Browsing by Author "Heller, Katherine"

High-dimensional probability distributions are important objects in a wide variety of applications.

Generative models provide an excellent manipulation method for training from rich available unlabeled data set and sampling new data points from underlying high-dimensional probability distributions.

The recent proposed Variational auto-encoders (VAE) framework is an efficient high-dimensional inference method to modeling complicated data manifold in an approximate Bayesian way, i.e., variational inference.

We first discuss how to design fast stochastic backpropagation algorithm for the VAE based amortized variational inference method.

Particularly, we propose second order Hessian-free optimization method for Gaussian latent variable models and provide a theoretical justification to the convergence of Monte Carlo estimation in our algorithm.

Then, we apply the amortized variational inference to a dynamic modeling application in flu diffusion task.

Compared with traditional approximate Gibbs sampling algorithm, we make less assumption to the infection rate.

Differing from the maximum likelihood approach of VAE, Generative Adversarial Networks (GAN) is trying to solve the generation problem from a game theoretical way.

From this viewpoint, we design a framework VAE+GAN, by placing a discriminator on top of auto-encoders based model and introducing an extra adversarial loss.

The adversarial training induced by the classification loss is to make the discriminator believe the sample from the generative model is as real as the one from the true dataset.

This trick can practically improve the quality of generation samples, demonstrated on images and text domains with elaborately designed architectures.

Additionally, we validate the importance of generative adversarial loss with the conditional generative model in two biological applications: approximate Turing pattern PDEs generation in synthetic/system biology, and automatic cardiovascular disease detection in medical imaging processing.

Clinical prediction models offer the ability to help physicians make better data-driven decisions that can improve patient outcomes. Given the wealth of data available with the widespread adoption of electronic health records, more flexible statistical models are required that can account for the messiness and complexity of this data. In this dissertation we focus on developing models for clinical time series, as most data within healthcare is collected longitudinally and it is important to take this structure into account. Models built off of Gaussian processes are natural in this setting of irregularly sampled, noisy time series with many missing values. In addition, they have the added benefit of accounting for and quantifying uncertainty, which can be extremely useful in medical decision making. In this dissertation, we develop new Gaussian process-based models for medical time series along with associated algorithms for efficient inference on large-scale electronic health records data. We apply these models to several real healthcare applications, using local data obtained from the Duke University healthcare system.

In Chapter 1 we give a brief overview of clinical prediction models, electronic health records, and Gaussian processes. In Chapter 2, we develop several Gaussian process models for clinical time series in the context of chronic kidney disease management. We show how our proposed joint model for longitudinal and time-to-event data and model for multivariate time series can make accurate predictions about a patient's future disease trajectory. In Chapter 3, we combine multi-output Gaussian processes with a downstream black-box deep recurrent neural network model from deep learning. We apply this modeling framework to clinical time series to improve early detection of sepsis among patients in the hospital, and show that the Gaussian process preprocessing layer both allows for uncertainty quantification and acts as a form of data augmentation to reduce overfitting. In Chapter 4, we again use multi-output Gaussian processes as a preprocessing layer in model-free deep reinforcement learning. Here the goal is to learn optimal treatments for sepsis given clinical time series and historical treatment decisions taken by clinicians, and we show that the Gaussian process preprocessing layer and use of a recurrent architecture offers improvements over standard deep reinforcement learning methods. We conclude in Chapter 5 with a summary of future areas for work, and a discussion on practical considerations and challenges involved in deploying machine learning models into actual clinical practice.

Objective

Materials and methods

Results

Conclusions

Deep learning models, which form the backbone of modern ML systems, generalize poorly to small changes to the data distribution. They are also bad at signalling failure, making predictions with high confidence when their training data or fragile assumptions make them unlikely to make reasonable decisions. This lack of robustness makes it difficult to trust their use in safety-critical settings. Accordingly, there is a pressing need to equip models with a notion of uncertainty to understand their failure modes and detect when their decisions cannot be used or require intervention. Uncertainty quantification is thus crucial for ML systems to work consistently on real-world data and fail loudly when they don’t.One growing line of research on uncertainty quantification is probabilistic modelling which is concerned with capturing model uncertainty by placing a distribution over the models which can be marginalized at test-time. This is especially useful in underspecified models which can have diverse near-optimal solutions, at training time, with similar population-level performance. However, probabilistic modelling approaches such as Bayesian neural networks (BNN) do not scale well in terms of memory and runtime and often underperform simple deterministic baselines in terms of accuracy. Furthermore, BNNs underperform deep ensembles as they fail to explore multiple modes, in the loss space, while being effective at capturing uncertainty within a single mode.

In this thesis, we develop multimodal representations of model uncertainty that can capture a diverse set of hypotheses. We first propose a scalable family of BNN priors (and corresponding approximate posteriors) that combine the local (i.e. within-mode) uncertainty with mode averaging to deliver robust and calibrated uncertainty estimates in addition to improving accuracy both in and out of distribution. We then leverage a multimodal representation of uncertainty to modulate the amount of information transfer between tasks in meta-learning. Our proposed framework integrates Bayesian non-parametric mixtures with deep learning to enable NNs to adapt their capacity as more data is observed which is crucial for lifelong learning. Finally, we propose to replace the reverse Kullback-Leibler divergence (RKL), known for its mode-seeking behavior and for underestimating posterior covariance, with the forward KL (FKL) divergence in a theoretically-guided novel inference procedure. This ensures the efficient combination of variational boosting with adaptive importance sampling. The proposed algorithm offers a well-defined compute-accuracy trade-off and is guaranteed to converge to the optimal multimodal variational solution as well as the optimal importance sampling proposal distribution.

We present a scalable end-to-end classifier that uses streaming physiological and medication data to accurately predict the onset of sepsis, a life-threatening complication from infections that has high mortality and morbidity. Our proposed framework models the multivariate trajectories of continuous-valued physiological time series using multitask Gaussian processes, seamlessly accounting for the high uncertainty, frequent missingness, and irregular sampling rates typically associated with real clinical data. The Gaussian process is directly connected to a black-box classifier that predicts whether a patient encounter will become septic, chosen in our case to be a recurrent neural network to account for the extreme variability in the length of patient encounters. We show how several approximations scale the computations associated with the Gaussian process in a manner so that the entire system can be trained discriminatively end-to-end using backpropagation. In a large cohort of heterogeneous inpatient encounters at our university health system we find that it outperforms several baselines at predicting sepsis, and yields 33\% and 195\% improved areas under the Receiver Operating Characteristic and Precision Recall curves as compared to the NEWS score currently in use on our own hospital wards.