Chart-based visual acuity measurements are used by billions of people to diagnose and guide treatment of vision impairment. However, the ubiquitous eye exam has no mechanism for reasoning about uncertainty and as such, suffers from a well-documented reproducibility problem. In this paper we uncover a new parametric probabilistic model of visual acuity response based on measurements of patients with eye disease. We present a state of the art eye exam which (1) reduces acuity exam error by 75\% without increasing exam length, (2) knows how confident it should be, (3) can trace predictions over time and incorporate prior beliefs and (4) provides insight for educational Item Response Theory. For patients with more serious eye disease, the novel ability to finely measure acuity from home could be a crucial part in early diagnosis. We provide a web implementation of our algorithm for anyone in the world to use.

We propose and investigate new complementary methodologies for estimating predictive variance networks in regression neural networks. We derive a locally aware mini-batching scheme that result in sparse robust gradients, and show how to make unbiased weight updates to a variance network. Further, we formulate a heuristic for robustly fitting both the mean and variance networks post hoc. Finally, we take inspiration from posterior Gaussian processes and propose a network architecture with similar extrapolation properties to Gaussian processes. The proposed methodologies are complementary, and improve upon baseline methods individually. Experimentally, we investigate the impact on predictive uncertainty on multiple datasets and tasks ranging from regression, active learning and generative modeling. Experiments consistently show significant improvements in predictive uncertainty estimation over state-of-the-art methods across tasks and datasets.

Policy gradients methods are perhaps the most widely used class of reinforcement learning algorithms. These methods apply to complex, poorly understood, control problems by performing stochastic gradient descent over a parameterized class of polices. Unfortunately, even for simple control problems solvable by classical techniques, policy gradient algorithms face non-convex optimization problems and are widely understood to converge only to local minima. This work identifies structural properties -- shared by finite MDPs and several classic control problems -- which guarantee that policy gradient objective function has no suboptimal local minima despite being non-convex. When these assumptions are relaxed, our work gives conditions under which any local minimum is near-optimal, where the error bound depends on a notion of the expressive capacity of the policy class.

Evaluation of the marginal likelihood plays an important role in model selection problems. The widely applicable Bayesian information criterion (WBIC) and singular Bayesian information criterion (sBIC) give approximations to the log marginal likelihood, which can be applied to both regular and singular models. When the real log canonical thresholds are known, the performance of sBIC is considered to be better than that of WBIC, but only few real log canonical thresholds are known. In this paper, we propose a new estimator of the real log canonical thresholds based on the variance of thermodynamic integration with an inverse temperature. In addition, we propose an application to make sBIC widely applicable. Finally, we investigate the performance of the estimator and model selection by simulation studies and application to real data.

While label fusion from multiple noisy annotations is a well understood concept in data wrangling (tackled for example by the Dawid-Skene (DS) model), we consider the extended problem of carrying out learning when the labels themselves are not consistently annotated with the same schema. We show that even if annotators use disparate, albeit related, label-sets, we can still draw inferences for the underlying full label-set. We propose the Inter-Schema AdapteR (ISAR) to translate the fully-specified label-set to the one used by each annotator, enabling learning under such heterogeneous schemas, without the need to re-annotate the data. We apply our method to a mouse behavioural dataset, achieving significant gains (compared with DS) in out-of-sample log-likelihood (-3.40 to -2.39) and F1-score (0.785 to 0.864).

Imagine you're sleeping, and you hear strange noises in your front lawn. You're very sleepy, so you hypothesize that the strange noises are being generated by a hungry dinosaur. You think to yourself, 'this is exactly what I would hear if there was a dinosaur outside in my front lawn'. But then as you think more about it, you realize that the likelihood of there actually being a dinosaur in your front lawn is extremely low; whereas the likelihood of hearing strange noises from the front lawn is likely pretty high. So you exhale as you realize that the actual probability of there being a dinosaur in your front lawn, aka your original hypothesis, given the evidence is extremely low.

Neural Networks (NNs) have been extensively used for a wide spectrum of real-world regression tasks, where the goal is to predict a numerical outcome such as revenue, effectiveness, or a quantitative result. In many such tasks, the point prediction is not enough, but also the uncertainty (i.e. risk, or confidence) of that prediction must be estimated. Standard NNs, which are most often used in such tasks, do not provide any such information. Existing approaches try to solve this issue by combining Bayesian models with NNs, but these models are hard to implement, more expensive to train, and usually do not perform as well as standard NNs. In this paper, a new framework called RIO is developed that makes it possible to estimate uncertainty in any pretrained standard NN. RIO models prediction residuals using Gaussian Process with a composite input/output kernel. The residual prediction and I/O kernel are theoretically motivated and the framework is evaluated in twelve real-world datasets. It is found to provide reliable estimates of the uncertainty, reduce the error of the point predictions, and scale well to large datasets. Given that RIO can be applied to any standard NN without modifications to model architecture or training pipeline, it provides an important ingredient in building real-world applications of NNs.

Making high quality inference on large, feature rich datasets under a constrained computational budget is arguably the primary goal of the learning community. This, however, comes with significant challenges. On the one hand, the exact computation of linear algebraic quantities may be prohibitively expensive, such as that of the log determinant. On the other hand, an analytic expression for the quantity of interest may not exist at all, such as the case for the entropy of a Gaussian mixture model, and approximate methods are often both inefficient and inaccurate.

Class-conditional generative models are an increasingly popular approach to achieve robust classification. They are a natural choice to solve discriminative tasks in a robust manner as they jointly optimize for predictive performance and accurate modeling of the input distribution. In this work, we investigate robust classification with likelihood-based conditional generative models from a theoretical and practical perspective. Our theoretical result reveals that it is impossible to guarantee detectability of adversarial examples even for near-optimal generative classifiers. Experimentally, we show that naively trained conditional generative models have poor discriminative performance, making them unsuitable for classification. This is related to overlooked issues with training conditional generative models and we show methods to improve performance. Finally, we analyze the robustness of our proposed conditional generative models on MNIST and CIFAR10. While we are able to train robust models for MNIST, robustness completely breaks down on CIFAR10. This lack of robustness is related to various undesirable model properties maximum likelihood fails to penalize. Our results indicate that likelihood may fundamentally be at odds with robust classification on challenging problems.

We propose to train Bayesian Neural Networks (BNNs) by empirical Bayes as an alternative to posterior weight inference. By approximately marginalizing out an i.i.d.\ realization of a finite number of sibling weights per data-point using the Central Limit Theorem (CLT), we attain a scalable and effective Bayesian deep predictor. This approach directly models the posterior predictive distribution, by-passing the intractable posterior weight inference step. However, it introduces a prohibitively large number of hyperparameters for stable training. As the prior weights are marginalized and hyperparameters are optimized, the model also no longer provides a means to incorporate prior knowledge. We overcome both of these drawbacks by deriving a trivial PAC bound that comprises the marginal likelihood of the predictor and a complexity penalty. The outcome integrates organically into the prior networks framework, bringing about an effective and holistic Bayesian treatment of prediction uncertainty. We observe on various regression, classification, and out-of-domain detection benchmarks that our scalable method provides an improved model fit accompanied with significantly better uncertainty estimates than the state-of-the-art.