This thesis presents an evaluation of the effectiveness of an agroforestry extension program to smallholder farmers on Leyte Island, the Philippines. The imperative for reforestation is well recognised in the Philippines and was the impetus for this program which provided farmers with assistance to establish and silviculturally manage timber trees on their land. Because the cost-effectiveness of agroforestry extension is increased if farmers develop self-efficacy without extensive training, the extension program was offered in two regimes to test the necessity for extended assistance. In the extended assistance regime, farmers were offered on-site assistance to collect seed, grow seedlings, prepare sites and establish trees, whereas in the limited assistance regime, farmers were only offered assistance to collect seed and grow seedlings. Descriptive statistics were collected of farmers’ acceptance of technology and the manner in which technology was adapted to suit their personal circumstances. Translated conversations between farmers and extension staff also provided a rich source of data which provided insights into farmers’ motivation. Extension activities were reviewed at a mid-program workshop, a final on-site inspection and an end-of-program workshop. Farmers responded positively to the extended assistance program which helped them to grow and out-plant seedlings. The limited assistance program was relatively unsuccessful. Overall, the extension program was successful in shifting the initiative for further planting from extension staff to participating farmers. However, farmers showed little interest in applying silvicultural thinning or pruning to existing plantations of trees because extension advice was not congruent with their existing mental models of these procedures. Systems modelling of socio-economic variables which had been found to affect program outcomes was used to predict critical success factors. A key constraint to program recruitment was found to be farmers’ perception of harvest security, even when their needs for technology and planting materials are met. Modelling also cast doubt on the usefulness of written extension materials and emphasised the necessity for extended face-to-face technical assistance. Although conducted in Leyte, the findings of this research provide guidance for issues which affect the adoption of agroforestry both in the Philippines and in other countries. The research found that it was possible to recruit and motivate farmers without providing material incentives. If farmers experienced unexpected problems, providing extended face-to-face contact and assistance was critical if catastrophic losses of participating farmers were to be avoided. The failure of attempts to introduce advanced-age silviculture also indicated a need to elicit farmers’ mental models as a precursor or parallel enquiry to extension activities. In a situation where little was initially known about farmers’ understanding of agroforestry technology or the variables which affect their acceptance or rejection of extension assistance, the results of this research have shown that it is possible to build the capacity of farmers to establish timber trees. This result is in contrast to the acknowledged failure of the logging concession system in the Philippines and the difficulties faced by some industrial plantations and community-based programs. This investigation has shown that an opportunity exists to lift the level of tree planting in Leyte, provided that system variables which are either critical success factors or impediments are addressed.

Restricted Boltzmann Machines (RBMs) and autoencoders have been used - in several variants - for similar tasks, such as reducing dimensionality or extracting features from signals. Even though their structures are quite similar, they rely on different training theories. Lately, they have been largely used as building blocks in deep learning architectures that are called deep belief networks (instead of stacked RBMs) and stacked autoencoders. In light of this, the student has worked on this thesis with the aim to understand the extent of the similarities and the overall pros and cons of using either RBMs, autoencoders or denoising autoencoders in deep networks. Important characteristics are tested, such as the robustness to noise, the influence on training of the availability of data and the tendency to overtrain. The author has then dedicated part of the thesis to study how the three deep networks in exam form their deep internal representations and how similar these can be to each other. In result of this, a novel approach for the evaluation of internal representations is presented with the name of F-Mapping. Results are reported and discussed.

Recent advancements in processing power, storage capabilities, and availability of data, has led to improvements in many applications through the use of machine learning. Using machine learning in control systems was first suggested in the 1990s, but is more recently being implemented. In this thesis, an embedded climate system, which is a type of control system, will be looked at. The ways in which machine learning can be used to replicate portions of the climate system is looked at. Deep Belief Networks are the machine learning models of choice. Firstly, the functionality of a PID controller is replicated using a Deep Belief Network. Then, the functionality of a more complex control path is replicated. The performance of the Deep Belief Networks are evaluated at how they compare to the original control portions, and the performance in hardware. It is found that the Deep Belief Network can quite accurately replicate the behaviour of a PID controller, whilst the performance is worse for the more complex control path. It was seen that the use of delays in input features gave better results than without. A climate system with a Deep Belief Network was also loaded onto hardware. The minimum requirements of memory usage and CPU usage were met. However, the CPU usage was greatly affected, and if this was to be used in practice, work should be done to decrease it. / Många applikationer har förbättras genom användningen av maskininlärning. Maskininlärning för reglersystem föreslogs redan på 1990-talet och har nu börjat tillämpas, eftersom processorkraft, lagringsmöjligheter och tillgänglighet till rådata ökat. I detta examensarbete användes ett inbäddat klimatsystem, som är en typ av reglersystem. Maskininlärningsmodellen Deep Belief Network användes för att undersöka hur delar av klimatsystemet skulle kunna återskapas. Först återskapades funktionaliteten hos en PID-regulator och sedan funktionaliteten av en mer komplex del av reglersystemet Prestandan hos nätverken utvärderades i jämförelse med prestandan i de ursprungliga kontrolldelarna och hårdvaran. Det visade sig att Deep Belief Network utmärkt kunde replikera PID-regulatorns beteende, medan prestandan var lägre för den komplexa delen av reglersystemet. Användningen av fördröjningar i indata till nätverken gav bättre resultat än utan. Ett klimatsystem med ett Deep Belief Network laddades också över på hårdvaran. Minimikrav för minnesanvändning och CPU- användning var uppfyllda, men CPU- användningen påverkades kraftigt. Detta gör, att om maskininlärning ska kunna användas i verkligheten, bör CPU-användningen minskas.

Multimodal biometric systems have been widely applied in many real-world applications due to its ability to deal with a number of significant limitations of unimodal biometric systems, including sensitivity to noise, population coverage, intra-class variability, non-universality, and vulnerability to spoofing. This PhD thesis is focused on the combination of both the face and the left and right irises, in a unified hybrid multimodal biometric identification system using different fusion approaches at the score and rank level. Firstly, the facial features are extracted using a novel multimodal local feature extraction approach, termed as the Curvelet-Fractal approach, which based on merging the advantages of the Curvelet transform with Fractal dimension. Secondly, a novel framework based on merging the advantages of the local handcrafted feature descriptors with the deep learning approaches is proposed, Multimodal Deep Face Recognition (MDFR) framework, to address the face recognition problem in unconstrained conditions. Thirdly, an efficient deep learning system is employed, termed as IrisConvNet, whose architecture is based on a combination of Convolutional Neural Network (CNN) and Softmax classifier to extract discriminative features from an iris image. Finally, The performance of the unimodal and multimodal systems has been evaluated by conducting a number of extensive experiments on large-scale unimodal databases: FERET, CAS-PEAL-R1, LFW, CASIA-Iris-V1, CASIA-Iris-V3 Interval, MMU1 and IITD and MMU1, and SDUMLA-HMT multimodal dataset. The results obtained have demonstrated the superiority of the proposed systems compared to the previous works by achieving new state-of-the-art recognition rates on all the employed datasets with less time required to recognize the person’s identity.Multimodal biometric systems have been widely applied in many real-world applications due to its ability to deal with a number of significant limitations of unimodal biometric systems, including sensitivity to noise, population coverage, intra-class variability, non-universality, and vulnerability to spoofing. This PhD thesis is focused on the combination of both the face and the left and right irises, in a unified hybrid multimodal biometric identification system using different fusion approaches at the score and rank level. Firstly, the facial features are extracted using a novel multimodal local feature extraction approach, termed as the Curvelet-Fractal approach, which based on merging the advantages of the Curvelet transform with Fractal dimension. Secondly, a novel framework based on merging the advantages of the local handcrafted feature descriptors with the deep learning approaches is proposed, Multimodal Deep Face Recognition (MDFR) framework, to address the face recognition problem in unconstrained conditions. Thirdly, an efficient deep learning system is employed, termed as IrisConvNet, whose architecture is based on a combination of Convolutional Neural Network (CNN) and Softmax classifier to extract discriminative features from an iris image. Finally, The performance of the unimodal and multimodal systems has been evaluated by conducting a number of extensive experiments on large-scale unimodal databases: FERET, CAS-PEAL-R1, LFW, CASIA-Iris-V1, CASIA-Iris-V3 Interval, MMU1 and IITD and MMU1, and SDUMLA-HMT multimodal dataset. The results obtained have demonstrated the superiority of the proposed systems compared to the previous works by achieving new state-of-the-art recognition rates on all the employed datasets with less time required to recognize the person’s identity. / Higher Committee for Education Development in Iraq

Cette thèse de doctorat est dédiée à l’analyse automatique de visages 3D, incluant la détection de points d’intérêt et la reconnaissance de l’expression faciale. En effet, l’expression faciale joue un rôle important dans la communication verbale et non verbale, ainsi que pour exprimer des émotions. Ainsi, la reconnaissance automatique de l’expression faciale offre de nombreuses opportunités et applications, et est en particulier au coeur d’interfaces homme-machine "intelligentes" centrées sur l’être humain. Par ailleurs, la détection automatique de points d’intérêt du visage (coins de la bouche et des yeux, ...) permet la localisation d’éléments du visage qui est essentielle pour de nombreuses méthodes d’analyse faciale telle que la segmentation du visage et l’extraction de descripteurs utilisée par exemple pour la reconnaissance de l’expression. L’objectif de cette thèse est donc d’élaborer des approches de détection de points d’intérêt sur les visages 3D et de reconnaissance de l’expression faciale pour finalement proposer une solution entièrement automatique de reconnaissance de l’activité faciale incluant l’expression et les unités d’action (ou Action Units). Dans ce travail, nous avons proposé un réseau de croyance bayésien (Bayesian Belief Network ou BBN) pour la reconnaissance d’expressions faciales ainsi que d’unités d’action. Un modèle statistique de caractéristiques faciales (Statistical Facial feAture Model ou SFAM) a également été élaboré pour permettre la localisation des points d’intérêt sur laquelle s’appuie notre BBN afin de permettre la mise en place d’un système entièrement automatique de reconnaissance de l’expression faciale. Nos principales contributions sont les suivantes. Tout d’abord, nous avons proposé un modèle de visage partiel déformable, nommé SFAM, basé sur le principe de l’analyse en composantes principales. Ce modèle permet d’apprendre à la fois les variations globales de la position relative des points d’intérêt du visage (configuration du visage) et les variations locales en terme de texture et de forme autour de chaque point d’intérêt. Différentes instances de visages partiels peuvent ainsi être produites en faisant varier les valeurs des paramètres du modèle. Deuxièmement, nous avons développé un algorithme de localisation des points d’intérêt du visage basé sur la minimisation d’une fonction objectif décrivant la corrélation entre les instances du modèle SFAM et les visages requête. Troisièmement, nous avons élaboré un réseau de croyance bayésien (BBN) dont la structure décrit les relations de dépendance entre les sujets, les expressions et les descripteurs faciaux. Les expressions faciales et les unités d’action sont alors modélisées comme les états du noeud correspondant à la variable expression et sont reconnues en identifiant le maximum de croyance pour tous les états. Nous avons également proposé une nouvelle approche pour l’inférence des paramètres du BBN utilisant un modèle de caractéristiques faciales pouvant être considéré comme une extension de SFAM. Finalement, afin d’enrichir l’information utilisée pour l’analyse de visages 3D, et particulièrement pour la reconnaissance de l’expression faciale, nous avons également élaboré un descripteur de visages 3D, nommé SGAND, pour caractériser les propriétés géométriques d’un point par rapport à son voisinage dans le nuage de points représentant un visage 3D. L’efficacité de ces méthodes a été évaluée sur les bases FRGC, BU3DFE et Bosphorus pour la localisation des points d’intérêt ainsi que sur les bases BU3DFE et Bosphorus pour la reconnaissance des expressions faciales et des unités d’action. / This Ph.D thesis work is dedicated to automatic facial analysis in 3D, including facial landmarking and facial expression recognition. Indeed, facial expression plays an important role both in verbal and non verbal communication, and in expressing emotions. Thus, automatic facial expression recognition has various purposes and applications and particularly is at the heart of "intelligent" human-centered human/computer(robot) interfaces. Meanwhile, automatic landmarking provides aprior knowledge on location of face landmarks, which is required by many face analysis methods such as face segmentation and feature extraction used for instance for expression recognition. The purpose of this thesis is thus to elaborate 3D landmarking and facial expression recognition approaches for finally proposing an automatic facial activity (facial expression and action unit) recognition solution.In this work, we have proposed a Bayesian Belief Network (BBN) for recognizing facial activities, such as facial expressions and facial action units. A StatisticalFacial feAture Model (SFAM) has also been designed to first automatically locateface landmarks so that a fully automatic facial expression recognition system can be formed by combining the SFAM and the BBN. The key contributions are the followings. First, we have proposed to build a morphable partial face model, named SFAM, based on Principle Component Analysis. This model allows to learn boththe global variations in face landmark configuration and the local ones in terms of texture and local geometry around each landmark. Various partial face instances can be generated from SFAM by varying model parameters. Secondly, we have developed a landmarking algorithm based on the minimization an objective function describing the correlation between model instances and query faces. Thirdly, we have designed a Bayesian Belief Network with a structure describing the casual relationships among subjects, expressions and facial features. Facial expression oraction units are modelled as the states of the expression node and are recognized by identifying the maximum of beliefs of all states. We have also proposed a novel method for BBN parameter inference using a statistical feature model that can beconsidered as an extension of SFAM. Finally, in order to enrich information usedfor 3D face analysis, and particularly 3D facial expression recognition, we have also elaborated a 3D face feature, named SGAND, to characterize the geometry property of a point on 3D face mesh using its surrounding points.The effectiveness of all these methods has been evaluated on FRGC, BU3DFEand Bosphorus datasets for facial landmarking as well as BU3DFE and Bosphorus datasets for facial activity (expression and action unit) recognition.

For the west coast of North America, from northern California to southern Washington, a habitat suitability prediction framework was developed to support wave energy device siting. Concern that wave energy devices may impact the seafloor and benthos has renewed research interest in the distribution of marine benthic invertebrates and factors influencing their distribution. A Bayesian belief network approach was employed for learning species-habitat associations for Rhabdus rectius, a tusk-shaped marine infaunal Mollusk. Environmental variables describing surficial geology and water depth were found to be most influential to the distribution of R. rectius. Water property variables, such as temperature and salinity, were less influential as distribution predictors. Species-habitat associations were used to predict habitat suitability probabilities for R. rectius, which were then mapped over an area of interest along the south-central Oregon coast. Habitat suitability prediction models tested well against data withheld for crossvalidation supporting our conclusion that Bayesian learning extracts useful information available in very small, incomplete data sets and identifies which variables drive habitat suitability for R. rectius. Additionally, Bayesian belief networks are easily updated with new information, quantitative or qualitative, which provides a flexible mechanism for multiple scenario analyses. The prediction framework presented here is a practical tool informing marine spatial planning assessment through visualization of habitat suitability. / Graduation date: 2012

Advances in unsupervised learning and deep neural networks have led to increased performance in a number of domains, and to the ability to draw strong comparisons between the biological method of self-organization conducted by the brain and computational mechanisms. This thesis aims to use real-world data to tackle two areas in the domain of computer vision which have biological equivalents: feature detection and motion tracking. The aforementioned advances have allowed efficient learning of feature representations directly from large sets of unlabeled data instead of using traditional handcrafted features. The first part of this thesis evaluates such representations by comparing regularization and preprocessing methods which incorporate local neighbouring information during training on a single-layer neural network. The networks are trained and tested on the Hollywood2 video dataset, as well as the static CIFAR-10, STL-10, COIL-100, and MNIST image datasets. The induction of topography or simple image blurring via Gaussian filters during training produces better discriminative features as evidenced by the consistent and notable increase in classification results that they produce. In the visual domain, invariant features are desirable such that objects can be classified despite transformations. It is found that most of the compared methods produce more invariant features, however, classification accuracy does not correlate to invariance. The second, and paramount, contribution of this thesis is a biologically-inspired model to explain the emergence of motion tracking behaviour in early development using unsupervised learning. The model’s self-organization is biased by an original concept called retinal constancy, which measures how similar visual contents are between successive frames. In the proposed two-layer deep network, when exposed to real-world video, the first layer learns to encode visual motion, and the second layer learns to relate that motion to gaze movements, which it perceives and creates through bi-directional nodes. This is unique because it uses general machine learning algorithms, and their inherent generative properties, to learn from real-world data. It also implements a biological theory and learns in a fully unsupervised manner. An analysis of its parameters and limitations is conducted, and its tracking performance is evaluated. Results show that this model is able to successfully follow targets in real-world video, despite being trained without supervision on real-world video.