Latent Dirichlet Allocation (LDA) is widely used for automatic discovery of latent topics in document corpora. However, output from analysis using an LDA topic model suffers from a lack of identifiability between topics not only across corpora, but across runs of the algorithm. The output is also isolated from enriching information from knowledge sources such as Wikipedia and is difficult for humans to interpret due to a lack of meaningful topic labels. This thesis introduces two methods for probabilistic explicit topic modeling that address these issues: Latent Dirichlet Allocation with Static Topic-Word Distributions (LDA-STWD), and Explicit Dirichlet Allocation (EDA). LDA-STWD directly substitutes precomputed counts for LDA topic-word counts, leveraging existing Gibbs sampler inference. EDA defines an entirely new explicit topic model and derives the inference method from first principles. Both of these methods approximate topic-word distributions a priori using word distributions from Wikipedia articles, with each article corresponding to one topic and the article title being used as a topic label. By this means, LDA-STWD and EDA overcome the nonidentifiability, isolation, and unintepretability of LDA output. We assess the effectiveness of LDA-STWD and EDA by means of three tasks: document classification, topic label generation, and document label generation. Label quality is quantified by means of user studies. We show that a competing non-probabilistic explicit topic model handily beats both LDA-STWD and EDA as a dimensionality reduction technique in a document classification task. Surprisingly, we find that topic labels from another approach using LDA and post hoc topic labeling (called LDA+Lau) are on one corpus preferred over topic labels prespecified from Wikipedia. Finally, we show that LDA-STWD improves substantially upon the performance of the state of the art in document labeling.

Many music composition algorithms attempt to compose music in a particular style. The resulting music is often impressive and indistinguishable from the style of the training data, but it tends to lack significant innovation. In an effort to increase innovation in the selection of pitches and rhythms, we present a system that discovers musical motifs by coupling machine learning techniques with an inspirational component. The inspirational component allows for the discovery of musical motifs that are unlikely to be produced by a generative model, while the machine learning component harnesses innovation. Candidate motifs are extracted from non-musical media such as images and audio. Machine learning algorithms select the motifs that best comply with patterns learned from training data. This process is validated by extracting motifs from real music scores, identifying themes in the piece according to a theme database, and measuring the probability of discovering thematic motifs verses non-thematic motifs. We examine the information content of the discovered motifs by comparing the entropy of the discovered motifs, candidate motifs, and training data. We measure innovation by comparing the probability of the training data and the probability of the discovered motifs given the model. We also compare the probabilities of media-inspired motifs with random motifs and find that media inspiration is more efficient than random generation.

The main research objective of this dissertation is to understand the sensing and communication challenges to achieving cooperative perception among autonomous vehicles, and then, using the insights gained, guide the design of the suitable format of data to be exchanged, reliable and efficient data fusion algorithms on vehicles. By understanding what and how data are exchanged among autonomous vehicles, from a machine learning perspective, it is possible to realize precise cooperative perception on autonomous vehicles, enabling massive amounts of sensor information to be shared amongst vehicles. I first discuss the trustworthy perception information sharing on connected and autonomous vehicles. Then how to achieve effective cooperative perception on autonomous vehicles via exchanging feature maps among vehicles is discussed in the following. In the last methodology part, I propose a set of mechanisms to improve the solution proposed before, i.e., reducing the amount of data transmitted in the network to achieve an efficient cooperative perception. The effectiveness and efficiency of our mechanism is analyzed and discussed.

Functional data analysis (FDA) refers to a broad collection of statistical and machine learning methods that deal with the data in the form of random functions. In general, functional data are assumed to lie in a constrained functional space, e.g., images, and smooth curves, rather than the conventional Euclidean space, e.g., scalar vectors. The explosion of massive data and high-performance computational resources brings exciting opportunities as well as new challenges to this field. On one hand, the rich information from modern functional data enables an investigation into the underlying data patterns at an unprecedented scale and resolution. On the other hand, the inherent complex structures and huge data sizes of modern functional data pose additional practical challenges to model building, model training, and model interpretation under various circumstances. This dissertation discusses recent advances in machine learning for analyzing complex structured functional data. Chapter 1 begins with a general introduction to examples of modern functional data and related data analysis challenges. Chapter 2 introduces a novel machine learning framework, artificial perceptual learning (APL), to tackle the problem of weakly supervised learning in functional remote sensing data. Chapter 3 develops a flexible function-on-scalar regression framework, Wasserstein distributional learning (WDL), to address the challenge of modeling density functional outputs. Chapter 4 concludes the dissertation and discusses future directions.

Undoubtedly, the advancement of IoT technology has created a plethora of new applications and a growing number of devices connected to the internet. Among these developments emerged the novel concept of smart farming. In this context, sensor nodes are used in farms to help farmers acquire a deeper insight into the environmental factors affecting their productivity. In recent years, we have witnessed an emerging trend of scholarly literature focused on smart farming. Some focus has been on system architecture for monitoring purposes, while another area of interest includes yield prediction. Humidity, air and soil temperature, solar radiation, and wind speed are some key weather elements monitored in smart farms. We introduce a mechanistic crop growth model to predict crop growth and subsequent yield, subject to weather, soil parameters, crop characteristics and management practices. We also seek to measure the influence of nitrogen on yield throughout the growing season. The machine learning models are trained to emulate the crop growth model in the state of Iowa (US). The multilayer perceptron (MLP) is chosen to evaluate the model prediction as it generates fewer errors. Furthermore, the MLP optimization model is used to maximize corn yield. The experiment was performed using different scenarios, stochastic gradient descent (SGD), and adaptive moment estimation (Adam) optimizers. The experiment results revealed that the SGD optimizer and the dataset with the scenario of unchanged parameters provided the highest crop yield compared to the mechanistic crop growth model.

No description available.

Curated music collection is a growing field as a result of the freedom and supply that streaming music services like Spotify provide us with. To be able to categorize music tracks based on subjective core values in a scalable manner, this thesis has explored if recommending such labels are possible through machine learning. When analysing 2464 tracks with one or more of the 22 different core values a profile was built up for each track by features from three different categories: editorial, cultural and acoustic. When classifying the tracks into core values different methods of multi-label classification were explored. By combining five different transformation approaches with three base classifiers and using two algorithm adaptations a total of 17 different configurations were constructed. The different configu- rations were evaluated with multiple measurements including (but not limited to) Hamming Loss, Ranking Loss, One error, F1 score, exact match and both training and testing time. The results showed that the problem transformation algorithm Label Powerset together with Sequential minimal optimization outper- formed the other configurations. We also found promising results for neural networks, something that should be investigated further in the future. / Kurerade musiksamlingar är ett växande område som en direkt följd av den frihet som strömmande musiktjänster som Spotify ger oss. För att kunna kategorisera låtar baserade på subjektiva värderingar på ett skalbart sätt har denna avhandling undersökt om rekommendationer av sådana etiketter är möjliga genom maskininlärning. När 2464 spår med ett eller flera av 22 olika kärnvärden analyserades byggdes en profil för varje spår upp av attribut från tre olika kategorier: redaktionella, kulturella och akustiska. Vid klassificering av spåren undersöktes flera olika metoder för fleretikettsklassificering. Genom att kombinera fem olika transformationsmetoder med tre bas-klassificerare och använda två algoritm-anpassningar konstruerades totalt 17 olika konfigurationer. De olika konfigurationerna utvärderades med flera olika mätvärden, inkluderat (men inte begränsat till) Hamming Loss, Ranking Loss, One error, F1 score, exakt matchning och både träningstid och testningstid. Resultaten visade att transformationsalgoritmen ”Label Powerset” tillsammans med Sekventiell Minimal Optimering utklassade de andra konfigurationerna. Vi fann också lovande resultat för artificiella neuronnät, något som bör undersökas ytterligare i framtiden.

Machine learning has gained a lot of interest over the past years and is now used extensively in various areas. Google has developed a framework called TensorFlow which simplifies the usage of machine learning without compromising the end result. However, it does not resolve the issue of neural network training being time consuming. The purpose of this thesis is to investigate with what accuracy training times can be predicted using TensorFlow. Essentially, how effectively one neural network in TensorFlow can be used to predict the training times of other neural networks, also in TensorFlow. In order to do this, training times for training different neural networks was collected. This data was used to create a neural network for prediction. The resulting neural network is capable of predicting training times with an average accuracy of 93.017%. / Maskininlärning har fått mycket uppmärksamhet de senaste åren och används nu i stor utsträckning inom olika områden. Google har utvecklat ramverket TensorFlow som förenklar användningen av maskininlärning utan att kompromissa slutresultatet. Det löser dock inte problemet med att det är tidskrävande att träna neurala nätverk. Syftet med detta examensarbete är att undersöka med vilken noggrannhet träningstiden kan förutsägas med TensorFlow. Alltså, hur effektivt kan ett neuralt nätverk i TensorFlow användas för att förutsäga träningstiderna av andra neurala nätverk, även dessa i TensorFlow. För att göra detta samlades träningstider för olika neurala nätverk. Datan användes sedan för att skapa ett neuralt nätverk för förutsägelse. Det resulterande neurala nätverket kan förutsäga träningstider med en genomsnittlig noggrannhet på 93,017%.

Modern machine learning techniques focus on extremely deep and multi-pathed networks, resulting in large memory and computational requirements. This thesis explores techniques for designing efficient convolutional networks including pixel shuffling, depthwise convolutions, and various activation fucntions. These techniques are then applied to two image processing domains: single-image super-resolution and image compression. The super-resolution model, TinyPSSR, is one-third the size of the next smallest model in literature while performing similar to or better than other larger models on representative test sets. The efficient deep image compression model is significantly smaller than any other model in literature and performs similarly in both computational cost and reconstruction quality to the JPEG standard.

The overarching goal of this work was to develop novel data collection and analysismethods to better understand how soldier burden affects the soldier survivability tradespace (i.e.,performance, musculoskeletal health, and susceptibility to enemy action). To achieve this goal,three studies were completed: 1) a mobile inertial measurement unit (IMU) suit was validatedagainst an optical motion capture (OPT) system; 2) data from the IMU suit was used to develop aframework for morphing movement patterns to represent intermediary body-borne load massesand personal characteristics; and 3) a single IMU was used to develop a human activity recognitionalgorithm and calculate tradespace metrics.In study one, a whole-body IMU suit (MVN Link, Xsens, Netherlands) was validatedagainst an OPT system (Vantage V5, Vicon, United Kingdom) for military-based movementsusing the root mean squared error (RMSE) of joint angles and Pearson correlation coefficients ofprincipal component (PC) scores. During a standard implementation (i.e., using differentbiomechanical models and not attempting to align them; VOPT vs. XIMU), average RMSE valuesacross all tasks were less than 9° for the lower limbs but up to 40.5° for the upper limbs. Whenusing the same biomechanical model and applying an alignment procedure (VOPT vs. VIMU-CAL),RMSE values decreased to an average of 2.5º and 17.5º for the lower and upper limbs, respectively.Of the 48 retained PCs, 38 (79%) had scores with a high or very high positive correlation (> +0.70)between the OPT and IMU systems, 15 (31%) of which had scores with a very high correlation (>+0.90). The average Pearson correlation coefficient was 0.81 (SD = 0.14). Given these results, theIMU system was deemed appropriate for collecting military-based movement patterns.In study two, principal component analysis (PCA) and linear discriminant analysis (LDA)were used to generate whole-body morphable movement patterns to represent intermediary body-ixborne loads and personal characteristics (sex, body mass, military experience). Reconstructedmovements were used for animation, musculoskeletal modelling, exposure time calculations, andsusceptibility calculations; all calculated values were comparable to previous research. Thisproject displayed that a relatively small representative dataset can be used to simulate the changein whole-body movement patterns caused by many different body-borne loads and personalcharacteristics not originally collected. By implementing this framework, defence scientists canreduce the amount and complexity of data collections needed to better understand the impact onthe survivability tradespace caused by all types of soldier burden.Study three focused on developing a deployable method for calculating tradespace metricsin the field. Three deep neural network (DNN) architectures were trained to identify eleven classlabels using data from a single IMU on the upper back. Data were collected during an indoorlaboratory-based protocol and an outdoor simulated two-person section attack. The predictionsmade by the DNNs were processed through a two-step logical algorithm to apply real-worldconstraints and expand the predictions to 19 class labels. The deep convolutional long short-termneural network architecture outperformed the convolutional neural network and fully-connectedneural network for all three approaches: indoor only, section attack only, and general. Movementswere identified with a high degree of accuracy (> 87% for accuracy and weighted F1-score), andtradespace metrics were calculated within 0.17 seconds, 0.21 shots, and 1.25% susceptibilitycompared to the tradespace metrics calculated from the ground truth labels.Overall, the data-driven methods developed throughout this dissertation can be used bydefence scientists and military leaders to improve the understanding of the survivabilitytradespace, which has the potential to improve the quality of life of soldiers, making them more fitand ready to fight, thus increasing the likelihood of mission success.