Recent successes in Reinforcement Learning have encouraged a fast-growing network of RL researchers and a number of breakthroughs in RL research. As the RL community and the body of RL work grows, so does the need for widely applicable benchmarks that can fairly and effectively evaluate a variety of RL algorithms. This need is particularly apparent in the realm of Hierarchical Reinforcement Learning (HRL). While many existing test domains may exhibit hierarchical action or state structures, modern RL algorithms still exhibit great difficulty in solving domains that necessitate hierarchical modeling and action planning, even when such domains are seemingly trivial. These difficulties highlight both the need for more focus on HRL algorithms themselves, and the need for new testbeds that will encourage and validate HRL research. Existing HRL testbeds exhibit a Goldilocks problem; they are often either too simple (e.g. Taxi) or too complex (e.g. Montezuma's Revenge from the Arcade Learning Environment). In this paper we present the Escape Room Domain (ERD), a new flexible, scalable, and fully implemented testing domain for HRL that bridges the "moderate complexity" gap left behind by existing alternatives. ERD is open-source and freely available through GitHub, and conforms to widely-used public testing interfaces for simple integration and testing with a variety of public RL agent implementations. We show that the ERD presents a suite of challenges with scalable difficulty to provide a smooth learning gradient from Taxi to the Arcade Learning Environment.

Future advancement of engineering applications is dependent on design of novel materials with desired properties. Enormous size of known chemical space necessitates use of automated high throughput screening to search the desired material. The high throughput screening uses quantum chemistry calculations to predict material properties, however, computational complexity of these calculations often imposes prohibitively high cost on the search for desired material. This critical bottleneck is resolved by using deep machine learning to emulate the quantum computations. However, the deep learning algorithms require a large training dataset to ensure an acceptable generalization, which is often unavailable a-priory. In this paper, we propose a deep Gaussian process based approach to develop an emulator for quantum calculations. We further propose a novel molecular descriptor that enables implementation of the proposed approach. As demonstrated in this paper, the proposed approach can be implemented using a small dataset. We demonstrate efficacy of our approach for prediction of formation energy of inorganic molecules.

This paper proposes a novel framework to regularize the highly ill-posed and non-linear Fourier ptychography problem using generative models. We demonstrate experimentally that our proposed algorithm, Deep Ptych, outperforms the existing Fourier ptychography techniques, in terms of quality of reconstruction and robustness against noise, using far fewer samples. We further modify the proposed approach to allow the generative model to explore solutions outside the range, leading to improved performance.

Recurrent neural networks (RNNs) have become one of the standard models in sequential data analysis [Rumelhart et al., 1986, Elman, 1990]. At each time step of the RNN, an observation is modeled via a neural network using the observations and hidden states from previous time points. Models such as the RNN, and also the hidden Markov model among others, often implicitly assume a sequence as having a fixed time interval between observations. They also often do not account for group-level effects when multiple sequences are observed and each sequence belongs to one of multiple groups. For example, consider data in the form of a sequence of discrete counts by a set of groups-- e.g., a sequence of purchases (market baskets) for a set of customers, with one sequence per customer. A vanilla RNN implementation would model these sequences using a network with the same parameters, which removes the customer-level information, and according to an enumerated indexing, which removes the time interval information between orders. However, this information is important: customer-specific effects can improve predictive performance for each customer, while an interval of one day versus one month between orders significantly impacts the items likely to be purchased next.

We consider a framework involving behavioral economics and machine learning. Rationally inattentive Bayesian agents make decisions based on their posterior distribution, utility function and information acquisition cost Renyi divergence which generalizes Shannon mutual information). By observing these decisions, how can an observer estimate the utility function and information acquisition cost? Using deep learning, we estimate framing information (essential extrinsic features) that determines the agent's attention strategy. Then we present a preference based inverse reinforcement learning algorithm to test for rational inattention: is the agent an utility maximizer, attention maximizer, and does an information cost function exist that rationalizes the data? The test imposes a Renyi mutual information constraint which impacts how the agent can select attention strategies to maximize their expected utility. The test provides constructive estimates of the utility function and information acquisition cost of the agent. We illustrate these methods on a massive YouTube dataset for characterizing the commenting behavior of users.

In structured output prediction tasks, labeling ground-truth training output is often expensive. However, for many tasks, even when the true output is unknown, we can evaluate predictions using a scalar reward function, which may be easily assembled from human knowledge or non-differentiable pipelines. But searching through the entire output space to find the best output with respect to this reward function is typically intractable. In this paper, we instead use efficient truncated randomized search in this reward function to train structured prediction energy networks (SPENs), which provide efficient test-time inference using gradient-based search on a smooth, learned representation of the score landscape, and have previously yielded state-of-the-art results in structured prediction. In particular, this truncated randomized search in the reward function yields previously unknown local improvements, providing effective supervision to SPENs, avoiding their traditional need for labeled training data.

Price responsiveness is a major feature of end use customers (EUCs) that participate in demand response (DR) programs, and has been conventionally modeled with static demand functions, which take the electricity price as the input and the aggregate energy consumption as the output. This, however, neglects the inherent temporal correlation of the EUC behaviors, and may result in large errors when predicting the actual responses of EUCs in real-time pricing (RTP) programs. In this paper, we propose a dynamical DR model so as to capture the temporal behavior of the EUCs. The states in the proposed dynamical DR model can be explicitly chosen, in which case the model can be represented by a linear function or a multi-layer feedforward neural network, or implicitly chosen, in which case the model can be represented by a recurrent neural network or a long short-term memory unit network. In both cases, the dynamical DR model can be learned from historical price and energy consumption data. Numerical simulation illustrated how the states are chosen and also showed the proposed dynamical DR model significantly outperforms the static ones.

Heart failure (HF) is one of the leading causes of hospital admissions in the US. Readmission within 30 days after a HF hospitalization is both a recognized indicator for disease progression and a source of considerable financial burden to the healthcare system. Consequently, the identification of patients at risk for readmission is a key step in improving disease management and patient outcome. In this work, we used a large administrative claims dataset to (1)explore the systematic application of neural network-based models versus logistic regression for predicting 30 days all-cause readmission after discharge from a HF admission, and (2)to examine the additive value of patients' hospitalization timelines on prediction performance. Based on data from 272,778 (49% female) patients with a mean (SD) age of 73 years (14) and 343,328 HF admissions (67% of total admissions), we trained and tested our predictive readmission models following a stratified 5-fold cross-validation scheme. Among the deep learning approaches, a recurrent neural network (RNN) combined with conditional random fields (CRF) model (RNNCRF) achieved the best performance in readmission prediction with 0.642 AUC (95% CI, 0.640-0.645). Other models, such as those based on RNN, convolutional neural networks and CRF alone had lower performance, with a non-timeline based model (MLP) performing worst. A competitive model based on logistic regression with LASSO achieved a performance of 0.643 AUC (95%CI, 0.640-0.646). We conclude that data from patient timelines improve 30 day readmission prediction for neural network-based models, that a logistic regression with LASSO has equal performance to the best neural network model and that the use of administrative data result in competitive performance compared to published approaches based on richer clinical datasets.

This work investigates the ways in which deep learning methods can benefit from random projection (RP), a classic linear dimensionality reduction method. We focus on two areas where, as we have found, employing RP techniques can improve deep models: training neural networks on high-dimensional data and initialization of network parameters. Training deep neural networks (DNNs) on sparse, high-dimensional data with no exploitable structure implies a network architecture with an input layer that has a huge number of weights, which often makes training infeasible. We show that this problem can be solved by prepending the network with an input layer whose weights are initialized with an RP matrix. We propose several modifications to the network architecture and training regime that makes it possible to efficiently train DNNs with learnable RP layer on data with as many as tens of millions of input features and training examples. In comparison to the state-of-the-art methods, neural networks with RP layer achieve competitive performance or improve the results on several extremely high-dimensional real-world datasets. The second area where the application of RP techniques can be beneficial for training deep models is weight initialization. Setting the initial weights in DNNs to elements of various RP matrices enabled us to train residual deep networks to higher levels of performance.

In this paper, we propose a new differentiable neural network alignment mechanism for text-dependent speaker verification which uses alignment models to produce a supervector representation of an utterance. Unlike previous works with similar approaches, we do not extract the embedding of an utterance from the mean reduction of the temporal dimension. Our system replaces the mean by a phrase alignment model to keep the temporal structure of each phrase which is relevant in this application since the phonetic information is part of the identity in the verification task. Moreover, we can apply a convolutional neural network as front-end, and thanks to the alignment process being differentiable, we can train the whole network to produce a supervector for each utterance which will be discriminative with respect to the speaker and the phrase simultaneously. As we show, this choice has the advantage that the supervector encodes the phrase and speaker information providing good performance in text-dependent speaker verification tasks. In this work, the process of verification is performed using a basic similarity metric, due to simplicity, compared to other more elaborate models that are commonly used. The new model using alignment to produce supervectors was tested on the RSR2015-Part I database for text-dependent speaker verification, providing competitive results compared to similar size networks using the mean to extract embeddings.