Computing and Information Systems - Theses

Permanent URI for this collection

http://hdl.handle.net/11343/351

Filters

2020 - 2022 2
1 1
Reset filters

Settings
Statistics
Citations
Type: PhD thesis × Subject: deep learning × Start date: 2020 × End date: 2022 ×

Search Results

Now showing 1 - 2 of 2
  • Item
    Thumbnail Image
    Towards Robust Medical Machine Learning
    He, Jiabo ( 2022)
    Machine learning systems have been developed to address a number of problems in numerous domains, among which medical solutions have been facilitated by machine learning approaches for decades. These approaches play important roles in automated disease diagnosis, medical image processing, and auxiliary surgical operation, etc. Despite the highly efficient diagnosis benefited from machine learning approaches, these methods may not be robust to common challenges in practical scenarios, such as special while crucial characteristics of medical data, annotation variations from multiple experts, noisy annotations, and multi-source datasets. Such problems impede machine learning methods from being applied accurately and safely to medical tasks. In the thesis, we introduce special while important medical problems that were not brought into the spotlight before. We then provide corresponding robust machine learning solutions for each problem when existing machine learning methods degrade significantly in these tasks. Specifically, the first problem is the similarity analysis for time series with large discontinuities, which is common in surgical time series. We thus propose a robust distance measurement for time series with large discontinuities when they disable the accurate measurement of local characteristics using existing algorithms. Second, surgical policies provided by different surgeons for the same patient/surgery may not be exactly the same. We then propose the reward-penalty Dice loss (RPDL) to learn non-unique surgical segmentation regions for deep vision networks. RPDL is robust to varying annotations for the same input, which enables the comprehensive learning of models from multiple experts. Third, medical datasets might be composed of limited examples and noisy annotations, making it challenging to train deep learning models. To address this challenge, we propose alpha-IoU, a family of power Intersection over Union (IoU) losses for bounding box (bbox) regression. We show that alpha-IoU losses are more robust to small datasets and noisy bboxes in lesion detection. Fourth, large-scale medical datasets are often collected from different institutions cooperated by a number of experts. In this case, we build a one-stage framework SpineOne for detecting degenerative discs and vertebrae from spinal MRIs, which implements both the keypoint localization and classification tasks simultaneously. SpineOne is a robust detector to multi-source MRI slices with various scales, numbers and quality. All four proposed machine learning approaches outperform existing baselines by a noticeable margin in specific medical tasks. In summary, four medical issues are thoroughly investigated in the thesis, i.e., the distance measurement for surgical time series with large discontinuities, the surgical region segmentation with a variety of clinician annotations, the lesion detection with limited examples and noisy bboxes, and the anatomical keypoint detection with multi-source medical data. Towards more robust medical machine learning, we then propose one robust machine learning approach for each corresponding problem.
  • Item
    Thumbnail Image
    A multi-faceted approach to document quality assessment
    Shen, Aili ( 2020)
    Document quality assessment, due to its complexity and subjectivity, requires considering information from multiple sources and aspects, to capture quality indicators. Grammaticality, readability, stylistics, structure, correctness, and expertise depth reflect the quality of documents from different aspects, with varying importance across different domains. Automatic quality assessment has obvious benefits in terms of time saving and tractability in contexts where the volume of documents is large. In the case of dynamic documents (possibly with multiple authors), such as in the case of Wikipedia, it is particularly pertinent, as any edit potentially has implications for the quality label of that document. In this thesis, we focusing on improving the performance of document quality assessment systems and measure the uncertainty of document quality assessment systems. This thesis addresses four research questions: (1) How can we capture visual features not present in the document text, such as images and visual layout, to enhance representations learned from text content? (2) How can we make use of hand-crafted features widely adopted in traditional machine learning approaches in the context of neural networks, to generate a more accurate document quality assessment system? (3) How can we model the inherent subjectivity of quality assessment in evaluating the performance of quality assessment systems? and (4) Can a quality assessment system detect whether there are intruder sentences in documents and identify the span of any such intruder sentences, given that they interrupt the coherence of documents, thereby lowering their quality? To address the first research question, we propose to use Inception V3 (Szegedy et al., 2016), a widely used visual model in computer vision, to capture visual features from visual renderings of documents, based on the observation that visual renderings of documents can capture these visual features. Inception V3 compares favourably to textual-based models over the Wikipedia and academic paper reviewing datasets. We further propose a joint model to predict document quality by combining visual and textual features. We observe further improvements over both Wikipedia and academic paper reviewing datasets, indicating complementary between visual and textual features, and the general applicability of our proposed method. Next, we propose two methods to enhance the capacity of neural models in predicting the quality of documents by utilising hand-crafted features. In the first method, we propose to concatenate hand-crafted features with neural learned high-level representations, assuming that neural model-learned features may not have captured all the information carried by these hand-crafted features. The second method, on the other hand, utilises hand-crafted features to guide neural model learning by explicitly attending to feature indicators when learning the relationship between the input and target variables, rather than simply concatenating hand-crafted features. Experimental results demonstrate the superiority of our proposed methods over baselines. To imitate people’s disagreement over the inherently subjective task of document quality assessment, we propose to measure the uncertainty in document quality predictions. We investigate two methods: Gaussian processes (GPs) (Rasmussen and Williams, 2006) and random forests (RFs) (Breiman, 2001), which provide not only a prediction of the document quality but also the uncertainty over their predictions. We also propose an asymmetric cost, considering the prediction uncertainty, which is used to measure the performance of two methods in the scenario, where decision-making processes based on model predictions can lead to different costs. Lastly, we propose a new task of detecting whether there is an intruder sentence in a document, generated by replacing an original sentence with a similar sentence from a second document. Existing datasets in coherence detection are not suitable for our task as they are either too small for training current data-hungry models on or do not specify the span of incoherent text. To benchmark model performance over this task, we construct a large-scale dataset consisting of documents from English Wikipedia and CNN news articles. Experimental results show that pre-trained language models which incorporate larger document contexts in pretraining perform remarkably well in-domain, but experience a substantial drop cross-domain. In follow-up analysis based on human annotations, substantial divergences from human intuitions were observed, pointing to limitations in their ability to model document coherence. Further results over a linguistic probe dataset show that pre-trained models fail to identify some linguistic characteristics that affect document coherence, suggesting room to improve for them to truly capture document coherence, and motivating the construction of a dataset with intruder text at the intra-sentential level.