Search

Rapid advances in infrastructure health monitoring and sensing technologies allow monitoring of assets continuously and in real-time throughout their life span. However, smart and automated techniques for decision-making (e.g. maintaining or improving infrastructure performance) are in their infancy. The revolution in new sensing capabilities has led to rapidly increasing volumes of data, which makes traditional data analysis techniques inadequate. Adoption of Big Data (BD) analytics and Artificial Intelligence (AI) techniques is urgently needed to automatically integrate information from multiple sensors, extract knowledge and inform decision-making. A novel damage identification and localisation approach is proposed, which utilises multi-level data fusion and anomaly detection techniques. Numerical models of a simply supported bridge and a continuous bridge were simulated for data generation. Accelerations, deflections and bending moments obtained at multiple sensor locations when the bridge is subjected to a moving vehicle were used as the input. A damagesensitive feature was established via coupling principal component analysis with the Mahalanobis distance, allowing for initial data dimensionality reduction and information integration. Anomaly detection using a deep convolutional autoencoder was performed to identify the presence of damage on the bridge. It is demonstrated that the proposed approach is independent of the mass and speed of the moving vehicles. The approach is highly successful in identifying and localising damage(s), even for extremely small damage severity (e.g. 1% reduction in the second moment of area of an element with 0.5m in length in the simply supported bridge, and 15% reduction in the second moment of area of an element with 0.166m in length in the continuous bridge), and multiple damaged cases on the bridges, achieving accuracies of over 93% for all tested cases of the simply supported bridge and over 91% for the majority of the tested cases of the continuous bridge. The proposed damage localisation approach can accurately determine damage location(s) for both single and multiple damage cases. It is shown how the proposed framework can be adopted for early damage detection to enable reliable decision-making and maximise structural safety.

While short-term interactions with robots benefit from the novelty effect, long-term interactions can suffer from a decrease in user interest and engagement. Based on the previous research within Human-Robot Interaction (HRI), the thesis presented here is that user experience in long-term human-robot interactions can be improved by personalising the interaction through recognising users and recalling previously learned information. User recognition is the first step towards personalising the interaction, however, there does not exist a reliable user recognition method for fully autonomous user recognition in long- term HRI for real-world applications. Correspondingly, this thesis proposes a Multi-modal Incremental Bayesian Network (MMIBN) model, which combines face recognition with soft biometrics and allows continuous, incremental and online learning of users, without the need for any preliminary training. We validated the robustness and reliability of this approach with a long-term (4-weeks) real-world study with 14 users and an artificially generated multi-modal long-term user recognition dataset with 200 users. Following on from this work, we explored personalisation of the interaction in service robotics and socially assistive robotics, based on earlier evidence for the impact of personalisation on long-term interactions. We created the text-based Barista Datasets that contain simulated generic and personalised dialogues for interactions with a barista that recall and suggest user preferences in subsequent interactions in a coffee shop. Based on these datasets, we designed fully autonomous barista robots with MMIBN, automatic speech recognition (ASR) and a rule-based dialogue manager, and evaluated these robots with a real-world long-term (5-day) study with 18 non-native English speakers. The study demonstrated that personalisation mitigates negative user experiences that arise from unreliable speech recognition and the inflexible structure of the rule-based dialogue manager. Consequently, we explored the potential of the state-of-the-art data-driven dialogue models based on the Barista Datasets. The results showed that while data-driven models perform remarkably well in generic task-oriented dialogue, no model could perform sufficiently well for personalisation in long-term interactions. Lastly, to demonstrate the real-world bene ts of long-term HRI, we design a personalised robot to improve user motivation and adherence to the cardiac rehabilitation programme, and evaluate with a study that ran for 2.5 years at a hospital in Colombia. The robot individually tracked the patients’ health progress and attendance throughout the programme, and provided personalised and immediate feedback based on continuous monitoring for 18 weeks. While the study could not be completed due to the outbreak of COVID-19, our initial findings (with 6 patients) showed that user engagement and motivation for the therapy and adherence were improved and maintained in the long-term interactions. Overall, the work undertaken provides supporting evidence for our thesis and contributes fundamental stepping stones for future research in personalised long-term HRI to develop robots that can meet and maintain user expectations.

The ability to organize concepts (e.g., dog, chair) into efficient mental representations, i.e., categories (e.g., animal, furniture) is a fundamental mechanism which allows humans to perceive, organize, and adapt to their world. Much research has been dedicated to the questions of how categories emerge and how they are represented. Experimental evidence suggests that (i) concepts and categories are represented through sets of features (e.g., dogs bark, chairs are made of wood) which are structured into different types (e.g, behavior, material); (ii) categories and their featural representations are learnt jointly and incrementally; and (iii) categories are dynamic and their representations adapt to changing environments. This thesis investigates the mechanisms underlying the incremental and dynamic formation of categories and their featural representations through cognitively motivated Bayesian computational models. Models of category acquisition have been extensively studied in cognitive science and primarily tested on perceptual abstractions or artificial stimuli. In this thesis, we focus on categories acquired from natural language stimuli, using nouns as a stand-in for their reference concepts, and their linguistic contexts as a representation of the concepts’ features. The use of text corpora allows us to (i) develop large-scale unsupervised models thus simulating human learning, and (ii) model child category acquisition, leveraging the linguistic input available to children in the form of transcribed child-directed language. In the first part of this thesis we investigate the incremental process of category acquisition. We present a Bayesian model and an incremental learning algorithm which sequentially integrates newly observed data. We evaluate our model output against gold standard categories (elicited experimentally from human participants), and show that high-quality categories are learnt both from child-directed data and from large, thematically unrestricted text corpora. We find that the model performs well even under constrained memory resources, resembling human cognitive limitations. While lists of representative features for categories emerge from this model, they are neither structured nor jointly optimized with the categories. We address these shortcomings in the second part of the thesis, and present a Bayesian model which jointly learns categories and structured featural representations. We present both batch and incremental learning algorithms, and demonstrate the model’s effectiveness on both encyclopedic and child-directed data. We show that high-quality categories and features emerge in the joint learning process, and that the structured features are intuitively interpretable through human plausibility judgment evaluation. In the third part of the thesis we turn to the dynamic nature of meaning: categories and their featural representations change over time, e.g., children distinguish some types of features (such as size and shade) less clearly than adults, and word meanings adapt to our ever changing environment and its structure. We present a dynamic Bayesian model of meaning change, which infers time-specific concept representations as a set of feature types and their prevalence, and captures their development as a smooth process. We analyze the development of concept representations in their complexity over time from child-directed data, and show that our model captures established patterns of child concept learning. We also apply our model to diachronic change of word meaning, modeling how word senses change internally and in prevalence over centuries. The contributions of this thesis are threefold. Firstly, we show that a variety of experimental results on the acquisition and representation of categories can be captured with computational models within the framework of Bayesian modeling. Secondly, we show that natural language text is an appropriate source of information for modeling categorization-related phenomena suggesting that the environmental structure that drives category formation is encoded in this data. Thirdly, we show that the experimental findings hold on a larger scale. Our models are trained and tested on a larger set of concepts and categories than is common in behavioral experiments and the categories and featural representations they can learn from linguistic text are in principle unrestricted.

Feature extraction is an extremely important pre-processing step to pattern recognition, and machine learning problems. This thesis highlights how one can best extract features from the data in an exhaustively online and purely adaptive manner. The solution to this problem is given for both labeled and unlabeled datasets, by presenting a number of novel on-line learning approaches. Specifically, the differential equation method for solving the generalized eigenvalue problem is used to derive a number of novel machine learning and feature extraction algorithms. The incremental eigen-solution method is used to derive a novel incremental extension of linear discriminant analysis (LDA). Further the proposed incremental version is combined with extreme learning machine (ELM) in which the ELM is used as a preprocessor before learning. In this first key contribution, the dynamic random expansion characteristic of ELM is combined with the proposed incremental LDA technique, and shown to offer a significant improvement in maximizing the discrimination between points in two different classes, while minimizing the distance within each class, in comparison with other standard state-of-the-art incremental and batch techniques. In the second contribution, the differential equation method for solving the generalized eigenvalue problem is used to derive a novel state-of-the-art purely incremental version of slow feature analysis (SLA) algorithm, termed the generalized eigenvalue based slow feature analysis (GENEIGSFA) technique. Further the time series expansion of echo state network (ESN) and radial basis functions (EBF) are used as a pre-processor before learning. In addition, the higher order derivatives are used as a smoothing constraint in the output signal. Finally, an online extension of the generalized eigenvalue problem, derived from James Stone’s criterion, is tested, evaluated and compared with the standard batch version of the slow feature analysis technique, to demonstrate its comparative effectiveness. In the third contribution, light-weight extensions of the statistical technique known as canonical correlation analysis (CCA) for both twinned and multiple data streams, are derived by using the same existing method of solving the generalized eigenvalue problem. Further the proposed method is enhanced by maximizing the covariance between data streams while simultaneously maximizing the rate of change of variances within each data stream. A recurrent set of connections used by ESN are used as a pre-processor between the inputs and the canonical projections in order to capture shared temporal information in two or more data streams. A solution to the problem of identifying a low dimensional manifold on a high dimensional dataspace is then presented in an incremental and adaptive manner. Finally, an online locally optimized extension of Laplacian Eigenmaps is derived termed the generalized incremental laplacian eigenmaps technique (GENILE). Apart from exploiting the benefit of the incremental nature of the proposed manifold based dimensionality reduction technique, most of the time the projections produced by this method are shown to produce a better classification accuracy in comparison with standard batch versions of these techniques - on both artificial and real datasets.

This research work develops real-time incremental learning decision tree solutions suitable for real-time embedded systems by virtue of having both a defined memory requirement and an upper bound on the computation time per training vector. In addition, the work provides embedded systems with the capabilities of rapid processing and training of streamed data problems, and adopts electronic hardware solutions to improve the performance of the developed algorithm. Two novel decision tree approaches, namely the Multi-Dimensional Frequency Table (MDFT) and the Hashed Frequency Table Decision Tree (HFTDT) represent the core of this research work. Both methods successfully incorporate a frequency table technique to produce a complete decision tree. The MDFT and HFTDT learning methods were designed with the ability to generate application specific code for both training and classification purposes according to the requirements of the targeted application. The MDFT allows the memory architecture to be specified statically before learning takes place within a deterministic execution time. The HFTDT method is a development of the MDFT where a reduction in the memory requirements is achieved within a deterministic execution time. The HFTDT achieved low memory usage when compared to existing decision tree methods and hardware acceleration improved the performance by up to 10 times in terms of the execution time.

Session-based recommendation has become a hot topic of intelligent system in recent years. As a sub-field of Recommending System, the session-based recommendation studies the sequential relationship of data in user's usage sessions. In some applications, the recommending system should focus more on the personalized usage feature in order to make better recommendations. This thesis analyzed the statistics of user's usage sessions and proposed the Statistics Enhanced FPMC algorithm to enhance the personalized usage pattern of users to improve the recommending performance of recommender system for in-vehicle infotainment system and APP manage system application. The proposed algorithm also addressed the user cold-start problem by incremental learning with a knowledge distillation method to alleviate the catastrophic forgetting problem. The user cold-start problem is defined as making recommendations to new users under cold-start conditions. While the usage data becomes available for new users, the model can continue to be updated to improve recommending performance.

Abstract: When speakers are asked to name a series of semantically related pictures (e.g., apple, pear, banana), response times increase as more pictures are named. This effect is known as cumulative semantic interference (CSI). Researchers who have studied these effects in adults have proposed that these interference effects are due to durable changes in the strength of representation of a word, otherwise known as incremental learning (Howard et al., 2006; Oppenheim, Dell & Schwartz, 2010). This effect has been well documented in adults, but few studies have focused on this phenomenon in children. This gap invited further investigation as cumulative semantic interference has the potential to offer a window into how children strengthen or change their representation and processing of words. The present study examines cumulative semantic interference in preschool-aged children using a continuous list method with unrelated, intervening items. Forty children (17 male, Mage = 3 years; 11 months, range 3;0 – 4;11) named 100 images, consisting of 72 experimental items (six exemplars drawn from 12 different semantic categories) and 28 filler items. There were two to eight intervening trials between related category members. Potential factors impacting cumulative semantic interference effects, such as vocabulary size, conceptual organization and inhibitory control, were also examined. Children demonstrated robust cumulative semantic interference effects as evidenced by increased response latencies and decreased response accuracy for each successive category image. The effects were persisting and independent of intervening naming experiences, consistent with the incremental learning account. Exploratory analyses were completed to investigate factors contributing to individual differences. Two variables were identified to have moderating effects: vocabulary size and age. As children’s vocabulary scores increased, predicted slopes suggested worsening accuracy over ordinal position; however, the degree of worsening was attenuated as compared to children with poorer vocabularies. Older children demonstrated a greater increase in reaction times as ordinal position increased, thereby indicating stronger cumulative semantic interference effects. This study supports the application of the incremental learning account to cumulative semantic interference effects in preschool aged children. While only vocabulary and age were identified as moderating factors, this indicates there are certain abilities and knowledge that contribute to the presence of these effects. This study provides important insights into lexical representation and processing in typical language development, and offers a foundation for future work with children with language delays.

Sound source localization is essential in robotics, which broadens the possibilities of human-robot interactions by enriching the robot's perceptual capabilities. Localizing an acoustic source in a complex indoor environment is especially challenging due to the high noise-to-signal ratio and reverberations. In this thesis, we present an incremental learning framework for mobile robots localizing the human sound source using a microphone array in a complex indoor environment consisting of multiple rooms. The framework allows robots to accumulate training data and improve the performance of the prediction model over time using an incremental learning scheme. A self-supervision process is developed such that the model ranks the priority of rooms to explore, assigns the ground truth label to the collected data, and updates the learned model on-the-fly. In experiments, we demonstrate that the framework can be directly deployed in real-world scenarios without extra human interventions, and can localize the sound source successfully.

Currently, the healthcare industry uses Big Data for essential patient care information. Electronic Health Records (EHR) store massive data and are continuously updated with information such as laboratory results, medication, and clinical events. There are various methods by which healthcare data is generated and collected, including databases, healthcare websites, mobile applications, wearable technologies, and sensors. The continuous flow of data will improve healthcare service, medical diagnostic research and, ultimately, patient care. Thus, it is important to implement advanced data analysis techniques to obtain more precise prediction results.Machine Learning (ML) has acquired an important place in Big Healthcare Data (BHD). ML has the capability to run predictive analysis, detect patterns or red flags, and connect dots to enhance personalized treatment plans. Because predictive models have dependent and independent variables, ML algorithms perform mathematical calculations to find the best suitable mathematical equations to predict dependent variables using a given set of independent variables. These model performances depend on datasets and response, or dependent, variable types such as binary or multi-class, supervised or unsupervised.The current research analyzed incremental, or streaming or online, algorithm performance with offline or batch learning (these terms are used interchangeably) using performance measures such as accuracy, model complexity, and time consumption. Batch learning algorithms are provided with the specific dataset, which always constrains the size of the dataset depending on memory consumption. In the case of incremental algorithms, data arrive sequentially, which is determined by hyperparameter optimization such as chunk size, tree split, or hoeffding bond. The model complexity of an incremental learning algorithm is based on a number of parameters, which in turn determine memory consumption.

The widespread usage of wearable sensors such as smart watches provide access to valuable objective physiological (such as Electrocardiogram(ECG)) signals ubiquitously. Healthcare domain has been tremendously benefited by the collection of physiological signals which can be used for health monitoring of patients. The signals from the wearable sensors enabled the researchers and data experts to process them and identify the human physiological state by classifying the human activities. This led to the growth and development of smart ecosystem in the healthcare domain.In this thesis, ECG signals have been investigated as the physiological measure to detect human activities. Various measures are extracted from ECG, such as heart rate variability, average heart rate etc. and their relationships with different human activities are investigated. To build a comprehensive analytical machine learning model for ECG signals and to enable the continuous monitoring of humans, one would need access to real time streaming of continuous data. So, the data would be unsupervised most of the time and it would be very expensive (almost practically impossible) to label all the data streaming in real time. Also, it is highly probable that the data is collected from different sessions and varying situations. Therefore, the machine learning models need to be able to adapt to new sessions. This would be a major challenge in human state monitoring provided that the conventional predictive models work only on the stationary data. Also, these models would fail to work on the data from multiple sessions. To provide a practical solution to address above issues, two advanced methods in machine learning have been discussed in this research: Incremental learning and Semi supervised learning. Incremental learning is a paradigm in Machine learning where the stream of input data is continuously used to extend the existing knowledge learnt by the model. The incremental learning module has been built in Apache Spark platform which provides a scalable cloud infrastructure to apply machine learning algorithms on streaming data. Semi supervised learning is another solution implemented in this thesis where some out of all the data points are labelled. Different semi supervised algorithms have been studied and applied which learn the relationship between features and adapts the model to data from multiple sessions. Finally, the results are compared and the implementation ideas for the discussed solutions have been proposed.

Incremental and parallel are two capabilities for machine learning algorithms to accommodate data from real world applications. Incremental learning addresses streaming data by constructing a learning model that is updated continuously in response to newly arrived samples. To solve the computational problems posed by large data sets, parallel learning distributes the computational efforts among multiple nodes within a cloud or cluster to speed up the calculation. With the rise of BigData, data become simultaneously large scale and streaming, which is the motivation to address incremental and parallel incremental (PI) learning in this work. Incremental and parallel are two capabilities for machine learning algorithms to This research first considers the incremental learning alone, in the graph max-flow/min-cut problem. An augmenting path based incremental max-flow algorithm is proposed. The proposed algorithm handles graph changes in a chunking manner, updating residual graph via augmentation and de-augmentation in response to edge capacity increase, decrease, edge/node adding and removal. The theoretical guarantee of our algorithm is that incremental max-flow is always equal to batch retraining. Experiments show the deterministic computational cost save (i.e., gain) of our algorithm with respect to batch retraining in handling graph edge adding. The proposed incremental max-flow is then applied to upgrade an existing batch semi-supervised learning algorithm known as graph minicuts to be incremental. In batch graph minicuts, a graph is learned from input labeled and unlabeled data, and then a min-cut is conducted on that graph to make the classification decision. In the proposed modification, the graph is updated dynamically for accommodating online data adding and retiring. Then the proposed incremental max-flow algorithm is adopted to learn min-cut from the resulting non-stationary graph. Empirical evaluation on real world data reveals that the proposed algorithm outperforms state-of-the-art stream classification algorithms. In the incremental max-flow, the training speed is not satisfactory when the data set is huge. A straightforward solution is to combine parallel data processing with incremental learning. Previously, parallel and incremental learning are often treated as two separate problems and solved one after another. Alternatively in this work, these two learning problems are solved in one process (i.e., PI integration). To simplify the learning, this research considers a base model in which incremental learning can be implemented by merging knowledge from incoming data and parallel learning can be performed by merging knowledge from simultaneous learners (i.e., in knowledge mergeable condition). As a result, this work develops a parallel incremental wESVM (weighted Extreme Support Vector Machine) algorithm, in which the parallel incremental learning of the base model is completed within a single process of knowledge merging. Specifically, the wESVM is reformulated such that knowledge from subsets of training data can be merged via simple matrix addition. As such, the proposed algorithm is able to conduct parallel incremental learning by merging knowledge from data slices arriving at each incremental stage. Both theoretical and experimental studies show the equivalence of the proposed algorithm to batch wESVM in terms of learning effectiveness. In particular, the algorithm demonstrates desired scalability and clear speed advantages to batch retraining. In the field of data stream knowledge discovery, this work investigates incremental machine learning and invents a wESVM-based parallel learning and incremental learning integrated system. The limitation of this work is that PI integration applies only to models that satisfy the knowledge mergeable condition. Future work should investigate how to release this constraint and expand PI integration to other models such as SVM and neural network.

Abstract: New machine learning methods and parallel computation opened the door to many applications in computer vision. While computer vision is progressing rapidly because of that, there are not as many and as successful real world applications in robotics. In this thesis, we investigate two possible causes of this problem and we provide potential solutions. In particular, using recurrent networks to deal with temporal information and using human robot interaction for incremental learning. Robotics is time dependent. The environment is perceived through time and tasks are being performed by time indexed trajectories. This important property has been generally neglected in the deep learning community. The main focus instead is on improving benchmarks on single image tasks or analyzing videos in batches. Consequently, real-time video analysis received less attention. But this is exactly what robots need to have. Processing single images will not provide sufficient information to observe the environment and processing a batch of images will be too delayed to be used for planning and decision-making in real-time. As a solution to this problem, we propose a recurrent fully convolutional neural network(RFCNN) for segmentation, which is very useful in different robotics scenarios. This type of network accepts a series of images ending with the current image to infer the segmentation for the last image. We showed how such networks can be designed and trained in an end-to-end fashion. An extensive set of experiments on several different architectures and benchmarks were made. We observed a consistent improvement by using RFCNN over non-recurrent counterparts. While deep learning methods are the most general machine learning solutions available, they still suffer from the change in the data distribution between training and test. Even though it is not limited to robotics, their effect is most apparent there. It is mostly due to the fact that robots need to learn complicated reasoning using a limited train data. This combination leads to severe overfitting which will become obvious during the test in a slightly different environment. To mitigate this problem, we suggest a novel paradigm in which new perception information can be thought to the robot through Human-Robot Interaction(HRI). A complete HRI system is developed that allows human- friendly communication using speech and gesture. An incremental learning method is designed to improve an object detection network by using human inputs. The system is tested on simulated data and real experiments with success. We showed that humans can easily teach new objects to the robot and the robot is able to learn and use this information later.

Abstract: Model-free off-policy temporal-difference (TD) algorithms form a powerful component of scalable predictive knowledge representation due to their ability to learn numerous counter- factual predictions in a computationally scalable manner. In this dissertation, we address and overcome two shortcomings of off-policy TD algorithms that preclude a widespread use in knowledge representation: they often produce high variance estimates and may become unstable. The issue of high variance occurs when off-policy TD algorithms use importance sampling (IS), a standard approach to adjusting estimates when the sample and the target distributions are different. IS performs this adjustment by scaling the estimate by the likelihood ratio, also known as the importance sampling ratio. However, this ratio also often produces high variance in the estimates. We provide two different approaches to overcome this shortcoming. The first approach relies on a well-known alternative to the ordinary form of IS, called the Weighted Importance Sampling (WIS), which is known to produce much less variance but has been under-utilized in reinforcement learning tasks. The main obstacle to using WIS in large-scale reinforcement learning tasks is that it is not known how to systematically extend a simple lookup-table-based estimator such as WIS to parametric function approximation with real-time computationally scalable updates. We overcome this obstacle by developing new techniques for deriving computationally efficient equivalent TD updates. This also allows us to simplify the derivations of some existing computationally efficient algorithms as well as produce new ones. In the second approach, we directly address the main factor underlying the high variance issue: the use of the IS ratios. Ratios play a key role in producing high variance estimation as they may vary drastically from one sample to another. We show that the ratios can be eliminated from the updates by varying the amount of bootstrapping, a sophisticated technique for allowing a spectrum of multi-step TD algorithms. The core idea is to allow the bootstrapping parameter, which controls the amount of bootstrapping, vary and shrink in a specific way. The elimination of ratios from the updates using this approach results directly in the reduction of variance compared to IS-based algorithms while still retaining the multi-step advantage of TD algorithms. We empirically evaluate the algorithms based on both approaches using several idealized experiments. We show that the WIS-based algorithms outperform exist- ing off-policy learning algorithms often with an order of magnitude margin. On the other hand, the bootstrapping-based methods can retain more effectively the multi-step advantage compared to the IS-based methods despite occasionally shrinking the bootstrapping parameter. The issue of instability with off-policy TD algorithms may appear when both function approximation and bootstrapping are used together, a desirable combination in reinforcement learning tasks. We discuss the key factors responsible for instability by analyzing the deterministic counterpart of TD updates. We illustrate that the on-policy TD algorithm may also be unstable if the updates are scaled or the bootstrapping parameter is set differently for different states but remains stable with a state-dependent discount factor. The TD algorithm can be made stable in all these cases by ensuring that a certain matrix involved in the deterministic TD update is always positive definite. We provide a systematic approach to ensuring positive definiteness of this matrix, which amounts to selectively emphasizing and de-emphasizing the updates of the stochastic counterpart in a particular manner. The resulting algorithm, which we call the emphatic TD algorithm, is stable, do not introduce extra tunable parameters and its additional memory or computational complexity is negligible. Finally, we provide computationally efficient methods to perform a search over features for improving the representation of learning systems. These search methods enable continual discovery and maintenance of relevant and useful predictions that can be used as a building block of scalable predictive knowledge representations in long-lived systems.

Understanding human activities in unconstrained natural videos is a widely studied problem, yet it remains as one of the most challenging problems in computer vision. It has many practical applications, including, but not limited to security, surveillance, and robotic vision. The primary requirement of these applications is to learn a recognition model that can differentiate among different types of activities. However, learning a recognition model requires lots of data and these data need to be labeled. Enormous amount of unlabeled videos are being generated from various sources. Labeling these video data is the biggest challenge that an activity understanding method encounters. State-of-the-art approaches generally learn fixed or static recognition model based on the assumptions that all the training data are labeled and available beforehand. But these assumptions are unrealistic for many applications where we have to deal with streaming videos or surveillance cameras. In such scenarios, new unlabeled video data may come over time and can be leveraged upon to incrementally improve the current model. In this thesis, we aim to develop frameworks for online activity understanding by taking advantage of unlabeled video data with a reduced labeling cost and without compromising performance.We present four distinct frameworks for continuous learning of activity recognition models. We leverage upon tools and techniques from various branches of machine learning and deep learning and effectively combine them together with computer vision in order to develop novel and efficient solutions. Our proposed methods reduce the labeling cost by a significant margin, yet their performances are highly competitive with the state-of-the-art methods. An integral part of activity understanding is to temporally segment the activities from a long video sequence. In this thesis, we also present an approach for activity segmentation that requires limited human supervision.The first approach we propose is based on an ensemble of SVM classifiers. Given the set of unlabeled data, we select the most informative queries to be labeled by a human annotator. Then, we train new SVMs and include them into the ensemble with updated weights. In the second approach, we take the advantage of contextual relationship among the activities and the objects in the video sequence. We encode contextual information using a conditional random field and then, present a novel active leaning algorithm that utilizes both of the entropy and the mutual information of the activity nodes. In the third approach, we further reduce human effort by making early prediction of the activity labels and providing dynamic suggestions to the annotator. State-of-the-art approaches do not scale with the growing number of video categories and they do not consider the cost of long viewing time for video labeling. Our proposed framework uses label propagation and LSTM based recurrent neural network that effectively selects informative queries and provides early suggestions to the annotator. In the fourth approach, we propose a continuous activity learning framework by intricately tying together deep hybrid feature models and active learning. This allows us to automatically select the most suitable features and to take the advantage of incoming unlabeled instances to improve the existing model incrementally. Finally, we propose a method for temporally segmenting meaningful activities that require very limited supervision. Perceiving these activities in a long video sequence is a challenging problem due to ambiguous definition of meaningfulness as well as clutters in the scene. We approach this problem by learning a generative model for regular motion patterns. We propose two methods that are built upon the autoencoders for their ability to work with unlabeled data.

This dissertation focuses on the novel learning strategy based on geometric support vector machines to address the difficulties of processing immense data set. Support vector machines find the hyper-plane that maximizes the margin between two classes, and the decision boundary is represented with a few training samples it becomes a favorable choice for incremental learning. The dissertation presents a novel method Geometric Incremental Support Vector Machines (GISVMs) to address both efficiency and accuracy issues in handling massive data sets. In GISVM, skin of convex hulls is defined and an efficient method is designed to find the best skin approximation given available examples. The set of extreme points are found by recursively searching along the direction defined by a pair of known extreme points. By identifying the skin of the convex hulls, the incremental learning will only employ a much smaller number of samples with comparable or even better accuracy. When additional samples are provided, they will be used together with the skin of the convex hull constructed from previous dataset. This results in a small number of instances used in incremental steps of the training process. Based on the experimental results with synthetic data sets, public benchmark data sets from UCI and endoscopy videos, it is evident that the GISVM achieved satisfactory classifiers that closely model the underlying data distribution. GISVM improves the performance in sensitivity in the incremental steps, significantly reduced the demand for memory space, and demonstrates the ability of recovery from temporary performance degradation.