The aim of this work is to introduce modular processing mechanisms for cortical functions implemented in networks of spiking neurons. Neural maps are a feature of cortical processing found to be generic throughout sensory cortical areas, and self-organisation to the fundamental properties of input spike trains has been shown to be an important property of cortical organisation. Additionally, oscillatory behaviour, temporal coding of information, and learning through spike timing dependent plasticity are all frequently observed in the cortex. The traditional self-organising map (SOM) algorithm attempts to capture the computational properties of this cortical self-organisation in a neural network. As such, a cognitive module for a spiking SOM using oscillations, phasic coding and STDP has been implemented. This model is capable of mapping to distributions of input data in a manner consistent with the traditional SOM algorithm, and of categorising generic input data sets. Higher-level cortical processing areas appear to feature a hierarchical category structure that is founded on a feature-based object representation. The spiking SOM model is therefore extended to facilitate input patterns in the form of sets of binary feature-object relations, such as those seen in the field of formal concept analysis. It is demonstrated that this extended model is capable of learning to represent the hierarchical conceptual structure of an input data set using the existing learning scheme. Furthermore, manipulations of network parameters allow the level of hierarchy used for either learning or recall to be adjusted, and the network is capable of learning comparable representations when trained with incomplete input patterns. Together these two modules provide related approaches to the generation of both topographic mapping and hierarchical representation of input spaces that can be potentially combined and used as the basis for advanced spiking neuron models of the learning of complex representations.

Natural images arise from complicated processes involving many factors of variation. They reflect the wealth of shapes and appearances of objects in our three-dimensional world, but they are also affected by factors such as distortions due to perspective, occlusions, and illumination, giving rise to structure with regularities at many different levels. Prior knowledge about these regularities and suitable representations that allow efficient reasoning about the properties of a visual scene are important for many image processing and computer vision tasks. This thesis focuses on models of image structure at intermediate levels of complexity as required, for instance, for image inpainting or segmentation. It aims at developing generative, probabilistic models of this kind of structure, and, in particular, at devising strategies for learning such models in a largely unsupervised manner from data. One hallmark of natural images is that they can often be decomposed into regions with very different visual characteristics. The main approach of this thesis is therefore to represent images in terms of regions that are characterized by their shapes and appearances, and an image is then composed from many such regions. We explore approaches to learn about the appearance of regions, to learn about region shapes, and ways to combine several regions to form a full image. To achieve this goal, we make use of some ideas for unsupervised learning developed in the literature on models of low-level image structure and in the “deep learning” literature. These models are used as building blocks of more structured model formulations that incorporate additional prior knowledge of how images are formed. The thesis makes the following contributions: Firstly, we investigate a popular, MRF based prior of natural image structure, the Field-of Experts, with respect to its ability to model image textures, and propose an extended formulation that is considerably more successful at this task. This formulation gives rise to a fully parametric, translation-invariant probabilistic generative model of image textures. We illustrate how this model can be used as a component of a more comprehensive model of images comprising multiple textured regions. Secondly, we develop a model of region shape. This work is an extension of the “Masked Restricted Boltzmann Machine” proposed by Le Roux et al. (2011) and it allows explicit reasoning about the independent shapes and relative depths of occluding objects. We develop an inference and unsupervised learning scheme and demonstrate how this shape model, in combination with the masked RBM gives rise to a good model of natural image patches. Finally, we demonstrate how this model of region shape can be extended to model shapes in large images. The result is a generative model of large images which are formed by composition from many small, partially overlapping and occluding objects.

"This thesis focuses on two key problems in reinforcement learning: How to design reward functions to obtain intended behaviors in autonomous systems using the learning-based control? Given complex mission specification, how to shape the reward function to achieve fast convergence and reduce sample complexity while learning the optimal policy? To answer these questions, the first part of this thesis investigates inverse reinforcement learning (IRL) method with a purpose of learning a reward function from expert demonstrations. However, existing algorithms often assume that the expert demonstrations are generated by the same reward function. Such an assumption may be invalid as one may need to aggregate data from multiple experts to obtain a sufficient set of demonstrations. In the first and the major part of the thesis, we develop a novel method, called Non-parametric Behavior Clustering IRL. This algorithm allows one to simultaneously cluster behaviors while learning their reward functions from demonstrations that are generated from more than one expert/behavior. Our approach is built upon the expectation-maximization formulation and non-parametric clustering in the IRL setting. We apply the algorithm to learn, from driving demonstrations, multiple driver behaviors (e.g., aggressive vs. evasive driving behaviors). In the second task, we study whether reinforcement learning can be used to generate complex behaviors specified in formal logic — Linear Temporal Logic (LTL). Such LTL tasks may specify temporally extended goals, safety, surveillance, and reactive behaviors in a dynamic environment. We introduce reward shaping under LTL constraints to improve the rate of convergence in learning the optimal and probably correct policies. Our approach exploits the relation between reward shaping and actor-critic methods for speeding up the convergence and, as a consequence, reducing training samples. We integrate compositional reasoning in formal methods with actor-critic reinforcement learning algorithms to initialize a heuristic value function for reward shaping. This initialization can direct the agent towards efficient planning subject to more complex behavior specifications in LTL. The investigation takes the initial step to integrate machine learning with formal methods and contributes to building highly autonomous and self-adaptive robots under complex missions."

Zero-resource speech processing is a growing research area which aims to develop methods that can discover linguistic structure and representations directly from unlabelled speech audio. Such unsupervised methods would allow speech technology to be developed in settings where transcriptions, pronunciation dictionaries, and text for language modelling are not available. Similar methods are required for cognitive models of language acquisition in human infants, and for developing robotic applications that are able to automatically learn language in a novel linguistic environment. There are two central problems in zero-resource speech processing: (i) finding frame-level feature representations which make it easier to discriminate between linguistic units (phones or words), and (ii) segmenting and clustering unlabelled speech into meaningful units. The claim of this thesis is that both top-down modelling (using knowledge of higher-level units to to learn, discover and gain insight into their lower-level constituents) as well as bottom-up modelling (piecing together lower-level features to give rise to more complex higher-level structures) are advantageous in tackling these two problems. The thesis is divided into three parts. The first part introduces a new autoencoder-like deep neural network for unsupervised frame-level representation learning. This correspondence autoencoder (cAE) uses weak top-down supervision from an unsupervised term discovery system that identifies noisy word-like terms in unlabelled speech data. In an intrinsic evaluation of frame-level representations, the cAE outperforms several state-of-the-art bottom-up and top-down approaches, achieving a relative improvement of more than 60% over the previous best system. This shows that the cAE is particularly effective in using top-down knowledge of longer-spanning patterns in the data; at the same time, we find that the cAE is only able to learn useful representations when it is initialized using bottom-up pretraining on a large set of unlabelled speech. The second part of the thesis presents a novel unsupervised segmental Bayesian model that segments unlabelled speech data and clusters the segments into hypothesized word groupings. The result is a complete unsupervised tokenization of the input speech in terms of discovered word types|the system essentially performs unsupervised speech recognition. In this approach, a potential word segment (of arbitrary length) is embedded in a fixed-dimensional vector space. The model, implemented as a Gibbs sampler, then builds a whole-word acoustic model in this embedding space while jointly performing segmentation. We first evaluate the approach in a small-vocabulary multi-speaker connected digit recognition task, where we report unsupervised word error rates (WER) by mapping the unsupervised decoded output to ground truth transcriptions. The model achieves around 20% WER, outperforming a previous HMM-based system by about 10% absolute. To achieve this performance, the acoustic word embedding function (which maps variable-duration segments to single vectors) is refined in a top-down manner by using terms discovered by the model in an outer loop of segmentation. The third and final part of the study extends the small-vocabulary system in order to handle larger vocabularies in conversational speech data. To our knowledge, this is the first full-coverage segmentation and clustering system that is applied to large-vocabulary multi-speaker data. To improve efficiency, the system incorporates a bottom-up syllable boundary detection method to eliminate unlikely word boundaries. We compare the system on English and Xitsonga datasets to several state-of-the-art baselines. We show that by imposing a consistent top-down segmentation while also using bottom-up knowledge from detected syllable boundaries, both single-speaker and multi-speaker versions of our system outperform a purely bottom-up single-speaker syllable-based approach. We also show that the discovered clusters can be made less speaker- and gender-specific by using features from the cAE (which incorporates both top-down and bottom-up learning). The system's discovered clusters are still less pure than those of two multi-speaker unsupervised term discovery systems, but provide far greater coverage. In summary, the different models and systems presented in this thesis show that both top-down and bottom-up modelling can improve representation learning, segmentation and clustering of unlabelled speech data.

The current trend in visual recognition research is to place a strict division between the supervised and unsupervised learning paradigms, which is problematic for two main reasons. On the one hand, supervised methods require training data for each and every category that the system learns; training data may not always be available and is expensive to obtain. On the other hand, unsupervised methods must determine the optimal visual cues and distance metrics that distinguish one category from another to group images into semantically meaningful categories; however, for unlabeled data, these are unknown a priori. I propose a visual category discovery framework that transcends the two paradigms and learns accurate models with few labeled exemplars. The main insight is to automatically focus on the prevalent objects in images and videos, and learn models from them for category grouping, segmentation, and summarization. To implement this idea, I first present a context-aware category discovery framework that discovers novel categories by leveraging context from previously learned categories. I devise a novel object-graph descriptor to model the interaction between a set of known categories and the unknown to-be-discovered categories, and group regions that have similar appearance and similar object-graphs. I then present a collective segmentation framework that simultaneously discovers the segmentations and groupings of objects by leveraging the shared patterns in the unlabeled image collection. It discovers an ensemble of representative instances for each unknown category, and builds top-down models from them to refine the segmentation of the remaining instances. Finally, building on these techniques, I show how to produce compact visual summaries for first-person egocentric videos that focus on the important people and objects. The system leverages novel egocentric and high-level saliency features to predict important regions in the video, and produces a concise visual summary that is driven by those regions. I compare against existing state-of-the-art methods for category discovery and segmentation on several challenging benchmark datasets. I demonstrate that we can discover visual concepts more accurately by focusing on the prevalent objects in images and videos, and show clear advantages of departing from the status quo division between the supervised and unsupervised learning paradigms. The main impact of my thesis is that it lays the groundwork for building large-scale visual discovery systems that can automatically discover visual concepts with minimal human supervision. / text

This thesis proposes and tests a recursive, adpative, and computationally inexpensive method for partitioning real-number spaces. When tested for proof-of-concept on both one- and two- dimensional classification and control problems, the Fern algorithm was found to work well in one dimension, moderately well for two-dimensional classification, and not at all for two-dimensional control. Testing ferns as pure discretizers - which would involve a secondary discrete learner - has been left to future work. / text

Unsupervised learning has become more and more important due to the recent explosion of data. Clustering, a key topic in unsupervised learning, is a well-studied task arising in many applications ranging from computer vision to computational biology to the social sciences. This thesis is a collection of work exploring two modern aspects of clustering: stability and scalability. In the first part, we study clustering under a stability property called perturbation resilience. As an alternative approach to worst case analysis, this novel theoretical framework aims at understanding the complexity of clustering instances that satisfy natural stability assumptions. In particular, we show how to correctly cluster instances whose optimal solutions are resilient to small multiplicative perturbations on the distances between data points, significantly improving existing guarantees. We further propose a generalized property that allows small changes in the optimal solutions after perturbations, and provide the first known positive results in this more challenging setting. In the second part, we study the problem of clustering large scale data distributed across nodes which communicate over the edges of a connected graph. We provide algorithms with small communication cost and provable guarantees on the clustering quality. We also propose algorithms for distributed principal component analysis, which can be used to reduce the communication cost of clustering high dimensional data while merely comprising the clustering quality. In the third part, we study community detection, the modern extension of clustering to network data. We propose a theoretical model of communities that are stable in the presence of noisy nodes in the network, and design an algorithm that provably detects all such communities. We also provide a local algorithm for large scale networks, whose running time depends on the sizes of the output communities but not that of the entire network.

Many high-dimensional time-varying signals can be modeled as a sequence of noisy nonlinear observations of a low-dimensional dynamical process. Given high-dimensional observations and a distribution describing the dynamical process, we present a computationally inexpensive approximate algorithm for estimating the inverse of this mapping. Once this mapping is learned, we can invert it to construct a generative model for the signals. Our algorithm can be thought of as learning a manifold of images by taking into account the dynamics underlying the low-dimensional representation of these images. It also serves as a nonlinear system identification procedure that estimates the inverse of the observation function in nonlinear dynamic system. Our algorithm reduces to a generalized eigenvalue problem, so it does not suffer from the computational or local minimum issues traditionally associated with nonlinear system identification, allowing us to apply it to the problem of learning generative models for video sequences.

Representing and handling high-dimensional data has been increasingly ubiquitous in many real world-applications, such as computer vision, machine learning, and data mining. High-dimensional data usually have intrinsic low-dimensional structures, which are suitable for subsequent data processing. As a consequent, it has been a common demand to find low-dimensional data representations in many machine learning and data mining problems. Factorization methods have been impressive in recovering intrinsic low-dimensional structures of the data. When seeking low-dimensional representation of the data, traditional methods mainly face two challenges: 1) how to discover the most variational features/information from the data; 2) how to measure accurate nonlinear relationships of the data. As a solution to these challenges, traditional methods usually make use of a two-step approach by performing feature selection and manifold construction followed by further data processing, which omits the dependence between these learning tasks and produce inaccurate data representation. To resolve these problems, we propose to integrate feature learning and graph learning with factorization model, which allows the goals of learning features, constructing manifold, and seeking new data representation to mutually enhance and lead to powerful data representation capability. Moreover, it has been increasingly common that 2-dimensional (2D) data often have high dimensions of features, where each example of 2D data is a matrix with its elements being features. For such data, traditional data usually convert them to 1-dimensional vectorial data before data processing, which severely damages inherent structures of such data. We propose to directly use 2D data for seeking new representation, which enables the model to preserve inherent 2D structures of the data. We propose to seek projection directions to find the subspaces, in which spatial information is maximumly preserved. Also, manifold and new data representation are learned in these subspaces, such that the manifold are clean and the new representation is discriminative. Consequently, seeking projections, learning manifold and constructing new representation mutually enhance and lead to powerful data representation technique.

The present thesis addresses machine learning in a domain of naturallanguage phrases that are names of universities. It describes two approaches to this problem and a software implementation that has made it possible to evaluate them and to compare them. In general terms, the system's task is to learn to 'understand' the significance of the various components of a university name, such as the city or region where the university is located, the scienti c disciplines that are studied there, or the name of a famous person which may be part of the university name. A concrete test for whether the system has acquired this understanding is when it is able to compose a plausible university name given some components that should occur in the name. In order to achieve this capability, our system learns the structure of available names of some universities in a given data set, i.e. it acquires a grammar for the microlanguage of university names. One of the challenges is that the system may encounter ambiguities due to multi meaning words. This problem is addressed using a small ontology that is created during the training phase. Both domain knowledge and grammatical knowledge is represented using decision trees, which is an ecient method for concept learning. Besides for inductive inference, their role is to partition the data set into a hierarchical structure which is used for resolving ambiguities. The present report also de nes some modi cations in the de nitions of parameters, for example a parameter for entropy, which enable the system to deal with cognitive uncertainties. Our method for automatic syntax acquisition, ADIOS, is an unsupervised learning method. This method is described and discussed here, including a report on the outcome of the tests using our data set. The software that has been implemented and used in this project has been implemented in C.