UTD Theses and Dissertations
Permanent URI for this collectionhttps://hdl.handle.net/10735.1/5608
Browse
Browsing UTD Theses and Dissertations by Subject "Active learning"
Now showing 1 - 3 of 3
- Results Per Page
- Sort Options
Item Cold Start Active Learning With Submodular Mutual Information for Imbalanced Text Classification(December 2023) Iyer, Adithya Sundararajan 1998-; Iyer, Rishabh; Du, Xinya; Gogate, VibhavThis study tackles the cold start problem in active learning for imbalanced binary text classification. Focusing on three datasets (YouTube spam, SMS spam, tweet sentiment) with class imbalances in their training data, we investigate the efficacy of Submodular Mutual Information (SMI) methods in the initial active learning stage. These methods aim to balance class representation using a query set around one percent of the unlabeled data size. We compare four SMI approaches (two facility location variants, log determinant, graph cut) with a custom regular expression matching baseline and five established baseline sampling strategies (Random, BADGE, Entropy, Least confidence, and Margin Sampling) across all datasets. Our experiments, conducted ten times per dataset reveal that SMI methods, on average, especially log determinant, outperform both regex matching and traditional baselines. Further analysis has also been done on the effect of the number of query samples used on performance. The work highlights the potential of SMI in efficiently addressing the cold start challenge in imbalanced text classification contexts.Item Domain Adaptation for Speech Based Emotion Recognition(2019-05) Abdelwahab, Mohammed; Busso-Recabarren, Carlos A.One of the main barriers in the deployment of speech emotion recognition systems in real applications is the lack of generalization of the emotion classifiers. The recognition performance achieved in controlled recordings drops when the models are tested with different speakers, channels, environments and domain conditions. Annotations of new data in the new domain is expensive and time-consuming. Therefore, it is important to design strategies that efficiently use limited amounts of labeled data in the new domain and extract as much useful information as possible from the available unlabeled data to improve the robustness of the system. This thesis studies approaches to generalize emotion classifiers to new domains. First, we explore supervised model adaptation, which modifies the trained model using labeled data from the new domain. We study the data requirements and different approaches for SVM adaptation in the context of supervised adaptation for speech based emotion recognition. The results indicate that even small portion of data used for adaptation can significantly improve the performance. Increasing the speaker diversity in the labeled data used for adaptation does not provide significant gain in performance. Also, we observe that classifiers trained with naturalistic or acted data achieve similar performance after adapting the models to the target domain. Second, we propose solutions for semi-supervised domain adaptation. We explore the use of active learning (AL) in speech emotion recognition. Active learning selects samples in the new domain that are used to adapt the classification models using domain adaptation. We consider two approaches. The first approach focuses on selecting samples that are more beneficial to the classifier. We propose a novel iterative fast converging incremental adaptation algorithm that only uses correctly classified samples at each iteration. This conservative framework creates sequences of smooth changes in the decision hyperplane, resulting in statistically significant improvements over conventional schemes that adapt the models at once using all the available data. The second approach focuses on selecting the features that optimize the performance in the new domain. The method combines AL along with feature selection to build a diverse ensemble that performs well in the new domain. The use of ensembles is an attractive solution, since they can be built to perform well across different mismatches. We study various data selection criteria, and different sample sizes to determine the best approach toward building a robust and diverse ensemble of classifiers. The results demonstrated that we can achieve a significant improvement by performing feature selection on a small set from the target domain. Finally, we explore unsupervised adaptation for speech emotion recognition. We propose to use adversarial multitask training to extract a common representation between training and testing domains. The primary task is to predict emotional attribute-based descriptors for arousal, valence or dominance. The secondary task is to learn a common representation where the train and test domains cannot be distinguished. By using a gradient reversal layer, the gradients coming from the domain classifier are used to bring the source and target domain representations closer. We show that exploiting unlabeled data consistently leads to better emotion recognition performance across all emotional dimensions. We visualize the effect of adversarial training on the feature representation across the proposed deep learning architecture. The analysis shows that the data representations for the train and test domains converge as the data is passed to deeper layers of the network. The proposed advances create appealing strategies to build robust speech emotion classifiers that generalize across domains, providing practical affective-aware solutions to real-life problems.Item Sample Efficient Cost-Aware Active Learning(2020-10-01) Das, Srijita; Natarajan, SriraamThe grand vision of Artificial Intelligence is to build agents that can continuously learn from experience and reason about how to act in the world. Traditional machine learning models generally assume access to huge amount of data. This allows effective learning from huge historical data. However, certain domains are inherently sparse, for example, healthcare. It is essential to build machine learning models that can learn and reason effectively in such sparse domains. This dissertation aims to address the problem of efficient learning with human in the loop in sparse, noisy and structured domains. To this effect, the dissertation has two different yet related directions. The first direction focuses on identifying the most informative training instances and features for learning in data-scarce domains. For example, in a medical domain, certain features like MRI, lab tests, genetic sequencing etc., are expensive. In such cases, it becomes crucial to identify the appropriate set of instances (clinical subjects) for whom these expensive features need to be solicited. It is also challenging to identify the appropriate set of features for different subjects since collecting all the expensive features might not be feasible due to budgetary restrictions (both time and cost budgets). We aim to develop strong predictive models at a reasonable cost. The second direction focuses on learning to act in noisy, structured domains. Deciding how to act in complex domains like healthcare should take into account the rich relational structure that exists between various interrelated entities present in the domain. For example, a patient’s current medical condition depends on the medical history of his/her family members. By capturing the existing structure in such domains, an agent can learn how to act and generalize well to unseen conditions. In this dissertation, to handle the challenge of data-scarcity, a unified active learning framework is developed which can identify the most informative samples for whom the expensive feature subset needs to be elicited. Further, the challenge of feature elicitation is addressed by identifying important feature subsets for different clusters of similar instances. Feature acquisition cost is taken into account in an optimization framework to handle the trade-off between acquisition cost and model performance. To address the second direction, an efficient symbolic reinforcement learning algorithm is developed to learn utility functions in complex structured domain. This approach is capable of capturing the various relations that exists among entities, thus resulting in learning effective and generalizable policies. Addressing these challenges in predictive modelling and decision making can help in building smart and resource efficient AI agents that can reason well in data scarce and structured domains.