Many imaging technologies rely on tomographic reconstruction, which requires solving a multidimensional inverse problem given a finite number of projections. Backprojection is a popular class of algorithm for tomographic reconstruction, however it typically results in poor image reconstructions when the projection angles are sparse and/or if the sensors characteristics are not uniform. Several deep learning based algorithms have been developed to solve this inverse problem and reconstruct the image using a limited number of projections. However these algorithms typically require examples of the ground-truth (i.e. examples of reconstructed images) to yield good performance. In this paper, we introduce an unsupervised sparse-view backprojection algorithm, which does not require ground-truth. The algorithm consists of two modules in a generator-projector framework; a convolutional neural network and a spatial transformer network. We evaluated our algorithm using computed tomography (CT) images of the human chest. We show that our algorithm significantly out-performs filtered backprojection when the projection angles are very sparse, as well as when the sensor characteristics vary for different angles. Our approach has practical applications for medical imaging and other imaging modalities (e.g. radar) where sparse and/or non-uniform projections may be acquired due to time or sampling constraints.

To predict an epileptic event means the ability to determine in advance the time of the seizure with the highest possible accuracy. A correct prediction benchmark for epilepsy events in clinical applications is a typical problem in biomedical signal processing that helps to an appropriate diagnosis and treatment of this disease. In this work, we use Pearson's product-moment correlation coefficient from generalized Gaussian distribution parameters coupled with a linear-based classifier to predict between seizure and non-seizure events in epileptic EEG signals. The performance in 36 epileptic events from 9 patients showing good performance with 100% of effectiveness for sensitivity and specificity greater than 83% for seizures events in all brain rhythms. Pearson's test suggests that all brain rhythms are highly correlated in non-seizure events but no during the seizure events. This suggests that our model can be scaled with the Pearson's product-moment correlation coefficient for the detection of epileptic seizures.

We develop a general non-asymptotic analysis of learning rates in kernel ridge regression (KRR), applicable for arbitrary Mercer kernels with multi-dimensional support. Our analysis is based on an operator-theoretic framework, at the core of which lies two error bounds under reproducing kernel Hilbert space norms encompassing a general class of kernels and regression functions, with remarkable extensibility to various inferential goals through augmenting results. When applied to KRR estimators, our analysis leads to error bounds under the stronger supremum norm, in addition to the commonly studied weighted $L_2$ norm; in a concrete example specialized to the Mat\'ern kernel, the established bounds recover the nearly minimax optimal rates. The wide applicability of our analysis is further demonstrated through two new theoretical results: (1) non-asymptotic learning rates for mixed partial derivatives of KRR estimators, and (2) a non-asymptotic characterization of the posterior variances of Gaussian processes, which corresponds to uncertainty quantification in kernel methods and nonparametric Bayes.

Submodular functions have been studied extensively in machine learning and data mining. In particular, the optimization of submodular functions over the integer lattice (integer submodular functions) has recently attracted much interest, because this domain relates naturally to many practical problem settings, such as multilabel graph cut, budget allocation and revenue maximization with discrete assignments. In contrast, the use of these functions for probabilistic modeling has received surprisingly little attention so far. In this work, we firstly propose the Generalized Multilinear Extension, a continuous DR-submodular extension for integer submodular functions. We study central properties of this extension and formulate a new probabilistic model which is defined through integer submodular functions. Then, we introduce a block-coordinate ascent algorithm to perform approximate inference for those class of models. Finally, we demonstrate its effectiveness and viability on several real-world social connection graph datasets with integer submodular objectives.

In multi-stage processes, decisions happen in an ordered sequence of stages. Many of them have the structure of dual funnel problem: as the sample size decreases from one stage to the other, the information increases. A related example is a selection process, where applicants apply for a position, prize, or grant. In each stage, more applicants are evaluated and filtered out, and from the remaining ones, more information is collected. In the last stage, decision-makers use all available information to make their final decision. To train a classifier for each stage becomes impracticable as they can underfit due to the low dimensionality in early stages or overfit due to the small sample size in the latter stages. In this work, we proposed a \textit{Multi-StaGe Transfer Learning} (MSGTL) approach that uses knowledge from simple classifiers trained in early stages to improve the performance of classifiers in the latter stages. By transferring weights from simpler neural networks trained in larger datasets, we able to fine-tune more complex neural networks in the latter stages without overfitting due to the small sample size. We show that it is possible to control the trade-off between conserving knowledge and fine-tuning using a simple probabilistic map. Experiments using real-world data demonstrate the efficacy of our approach as it outperforms other state-of-the-art methods for transfer learning and regularization.

We present a new method for understanding the performance of a model in virtual drug screening tasks. While most virtual screening problems present as a mix between ranking and classification, the models are typically trained as regression models presenting a problem requiring either a choice of a cutoff or ranking measure. Our method, regression enrichment surfaces (RES), is based on the goal of virtual screening: to detect as many of the top-performing treatments as possible. We outline history of virtual screening performance measures and the idea behind RES. We offer a python package and details on how to implement and interpret the results.

A common assumption of novelty detection is that the distribution of both "normal" and "novel" data are static. However, this is often not the case in scenarios where data evolves over time, or when the definition of normal and novel depends on contextual information, leading to changes in these distributions. This can lead to significant difficulties when attempting to train a model on datasets where the distribution of normal data in one scenario is similar to that of novel data in another scenario. In this paper we propose a context-aware approach to novelty detection for deep autoencoders. We create a semi-supervised network architecture which utilises auxiliary labels in order to reveal contextual information and allows the model to adapt to a variety of normal and novel scenarios. We evaluate our approach on both synthetic image data and real world audio data displaying these characteristics.

The two dominant approaches to neural text generation are fully autoregressive models, using serial beam search decoding, and non-autoregressive models, using parallel decoding with no output dependencies. This work proposes an autoregressive model with sub-linear parallel time generation. Noting that conditional random fields with bounded context can be decoded in parallel, we propose an efficient cascaded decoding approach for generating high-quality output. To parameterize this cascade, we introduce a Markov transformer, a variant of the popular fully autoregressive model that allows us to simultaneously decode with specific autoregressive context cutoffs. This approach requires only a small modification from standard autoregressive training, while showing competitive accuracy/speed tradeoff compared to existing methods on five machine translation datasets.

This paper studies model-based reinforcement learning (RL) for regret minimization. We focus on finite-horizon episodic RL where the transition model $P$ belongs to a known family of models $\mathcal{P}$, a special case of which is when models in $\mathcal{P}$ take the form of linear mixtures: $P_{\theta} = \sum_{i=1}^{d} \theta_{i}P_{i}$. We propose a model based RL algorithm that is based on optimism principle: In each episode, the set of models that are `consistent' with the data collected is constructed. The criterion of consistency is based on the total squared error of that the model incurs on the task of predicting \emph{values} as determined by the last value estimate along the transitions. The next value function is then chosen by solving the optimistic planning problem with the constructed set of models. We derive a bound on the regret, which, in the special case of linear mixtures, the regret bound takes the form $\tilde{\mathcal{O}}(d\sqrt{H^{3}T})$, where $H$, $T$ and $d$ are the horizon, total number of steps and dimension of $\theta$, respectively. In particular, this regret bound is independent of the total number of states or actions, and is close to a lower bound $\Omega(\sqrt{HdT})$. For a general model family $\mathcal{P}$, the regret bound is derived using the notion of the so-called Eluder dimension proposed by Russo & Van Roy (2014).

It is proposed to use analogs of the forecast mean to generate an ensemble of perturbations for use in ensemble optimal interpolation (EnOI) or ensemble variational (EnVar) methods. A new method of constructing analogs using variational autoencoders (VAEs; a machine learning method) is proposed. The resulting analog methods using analogs from a catalog (AnEnOI), and using constructed analogs (cAnEnOI), are tested in the context of a multiscale Lorenz-`96 model, with standard EnOI and an ensemble square root filter for comparison. The use of analogs from a modestly-sized catalog is shown to improve the performance of EnOI, with limited marginal improvements resulting from increases in the catalog size. The method using constructed analogs (cAnEnOI) is found to perform as well as a full ensemble square root filter, and to be robust over a wide range of tuning parameters.