Browsing by Author "Solin, Arno, Prof., Aalto University, Department of Computer Science, Finland"
Now showing 1 - 3 of 3
- Results Per Page
- Sort Options
- Machine Learning for Healthcare
School of Science | Doctoral dissertation (article-based)(2022) Jaskari, JoelMachine learning has been recently proposed for various medical applications. Especially the deep neural network based approach has been found to achieve state-of-the-art performance in various classification tasks. However, many of these studies use simplified classification systems, for example, the referable/non-referable system in the case of diabetic retinopathy classification. Moreover, the studies that have used clinical classification systems have not considered the uncertainty of the classifiers, which is of paramount interest in the medical field. In addition, extensive analysis of automatic segmentation algorithms that includes comparison to the interobserver variability of multiple radiologists' segmentations has not yet been performed for some challenging tasks, such as the automatic segmentation of the mandibular canals. The machine learning algorithms should also be able to be trained on local hospital data, which can pose issues relating to the amount of available training data. This thesis considers machine learning for various tasks in healthcare using Finnish hospital data. Deep convolutional neural networks (CNNs) are utilized for diabetic retinopathy and macular edema classification based on clinical severity scales. In addition, approximate Bayesian deep learning approaches are systematically studied for uncertainty-aware diabetic retinopathy classification of clinical data. A connection is derived between the referral of uncertain classifications and reject option classification, and it is used to develop a novel uncertainty measure. A CNN approach will also be introduced for the segmentation of the mandibular canal in cone beam computed tomography volumes. The approach is then compared to the interobserver variability of multiple radiologists' segmentations of the canal. Lastly, this thesis will examine multiple machine learning approaches for very low birth weight neonate mortality and morbidity prediction. The results suggest that even a relatively small set of Finnish hospital data can be utilized to train deep learning classifiers for diabetic retinopathy and macular edema classification with clinical classification systems. It also turns out that approximate Bayesian neural networks and the derived novel uncertainty measure can be used to accurately estimate the uncertainty in clinical diabetic retinopathy classification. The deep learning approach is shown to set a new state-of-the-art for the mandibular canal segmentation task and it is also found to localize the canals with lower variability than the interobserver variability of four radiologists. A random forest classifier turned out to outperform other methods in neonatal mortality and morbidity prediction. - Rethinking Inference in Gaussian Processes: A Dual Parameterization Approach
School of Science | Doctoral dissertation (article-based)(2024) Chang, Paul E.Uncertainty quantification is a vital aspect of machine learning, especially when accurate estimates of uncertainty are crucial for making informed decisions. Gaussian Processes (GPs), known for their versatility as function space priors, find wide-ranging applications in diverse fields, including climate modelling and epidemiology. GPs are particularly useful due to their non-linearity, allowing them to adapt to various data patterns and their ability to integrate domain-specific knowledge. As probabilistic models, they offer predictions and quantify the uncertainty within these predictions, an essential feature in scenarios demanding reliable forecasts. This thesis focuses on applying GPs to large-scale, non-Gaussian sequential data. Due to their non-parametric nature, GPs face increasing computational demands as data size grows. The requirement for a conjugate Gaussian likelihood for computational tractability presents further challenges. Therefore, it is common to use approximate inference for applying GPs to non-Gaussian likelihoods alongside scalable model formulations to handle complex data distributions in real- world applications. The theme connecting the papers is an innovative approach to parameterizing the optimization problem in approximate inference, centring on a forgotten parameterization termed the dual parameters. This fresh perspective offers methods to enhance the efficiency of GPs in applying them to large and complex datasets, particularly in the context of sequential data. This approach addresses the pivotal challenges of tractability and scalability inherent in GPs in the sequential setting. The concept of dual parameters serves as a unifying framework, linking all approximate inference techniques through their various likelihood approximations. In addition, the thesis shows the application of dual parameterization methods in a range of GP model formulations and problem settings. It introduces a new algorithm for inference and learning in non-Gaussian time series data and the sparse GP framework. The applications discussed extend to areas such as Bayesian optimization and continual learning, highlighting the adaptability and potential of GPs in contemporary machine learning. - Uncertainty Quantification in Deep Learning
School of Science | Doctoral dissertation (article-based)(2023) Meronen, LassiDeep learning has recently become a popular method for solving problems involving large data sets, and in many applications, human performance has been exceeded. However, deep learning models tend to be overconfident in their predictions, especially when encountering new input samples that differ from anything the model has learned during training. This thesis aims to address this problem by developing uncertainty quantification techniques that allow deep learning models to recognise the limits of their capabilities better and when they should be uncertain in their predictions. Improved uncertainty quantification would enable deep learning models to be used in safety-critical applications that require reliable uncertainty estimates. Uncertainty quantification is improved through a Bayesian perspective, and making connections between neural networks and Gaussian processes is at the core of this research. Gaussian processes are principled Bayesian models that are known to provide reliable uncertainty estimates for their predictions, and the aim is to bring these desirable properties to deep learning models. Another key benefit of Gaussian processes in terms of uncertainty quantification is the possibility of including prior assumptions into the model through a covariance function. The results in this thesis show that similar prior assumptions can be induced into deep learning models through activation functions. This allows neural networks to replicate stationary Gaussian process behaviour with a Matérn covariance. This result fills a gap in research connecting Gaussian processes and neural networks that has existed for over twenty years. Matérn covariance is arguably the most used covariance function in Gaussian processes, making this result impactful. This thesis considers two distinct parts contributing to uncertainty quantification: 1. encoding meaningful priors and 2. approximate inference. The main focus is on meaningful priors, but approximate inference is also focused on, as it is required to use Bayesian deep learning models in practice. Publications in this thesis show theoretical results that progress uncertainty quantification through model design, which allows the encoding of conservative behaviour into the model. In addition, this thesis tackles the problem of increasing size and computational requirements of modern deep learning models. This is also done with uncertainty quantification methods by applying them to dynamic neural networks that attempt to achieve improved performance for a limited computational budget. Computationally efficient uncertainty quantification methods that fit into the computationally restricted regime of dynamic neural networks are introduced. The results show that uncertainty quantification improves decision-making in dynamic neural networks, which leads to better predictive performance. This means high performance is achieved at a lower computational cost, making high-end deep learning models available on hardware with limited computational capacity, such as mobile devices. Improving dynamic neural network performance also helps decrease the energy consumption of large deep learning models.