improvements needto be made. To see the load on different water sources at different times of theday and during the year, and to know how many people that are visiting thewells, is of particular interest. In this paper, a method is proposed for countingpeople around the wells. The goal is to choose a suitable method for detectinghumans in images and evaluate how it performs. The area of counting humansin images is not a new topic, though it needs to be taken into account that thesituation implies some restrictions. A Raspberry Pi with an associated camerais used, which is a small embedded system that cannot handle large and complexsoftware. There is also a limited amount of data in the project. The methodproposed in this project uses a pre-trained convolutional neural network basedobject detector called the Single Shot Detector, which is adapted to suit smallerdevices and applications. The pre-trained network that it is based on is calledMobileNet, a network that is developed to be used on smaller systems. To see howgood the chosen detector performs it will be compared with some other models.Among them a detector based on the Inception network, a significantly larger networkthan the MobileNet. The base network is modified by transfer learning.Results shows that a fine-tuned and modified network can achieve better result,from a F1-score of 0.49 for a non-fine-tuned model to 0.66 for the fine-tuned one.

We created and analyzed a text classification dataset from freely-available web documents from the United Nation's Sustainable Development Goals. We then used it to train and compare different multi-label text classifiers with the aim of exploring the alternatives for methods that facilitate the search of information of this type of documents. We explored the effectiveness of deep learning and transfer learning in text classification by fine-tuning different pre-trained language representations — Word2Vec, GloVe, ELMo, ULMFiT and BERT. We also compared these approaches against a baseline of more traditional algorithms without using transfer learning. More specifically, we used multinomial Naive Bayes, logistic regression, k-nearest neighbors and Support Vector Machines. We then analyzed the results of our experiments quantitatively and qualitatively. The best results in terms of micro-averaged F1 scores and AUROC are obtained by BERT. However, it is also interesting that the second best classifier in terms of micro-averaged F1 scores is the Support Vector Machines, closely followed by the logistic regression classifier, which both have the advantage of being less computationally expensive than BERT. The results also show a close relation between our dataset size and the effectiveness of the classifiers.

The purpose of this dissertation is to understand how algorithms can efficiently learn to solve new tasks based on previous experience, instead of being explicitly programmed with a solution for each task that we want it to solve. Here a task is a series of decisions, such as a robot vacuum deciding which room to clean next or an intelligent car deciding to stop at a traffic light. In such a case, state-of-the-art learning algorithms are difficult to employ in practice because they often make thou- sands of mistakes before reliably solving a task. However, humans learn solutions to novel tasks, often making fewer mistakes, which suggests that efficient learning algorithms may exist. One advantage that humans have over state- of-the-art learning algorithms is that, while learning a new task, humans can apply knowledge gained from previously solved tasks. The central hypothesis investigated by this dissertation is that learning algorithms can solve new tasks more efficiently when they take into consideration knowledge learned from solving previous tasks. Al- though this hypothesis may appear to be obviously true, what knowledge to use and how to apply that knowledge to new tasks is a challenging, open research problem. I investigate this hypothesis in three ways. First, I developed a new learning algorithm that is able to use prior knowledge to constrain the exploration space. Second, I extended a powerful theoretical framework in machine learning, called Probably Approximately Correct, so that I can formally compare the efficiency of algorithms that solve only a single task to algorithms that consider knowledge from previously solved tasks. With this framework, I found sufficient conditions for using knowledge from previous tasks to improve efficiency of learning to solve new tasks and also identified conditions where transferring knowledge may impede learning. I present situations where transfer learning can be used to intelligently constrain the exploration space so that optimality loss can be minimized. Finally, I tested the efficiency of my algorithms in various experimental domains. These theoretical and empirical results provide support for my central hypothesis. The theory and experiments of this dissertation provide a deeper understanding of what makes a learning algorithm efficient so that it can be widely used in practice. Finally, these results also contribute the general goal of creating autonomous machines that can be reliably employed to solve complex tasks.