The 19th-century physicist Hermann von Helmholtz compared his progress in solving a problem to that of a mountain climber "compelled to retrace his steps because his progress stopped." A mountain climber, von Helmholtz said, "hits upon traces of a fresh path, which again leads him a little further." The physicist's introspection provokes the question: How do creative minds overcome valleys to get to the next higher peak? Because thinking minds are different from evolving organisms and self-assembling molecules, we cannot expect them to use the same means--mechanisms like genetic drift and thermal vibrations--to overcome deep valleys in the landscapes they explore. But they must have some way to achieve the same purpose.

Recently, several studies have proven the global convergence and generalization abilities of the gradient descent method for two-layer ReLU networks by making a positivity assumption of the Gram-matrix of the neural tangent kernel. However, the performance of gradient descent on classification problems has not been well studied, and further investigation of the problem structure is possible. In this work, we present a partially stronger but reasonable assumption for binary classification problems compared to the positivity assumption of the Gram-matrix, where a data distribution can be perfectly classifiable by a tangent model, and we provide a refined generalization analysis of the gradient descent method for two-layer networks with smooth activations. A remarkable point of this study is that our generalization bound has much better dependence on the network width compared to existing results. As a result, our theory significantly enlarges a class of over-parameterized networks having provable generalization ability, with respect to network width, while most studies require much higher over-parameterization.

Multi-hop Reading Comprehension (RC) requires reasoning and aggregation across several paragraphs. We propose a system for multi-hop RC that decomposes a compositional question into simpler sub-questions that can be answered by off-the-shelf single-hop RC models. Since annotations for such decomposition are expensive, we recast sub-question generation as a span prediction problem and show that our method, trained using only 400 labeled examples, generates sub-questions that are as effective as human-authored sub-questions. We also introduce a new global rescoring approach that considers each decomposition (i.e. the sub-questions and their answers) to select the best final answer, greatly improving overall performance. Our experiments on HotpotQA show that this approach achieves the state-of-the-art results, while providing explainable evidence for its decision making in the form of sub-questions.

The timeliness of detection of a sepsis event in progress is a crucial factor in the outcome for the patient. Machine learning models built from data in electronic health records can be used as an effective tool for improving this timeliness, but so far the potential for clinical implementations has been largely limited to studies in intensive care units. This study will employ a richer data set that will expand the applicability of these models beyond intensive care units. Furthermore, we will circumvent several important limitations that have been found in the literature: 1) Models are evaluated shortly before sepsis onset without considering interventions already initiated. 2) Machine learning models are built on a restricted set of clinical parameters, which are not necessarily measured in all departments. 3) Model performance is limited by current knowledge of sepsis, as feature interactions and time dependencies are hardcoded into the model. In this study, we present a model to overcome these shortcomings using a deep learning approach on a diverse multicenter data set. We used retrospective data from multiple Danish hospitals over a seven-year period. Our sepsis detection system is constructed as a combination of a convolutional neural network and a long short-term memory network. We suggest a retrospective assessment of interventions by looking at intravenous antibiotics and blood cultures preceding the prediction time. Results show performance ranging from AUROC 0.856 (3 hours before sepsis onset) to AUROC 0.756 (24 hours before sepsis onset). We present a deep learning system for early detection of sepsis that is able to learn characteristics of the key factors and interactions from the raw event sequence data itself, without relying on a labor-intensive feature extraction work.

One of the biggest challenges in robotics systems is interacting under uncertainty. Unlike robots, humans learn, adapt and perceive their body as a unity when interacting with the world. We hypothesize that the nervous system counteracts sensor and motor uncertainties by unconscious processes that robustly fuse the available information for approximating their body and the world state. Being able to unite perception and action under a common principle has been sought for decades and active inference is one of the potential unification theories. In this work, we present a humanoid robot interacting with the world by means of a human brain-like inspired perception and control algorithm based on the free-energy principle. Until now, active inference was only tested in simulated examples. Their application on a real robot shows the advantages of such an algorithm for real world applications. The humanoid robot iCub was capable of performing robust reaching behaviors with both arms and active head object tracking in the visual field, despite the visual noise, the artificially introduced noise in the joint encoders (up to 40 degrees deviation), the differences between the model and the real robot and the misdetections of the hand.

Accurate estimation of the run time of computational codes has a number of significant advantages for scientific computing. It is required information for optimal resource allocation, improving turnaround times and utilization of science gateways. Furthermore, it allows users to better plan and schedule their research, streamlining workflows and improving the overall productivity of cyberinfrastructure. Predicting run time is challenging, however. The inputs to scientific codes can be complex and high dimensional. Their relationship to the run time may be highly non-linear, and, in the most general case is completely arbitrary and thus unpredictable (i.e., simply a random mapping from inputs to run time). Most codes are not so arbitrary, however, and there has been significant prior research on predicting the run time of applications and workloads. Such predictions are generally application-specific, however. In this paper, we focus on the Gaussian computational chemistry code. We characterize a data set of runs from the SEAGrid science gateway with a number of different studies. We also explore a number of different potential regression methods and present promising future directions.

Multi-hop reading comprehension (RC) questions are challenging because they require reading and reasoning over multiple paragraphs. We argue that it can be difficult to construct large multi-hop RC datasets. For example, even highly compositional questions can be answered with a single hop if they target specific entity types, or the facts needed to answer them are redundant. Our analysis is centered on HotpotQA, where we show that single-hop reasoning can solve much more of the dataset than previously thought. We introduce a single-hop BERT-based RC model that achieves 67 F1---comparable to state-of-the-art multi-hop models. We also design an evaluation setting where humans are not shown all of the necessary paragraphs for the intended multi-hop reasoning but can still answer over 80% of questions. Together with detailed error analysis, these results suggest there should be an increasing focus on the role of evidence in multi-hop reasoning and possibly even a shift towards information retrieval style evaluations with large and diverse evidence collections.

We propose a scalable framework for inference in an inhomogeneous Poisson process modeled by a continuous sigmoidal Cox process that assumes the corresponding intensity function is given by a Gaussian process (GP) prior transformed with a scaled logistic sigmoid function. We present a tractable representation of the likelihood through augmentation with a superposition of Poisson processes. This view enables a structured variational approximation capturing dependencies across variables in the model. Our framework avoids discretization of the domain, does not require accurate numerical integration over the input space and is not limited to GPs with squared exponential kernels. We evaluate our approach on synthetic and real-world data showing that its benefits are particularly pronounced on multivariate input settings where it overcomes the limitations of mean-field methods and sampling schemes. We provide the state of-the-art in terms of speed, accuracy and uncertainty quantification trade-offs.

General purpose relation extractors, which can model arbitrary relations, are a core aspiration in information extraction. Efforts have been made to build general purpose extractors that represent relations with their surface forms, or which jointly embed surface forms with relations from an existing knowledge graph. However, both of these approaches are limited in their ability to generalize. In this paper, we build on extensions of Harris' distributional hypothesis to relations, as well as recent advances in learning text representations (specifically, BERT), to build task agnostic relation representations solely from entity-linked text. We show that these representations significantly outperform previous work on exemplar based relation extraction (FewRel) even without using any of that task's training data. We also show that models initialized with our task agnostic representations, and then tuned on supervised relation extraction datasets, significantly outperform the previous methods on SemEval 2010 Task 8, KBP37, and TACRED.

In this paper, we empirically investigate applying word-level weights to adapt neural machine translation to e-commerce domains, where small e-commerce datasets and large out-of-domain datasets are available. In order to mine in-domain like words in the out-of-domain datasets, we compute word weights by using a domain-specific and a non-domain-specific language model followed by smoothing and binary quantization. The baseline model is trained on mixed in-domain and out-of-domain datasets. Experimental results on English to Chinese e-commerce domain translation show that compared to continuing training without word weights, it improves MT quality by up to 2.11% BLEU absolute and 1.59% TER. We have also trained models using fine-tuning on the in-domain data. Pre-training a model with word weights improves fine-tuning up to 1.24% BLEU absolute and 1.64% TER, respectively.