Sentence labeling is a type of pattern recognition task that involves the assignment of a categorical label to each member of a sentence of observed words. Standard supervised sentence-labeling systems often have poor generalization: it is difficult to estimate parameters for words which appear in the test set, but seldom (or never) appear in the training set, because they only use words as features in their prediction tasks. Representation learning is a promising technique for discovering features that allow a supervised classifier to generalize from a source domain dataset to arbitrary new domains. We demonstrate that features which are learned from distributional representations of unlabeled data can be used to improve performance on out-of-vocabulary words and help the model to generalize. We also argue that it is important for a representation learner to be able to incorporate expert knowledge during its search for helpful features. We investigate techniques for building open-domain sentence labeling systems that approach the ideal of a system whose accuracy is high and consistent across domains. In particular, we investigate unsupervised techniques for language model representation learning that provide new features which are stable across domains, in that they are predictive in both the training and out-of-domain test data. In experiments, our best system with the proposed techniques reduce error by as much as 11.4% relative to the previous system using traditional representations on the Part-of-Speech tagging task. Moreover, we leverage the Posterior Regularization framework, and develop an architecture for incorporating biases from prior knowledge into representation learning. We investigate three types of biases: entropy bias, distance bias and predictive bias. Experiments on two domain adaptation tasks show that our biased learners identify significantly better sets of features than unbiased learners. This results in a relative reduction in error of more than 16% for both tasks with respect to existing state-of-the-art representation learning techniques. We also extend the idea of using additional unlabeled data to improve the system's performance on a different NLP task, word alignment. Traditional word alignment only takes a sentence-level aligned parallel corpus as input and generates the word-level alignments. However, as the integration of different cultures, more and more people are competent in multiple languages, and they often use elements of multiple languages in conversations. Linguist Code Switching (LCS) is such a situation where two or more languages show up in the context of a single conversation. Traditional machine translation (MT) systems treat LCS data as noise, or just as regular sentences. However, if LCS data is processed intelligently, it can provide a useful signal for training word alignment and MT models. In this work, we first extract constraints from this code switching data and then incorporate them into a word alignment model training procedure. We also show that by using the code switching data, we can jointly train a word alignment model and a language model using co-training. Our techniques for incorporating LCS data improve by 2.64 in BLEU score over a baseline MT system trained using only standard sentence-aligned corpora. / Computer and Information Science

Active learning (AL) aims at reducing labeling effort by identifying the most valuable unlabeled data points from a large pool. Traditional AL frameworks have two limitations: First, they perform data selection in a multi-round manner, which is time-consuming and impractical. Second, they usually assume that there are a small amount of labeled data points available in the same domain as the data in the unlabeled pool. In this thesis, we initiate the study of one-round active learning to solve the first issue. We propose DULO, a general framework for one-round setting based on the notion of data utility functions, which map a set of data points to some performance measure of the model trained on the set. We formulate the one-round active learning problem as data utility function maximization. We then propose D²ULO on the basis of DULO as a solution that solves both issues. Specifically, D²ULO leverages the idea of domain adaptation (DA) to train a data utility model on source labeled data. The trained utility model can then be used to select high-utility data in the target domain and at the same time, provide an estimate for the utility of the selected data. Our experiments show that the proposed frameworks achieves better performance compared with state-of-the-art baselines in the same setting. Particularly, D²ULO is applicable to the scenario where the source and target labels have mismatches, which is not supported by the existing works. / M.S. / Machine Learning (ML) has achieved huge success in recent years. Machine Learning technologies such as recommendation system, speech recognition and image recognition play an important role on human daily life. This success mainly build upon the use of large amount of labeled data: Compared with traditional programming, a ML algorithm does not rely on explicit instructions from human; instead, it takes the data along with the label as input, and aims to learn a function that can correctly map data to the label space by itself. However, data labeling requires human effort and could be time-consuming and expensive especially for datasets that contain domain-specific knowledge (e.g., disease prediction etc.) Active Learning (AL) is one of the solution to reduce data labeling effort. Specifically, the learning algorithm actively selects data points that provide more information for the model, hence a better model can be achieved with less labeled data. While traditional AL strategies do achieve good performance, it requires a small amount of labeled data as initialization and performs data selection in multi-round, which pose great challenge to its application, as there is no platform provide timely online interaction with data labeler and the interaction is often time inefficient. To deal with the limitations, we first propose DULO which a new setting of AL is studied: data selection is only allowed to be performed once. To further broaden the application of our method, we propose D²ULO which is built upon DULO and Domain Adaptation techniques to avoid the use of initial labeled data. Our experiments show that both of the proposed two frameworks achieve better performance compared with state-of-the-art baselines.

Machine learning models are becoming popular in the industry since the technology has developed to solve numerous problems, such as classification [1], detection [2], and segmentation [3]. These algorithms require training with a large dataset which includes correct class labels to perform well on unseen data. One way to get access to large sets of annotated data is to use data from simulation engines. However this data is often not as complex and rich as real data, and for images, for examples, there can be a need to make these look more photorealistic. One approach to do this is denoted Domain adaptation. In collaboration with SAAB Aeronautics, which funds this research, this study aims to explore available domain adaptation frameworks, implement a framework and use it to make a transformation from simulation to real- life. A state-of-the-art framework CyCADA was re-implemented from scratch using Python and TensorFlow as a Deep Learning package. The CyCADA implementation was successfully verified by reproducing the digit adaptation result demonstrated in the original paper, making domain adaptations between MNIST, USPS, and SVHN. CyCADA was used to domain adapt landscape images from simulation to real-life. Domain-adapted images were used to train an object detector to evaluate whether CyCADA allows a detector to perform more accurately in real-life data. Statistical measurements, unfortunately, showed that domain-adapted images became less photorealistic with CyCADA, 88.68 FID on domain-adapted images compared to 80.43 FID on simulations, and object detection performed better on real-life data without CyCADA, 0.131 mAP with a detector trained on domain-adapted images compared to 0.681 mAP with simulations. Since CyCADA produced effective domain adaptation results between digits, there remains a possibility to try multiple hyperparameter settings and neural network architecture to produce effective results with landscape images. / Denna studie genomfördes i ett samarbete med SAAB Aeronautics och handlar om att utveckla en Domain Adaptation-modul som förbättrar prestandan av ett nätverk för objektdetektering. När ett objektdetekteringsnätverk är tränat med data från en domän så är det inte givet att samma nätverk presterar bra på en annan domän. Till exempel, ritningar och fotografier av frukter. Forskare löser problemet genom att samla data från varje domän och träna flera maskininlärningsalgoritmer, vilket är en lösning som kräver tid och energi. Detta problem kallas för domänskiftesproblem. Ett hett ämne inom djupinlärning handlar om att lösa just detta problem med domänskift och det finns en rad algoritmer som faller i kategorin Domain Adaptation. Denna studie utvecklar CyCADA som metod att evaluera en toppmodern Domain Adaptation-algoritm. Återimplementering av CyCADA blev lyckad, eftersom flera resultat var återskapade från den originala artikeln. CyCADA producerade effektiva domänskiften på bilder av siffror. CyCADA användes med landskapsbilder från en simulator för att öka verklighetsförankringen på bilderna. Domänskiftade landskapsbilder blev suddiga med CyCADA. FID värdet av domänskiftade bilder, ett utvärderingsmått som evaluerar fotorealism av bilder, blev lägre i jämförelse med endast simulerade bilder. Objektdetekteringsnätverket presterade bättre utan användning av CyCADA. Givet att CyCADA presterade bra i att transformera bilder av siffror från en domän till en annan finns det hopp om att ramverket kan prestera bra med landskapsbilder med fler försök i att ställa in hyperparameterar.

Framgången inom maskininlärning och djupinlärning beror till stor del på stora, annoterade dataset. Att tilldela etiketter till data är väldigt resurskrävande och kan till viss del undvikas genom att utnyttja datans statistiska egenskaper. En maskininlärningsmodell kan lära sig att klassificera bilder från en domän utifrån träningsexempel som innehåller bilder, samt etiketter som berättar vad bilder föreställer. Men vad gör man om datan inte har tilldelade etiketter? En maskininlärningsmodell som lär sig en uppgift utifrån annoterad data från en källdomän, kan med hjälp av information från måldomänen (som inte har tilldelade etiketter), anpassas till att prestera bättre på data från måldomänen. Forskningsområdet som studerar hur man anpassar och generaliserar en modell mellan två olika domäner heter domänanpassning, eller domain adaptation, på engelska.   Detta examensarbete är utfört på Scanias forskningsavdelning för autonom transport och handlar om hur modeller för bildklassificering som tränas på kamerabilder med etiketter, kan anpassas till att få ökad noggrannhet på ett dataset med LiDAR bilder, som inte har etiketter. Två metoder för domänanpassning har jämförts med varandra, samt en model tränad på kameradata genom övervakad inlärning utan domänanpassning. Alla metoder opererar på något vis med ett djupt faltningsnätverk (CNN) där uppgiften är att klassificera bilder utav bilar eller fotgängare. Kovariansen utav datan från käll- och måldomänen är det centrala måttet för domänanpassningsmetoderna i detta projekt. Den första metoden är en så kallad ytlig metod, där själva anpassningsmetoden inte ingår inuti den djupa arkitekturen av modellen, utan är ett mellansteg i processen. Den andra metoden förenar domänanpassningsmetoden med klassificeringen i den djupa arkitekturen. Den tredje modellen består endast utav faltningsnätverket, utan en metod för domänanpassning och används som referens.    Modellen som tränades på kamerabilderna utan en domänanpassningsmetod klassificerar LiDAR-bilderna med en noggrannhet på 63.80%, samtidigt som den ”ytliga” metoden når en noggrannhet på 74.67% och den djupa metoden presterar bäst med 80.73%. Resultaten visar att det är möjligt att anpassa en modell som tränas på data från källdomänen, till att få ökad klassificeringsnoggrannhet i måldomänen genom att använda kovariansen utav datan från de två domänerna. Den djupa metoden för domänanpassning tillåter även användandet utav andra statistiska mått som kan vara mer framgångsrika i att generalisera modellen, beroende på hur datan är fördelad. Överlägsenheten hos den djupa metoden antyder att domänanpassning med fördel kan bäddas in i den djupa arkitekturen så att modelparametrarna blir uppdaterade för att lära sig en mer robust representation utav måldomänen.

Master of Science / Department of Computing and Information Sciences / Doina Caragea / Several computational biology and bioinformatics problems involve DNA sequence classification using supervised machine learning algorithms. The performance of these algorithms is largely dependent on the availability of labeled data and the approach used to represent DNA sequences as {\it feature vectors}. For many organisms, the labeled DNA data is scarce, while the unlabeled data is easily available. However, for a small number of well-studied model organisms, large amounts of labeled data are available. This calls for {\it domain adaptation} approaches, which can transfer knowledge from a {\it source} domain, for which labeled data is available, to a {\it target} domain, for which large amounts of unlabeled data are available. Intuitively, one approach to domain adaptation can be obtained by extracting and representing the features that the source domain and the target domain sequences share. \emph{Latent Dirichlet Allocation} (LDA) is an unsupervised dimensionality reduction technique that has been successfully used to generate features for sequence data such as text. In this work, we explore the use of LDA for generating predictive DNA sequence features, that can be used in both supervised and domain adaptation frameworks. More precisely, we propose two dimensionality reduction approaches, LDA Words (LDAW) and LDA Distribution (LDAD) for DNA sequences. LDA is a probabilistic model, which is generative in nature, and is used to model collections of discrete data such as document collections. For our problem, a sequence is considered to be a ``document" and k-mers obtained from a sequence are ``document words". We use LDA to model our sequence collection. Given the LDA model, each document can be represented as a distribution over topics (where a topic can be seen as a distribution over k-mers). In the LDAW method, we use the top k-mers in each topic as our features (i.e., k-mers with the highest probability); while in the LDAD method, we use the topic distribution to represent a document as a feature vector. We study LDA-based dimensionality reduction approaches for both supervised DNA sequence classification, as well as domain adaptation approaches. We apply the proposed approaches on the splice site predication problem, which is an important DNA sequence classification problem in the context of genome annotation. In the supervised learning framework, we study the effectiveness of LDAW and LDAD methods by comparing them with a traditional dimensionality reduction technique based on the information gain criterion. In the domain adaptation framework, we study the effect of increasing the evolutionary distances between the source and target organisms, and the effect of using different weights when combining labeled data from the source domain and with labeled data from the target domain. Experimental results show that LDA-based features can be successfully used to perform dimensionality reduction and domain adaptation for DNA sequence classification problems.

The use of machine learning to predict output from data, using a model, is a well studied area. There are, however, a number of real-world applications that require a model to be produced but have little or no data available of the specific environment. These situations are prominent in Intelligent Environments (IEs). The sparsity of the data can be a result of the physical nature of the implementation, such as sensors placed into disaster recovery scenarios, or where the focus of the data acquisition is on very defined user groups, in the case of disabled individuals. Standard machine learning approaches focus on a need for training data to come from the same domain. The restrictions of the physical nature of these environments can severely reduce data acquisition making it extremely costly, or in certain situations, impossible. This impedes the ability of these approaches to model the environments. It is this problem, in the area of IEs, that this thesis is focussed. To address complex and uncertain environments, humans have learnt to use previously acquired information to reason and understand their surroundings. Knowledge from different but related domains can be used to aid the ability to learn. For example, the ability to ride a road bicycle can help when acquiring the more sophisticated skills of mountain biking. This humanistic approach to learning can be used to tackle real-world problems where a-priori labelled training data is either difficult or not possible to gain. The transferral of knowledge from a related, but differing context can allow for the reuse and repurpose of known information. In this thesis, a novel composition of methods are brought together that are broadly based on a humanist approach to learning. Two concepts, Transfer Learning (TL) and Fuzzy Logic (FL) are combined in a framework, Fuzzy Transfer Learning (FuzzyTL), to address the problem of learning tasks that have no prior direct contextual knowledge. Through the use of a FL based learning method, uncertainty that is evident in dynamic environments is represented. By combining labelled data from a contextually related source task, and little or no unlabelled data from a target task, the framework is shown to be able to accomplish predictive tasks using models learned from contextually different data. The framework incorporates an additional novel five stage online adaptation process. By adapting the underlying fuzzy structure through the use of previous labelled knowledge and new unlabelled information, an increase in predictive performance is shown. The framework outlined is applied to two differing real-world IEs to demonstrate its ability to predict in uncertain and dynamic environments. Through a series of experiments, it is shown that the framework is capable of predicting output using differing contextual data.

While parsing performance on in-domain text has developed steadily in recent years, out-of-domain text and grammatically noisy text remain an obstacle and often lead to significant decreases in parsing accuracy. In this thesis, we focus on the parsing of noisy content, such as user-generated content in services like Twitter. We investigate the question whether a preprocessing step based on machine translation techniques and unsupervised models for text-normalization can improve parsing performance on noisy data. Existing data sets are evaluated and a new data set for dependency parsing of grammatically noisy Twitter data is introduced. We show that text-normalization together with a combination of domain-specific and generic part-of-speech taggers can lead to a significant improvement in parsing accuracy. Powered by TCPDF (www.tcpdf.org)

Identifying semantic relations in natural language text is an important component of many knowledge extraction systems. This thesis studies the task of hypernym discovery, i.e discovering terms that are related by the hypernymy (is-a) relation. Specifically, this thesis explores how state-of-the-art methods for hypernym discovery perform when applied in specific language domains. In recent times, state-of-the-art methods for hypernym discovery are mostly made up by supervised machine learning models that leverage distributional word representations such as word embeddings. These models require labeled training data in the form of term pairs that are known to be related by hypernymy. Such labeled training data is often not available when working with a specific language domain. This thesis presents experiments with an automatic training data collection algorithm. The algorithm leverages a pre-defined domain-specific vocabulary, and the lexical resource WordNet, to extract training pairs automatically. This thesis contributes by presenting experimental results when attempting to leverage such automatically collected domain-specific training data for the purpose of domain adaptation. Experiments are conducted in two different domains: One domain where there is a large amount of text data, and another domain where there is a much smaller amount of text data. Results show that the automatically collected training data has a positive impact on performance in both domains. The performance boost is most significant in the domain with a large amount of text data, with mean average precision increasing by up to 8 points.

In order to enable more widespread application of robots, we are required to reduce the human effort for the introduction of existing robotic platforms to new environments and tasks. In this thesis, we identify three complementary strategies to address this challenge, via the use of imitation learning, domain adaptation, and transfer learning based on simulations. The overall work strives to reduce the effort of generating training data by employing inexpensively obtainable labels and by transferring information between different domains with deviating underlying properties. Imitation learning enables a straightforward way for untrained personnel to teach robots to perform tasks by providing demonstrations, which represent a comparably inexpensive source of supervision. We develop a scalable approach to identify the preferences underlying demonstration data via the framework of inverse reinforcement learning. The method enables integration of the extracted preferences as cost maps into existing motion planning systems. We further incorporate prior domain knowledge and demonstrate that the approach outperforms the baselines including manually crafted cost functions. In addition to employing low-cost labels from demonstration, we investigate the adaptation of models to domains without available supervisory information. Specifically, the challenge of appearance changes in outdoor robotics such as illumination and weather shifts is addressed using an adversarial domain adaptation approach. A principal advantage of the method over prior work is the straightforwardness of adapting arbitrary, state-of-the-art neural network architectures. Finally, we demonstrate performance benefits of the method for semantic segmentation of drivable terrain. Our last contribution focuses on simulation to real world transfer learning, where the characteristic differences are not only regarding the visual appearance but the underlying system dynamics. Our work aims at parallel training in both systems and mutual guidance via auxiliary alignment rewards to accelerate training for real world systems. The approach is shown to outperform various baselines as well as a unilateral alignment variant.

In this thesis, we propose a new framework for the generation of training data for machine learning techniques used for classification in communications applications. Machine learning-based signal classifiers do not generalize well when training data does not describe the underlying probability distribution of real signals. The simplest way to accomplish statistical similarity between training and testing data is to synthesize training data passed through a permutation of plausible forms of noise. To accomplish this, a framework is proposed that implements arbitrary channel conditions and baseband signals. A dataset generated using the framework is considered, and is shown to be appropriately sized by having $11\%$ lower entropy than state-of-the-art datasets. Furthermore, unsupervised domain adaptation can allow for powerful generalized training via deep feature transforms on unlabeled evaluation-time signals. A novel Deep Reconstruction-Classification Network (DRCN) application is introduced, which attempts to maintain near-peak signal classification accuracy despite dataset bias, or perturbations on testing data unforeseen in training. Together, feature transforms and diverse training data generated from the proposed framework, teaching a range of plausible noise, can train a deep neural net to classify signals well in many real-world scenarios despite unforeseen perturbations.