Our aim is to establish a framework where reinforcement learning (RL) of optimizing interventions retrospectively allows us a regulatory compliant pathway to prospective clinical testing of the learned policies in a clinical deployment. We focus on infections in intensive care units which are one of the major causes of death and difficult to treat because of the complex and opaque patient dynamics, and the clinically debated, highly-divergent set of intervention policies required by each individual patient, yet intensive care units are naturally data rich. In our work, we build on RL approaches in healthcare ("AI Clinicians"), and learn off-policy continuous dosing policy of pharmaceuticals for sepsis treatment using historical intensive care data under partially observable MDPs (POMDPs). POMPDs capture uncertainty in patient state better by taking in all historical information, yielding an efficient representation, which we investigate through ablations. We compensate for the lack of exploration in our retrospective data by evaluating each encountered state with a best-first tree search. We mitigate state distributional shift by optimizing our policy in the vicinity of the clinicians' compound policy. Crucially, we evaluate our model recommendations using not only conventional policy evaluations but a novel framework that incorporates human experts: a model-agnostic pre-clinical evaluation method to estimate the accuracy and uncertainty of clinician's decisions versus our system recommendations when confronted with the same individual patient history ("shadow mode").

Spin glass models, such as the Sherrington-Kirkpatrick, Hopfield and Ising models, are all well-studied members of the exponential family of discrete distributions, and have been influential in a number of application domains where they are used to model correlation phenomena on networks. Conventionally these models have quadratic sufficient statistics and consequently capture correlations arising from pairwise interactions. In this work we study extensions of these to models with higher-order sufficient statistics, modeling behavior on a social network with peer-group effects. In particular, we model binary outcomes on a network as a higher-order spin glass, where the behavior of an individual depends on a linear function of their own vector of covariates and some polynomial function of the behavior of others, capturing peer-group effects. Using a {\em single}, high-dimensional sample from such model our goal is to recover the coefficients of the linear function as well as the strength of the peer-group effects. The heart of our result is a novel approach for showing strong concavity of the log pseudo-likelihood of the model, implying statistical error rate of $\sqrt{d/n}$ for the Maximum Pseudo-Likelihood Estimator (MPLE), where $d$ is the dimensionality of the covariate vectors and $n$ is the size of the network (number of nodes). Our model generalizes vanilla logistic regression as well as the peer-effect models studied in recent works, and our results extend these results to accommodate higher-order interactions.

Learning continuous representations of discrete objects such as text, users, and URLs lies at the heart of many applications including language and user modeling. When using discrete objects as input to neural networks, we often ignore the underlying structures (e.g. natural groupings and similarities) and embed the objects independently into individual vectors. As a result, existing methods do not scale to large vocabulary sizes. In this paper, we design a Bayesian nonparametric prior for embeddings that encourages sparsity and leverages natural groupings among objects. We derive an approximate inference algorithm based on Small Variance Asymptotics which yields a simple and natural algorithm for learning a small set of anchor embeddings and a sparse transformation matrix. We call our method Anchor & Transform (ANT) as the embeddings of discrete objects are a sparse linear combination of the anchors, weighted according to the transformation matrix. ANT is scalable, flexible, end-to-end trainable, and allows the user to incorporate domain knowledge about object relationships. On text classification and language modeling benchmarks, ANT demonstrates stronger performance with fewer parameters as compared to existing compression baselines.

There is a rich literature on Bayesian nonparametric methods for unknown densities. The most popular approach relies on Dirichlet process mixture models. These models characterize the unknown density as a kernel convolution with an unknown almost surely discrete mixing measure, which is given a Dirichlet process prior. Such models are very flexible and have good performance in many settings, but posterior computation relies on Markov chain Monte Carlo algorithms that can be complex and inefficient. As a simple and general alternative, we propose a class of nearest neighbor-Dirichlet processes. The approach starts by grouping the data into neighborhoods based on standard algorithms. Within each neighborhood, the density is characterized via a Bayesian parametric model, such as a Gaussian with unknown parameters. Assigning a Dirichlet prior to the weights on these local kernels, we obtain a simple pseudo-posterior for the weights and kernel parameters. A simple and embarrassingly parallel Monte Carlo algorithm is proposed to sample from the resulting pseudo-posterior for the unknown density. Desirable asymptotic properties are shown, and the methods are evaluated in simulation studies and applied to a motivating dataset in the context of classification.

Label shift describes the setting where although the label distribution might change between the source and target domains, the class-conditional probabilities (of data given a label) do not. There are two dominant approaches for estimating the label marginal. BBSE, a moment-matching approach based on confusion matrices, is provably consistent and provides interpretable error bounds. However, a maximum likelihood estimation approach, which we call MLLS, dominates empirically. In this paper, we present a unified view of the two methods and the first theoretical characterization of the likelihood-based estimator. Our contributions include (i) conditions for consistency of MLLS, which include calibration of the classifier and a confusion matrix invertibility condition that BBSE also requires; (ii) a unified view of the methods, casting the confusion matrix as roughly equivalent to MLLS for a particular choice of calibration method; and (iii) a decomposition of MLLS's finite-sample error into terms reflecting the impacts of miscalibration and estimation error. Our analysis attributes BBSE's statistical inefficiency to a loss of information due to coarse calibration. We support our findings with experiments on both synthetic data and the MNIST and CIFAR10 image recognition datasets.

Many machine learning problems consider data where each instance is, itself, an unordered set of elements; i.e., such that each observation is a set. Data of this kind arises in a variety of applications, ranging from document modeling (Blei et al., 2003; Garnelo et al., 2018a) and multi-task learning (Zaheer et al., 2017; Edwards & Storkey, 2016; Liu et al., 2019) to 3D point cloud modeling (Li et al., 2018; Yang et al., 2019). In unsupervised settings, a dataset typically consists of a set of such sets, while in supervised learning, it consists of a set of (set, label) pairs. Modeling a distribution over a space of instances, where each instance is, itself, an unordered set of elements involves two key considerations: (1) the elements within a single instance are exchangeable, i.e., the elements are order invariant; and (2) the cardinalities of the instances (sets) vary, i.e., instances need not exhibit the same cardinality. Modeling both unconditional and conditional distributions over instances (sets) are relevant to consider, since these support unsupervised and supervised tasks respectively. For unconditional distribution modeling, there has been significant prior work on modeling set distributions, which has sought to balance competing needs to expand model flexibility and preserve tractability on the one hand, with respecting exchangeability and varying instance cardinalities on the other hand. However, managing these tradeoffs has proved to be quite difficult, and current approaches remain limited in different respects. For example, a particularly straightforward strategy for modeling distributions over instances x {x 1,..., x n } without assuming fixed cardinality is simply to use a recurrent neural network (RNNs) to encode instance probability auto-regressively via p (x) n

We consider how to effectively use prior knowledge when learning a Bayesian model from streaming environments where the data come infinitely and sequentially. This problem is highly important in the era of data explosion and rich sources of precious external knowledge such as pre-trained models, ontologies, Wikipedia, etc. We show that some existing approaches can forget any knowledge very fast. We then propose a novel framework that enables to incorporate the prior knowledge of different forms into a base Bayesian model for data streams. Our framework subsumes some existing popular models for time-series/dynamic data. Extensive experiments show that our framework outperforms existing methods with a large margin. In particular, our framework can help Bayesian models generalize well on extremely short text while other methods overfit. The implementation of our framework is available at https://github.com/bachtranxuan/TPS.git.

Learning hidden topics in data streams has been paid a great deal of attention by researchers with a lot of proposed methods, but exploiting prior knowledge in general and a knowledge graph in particular has not been taken into adequate consideration in these methods. Prior knowledge that is derived from human knowledge (e.g. Wordnet) or a pre-trained model (e.g.Word2vec) is very valuable and useful to help topic models work better, especially on short texts. However, previous work often ignores this resource, or it can only utilize prior knowledge of a vector form in a simple way. In this paper, we propose a novel graph convolutional topic model (GCTM) which integrates graph convolutional networks (GCN) into a topic model and a learning method which learns the networks and the topic model simultaneously for data streams. In each minibatch, our method not only can exploit an external knowledge graph but also can balance between the external and old knowledge to perform well on new data. We conduct extensive experiments to evaluate our method with both human graph knowledge(Wordnet) and a graph built from pre-trained word embeddings (Word2vec). The experimental results show that our method achieves significantly better performances than the state-of-the-art baselines in terms of probabilistic predictive measure and topic coherence. In particular, our method can work well when dealing with short texts as well as concept drift. The implementation of GCTM is available at https://github.com/bachtranxuan/GCTM.git.

Next location prediction is of great importance for many location-based applications and provides essential intelligence to business and governments. In existing studies, a common approach to next location prediction is to learn the sequential transitions with massive historical trajectories based on conditional probability. Unfortunately, due to the time and space complexity, these methods (e.g., Markov models) only use the just passed locations to predict next locations, without considering all the passed locations in the trajectory. In this paper, we seek to enhance the prediction performance by considering the travel time from all the passed locations in the query trajectory to a candidate next location. In particular, we propose a novel method, called Travel Time Difference Model (TTDM), which exploits the difference between the shortest travel time and the actual travel time to predict next locations. Further, we integrate the TTDM with a Markov model via a linear interpolation to yield a joint model, which computes the probability of reaching each possible next location and returns the top-rankings as results. We have conducted extensive experiments on two real datasets: the vehicle passage record (VPR) data and the taxi trajectory data. The experimental results demonstrate significant improvements in prediction accuracy over existing solutions. For example, compared with the Markov model, the top-1 accuracy improves by 40% on the VPR data and by 15.6% on the Taxi data.

Providing explanations of chosen robotic actions can help to increase the transparency of robotic planning and improve users' trust. Social sciences suggest that the best explanations are contrastive, explaining not just why one action is taken, but why one action is taken instead of another. We formalize the notion of contrastive explanations for robotic planning policies based on Markov decision processes, drawing on insights from the social sciences. We present methods for the automated generation of contrastive explanations with three key factors: selectiveness, constrictiveness, and responsibility. The results of a user study with 100 participants on the Amazon Mechanical Turk platform show that our generated contrastive explanations can help to increase users' understanding and trust of robotic planning policies while reducing users' cognitive burden.