Search

We propose a new setting for question answering (QA) in which users can query the system using both natural language and direct interactions within a graphical user interface that displays multiple time series associated with an entity of interest. The user interacts with the interface in order to understand the entity’s state and behavior, entailing sequences of actions and questions whose answers may depend on previous factual or navigational interactions. We describe a pipeline implementation where spoken questions are first transcribed into text which is then semantically parsed into logical forms that can be used to automatically extract the answer from the underlying database. The speech recognition module is implemented by adapting a pre-trained long short-term memory (LSTM)-based architecture to the user’s speech, whereas for the semantic parsing component we introduce an LSTM-based encoder–decoder architecture that models context dependency through copying mechanisms and multiple levels of attention over inputs and previous outputs. When evaluated separately, with and without data augmentation, both models are shown to substantially outperform several strong baselines. Furthermore, the full pipeline evaluation shows only a small degradation in semantic parsing accuracy, demonstrating that the semantic parser is robust to mistakes in the speech recognition output. The new QA paradigm proposed in this paper has the potential to improve the presentation and navigation of the large amounts of sensor data and life events that are generated in many areas of medicine.

As the number of processing cores and associated threads in chip multiprocessors (CMPs) continues to scale out, on-chip memory access latency dominates application execution time due to increased data movement. Although tiled CMP architectures with distributed shared caches provide a scalable design, increased physical distance between requesting and responding cores has led to both increased on-chip memory access latency and excess energy consumption. Near data processing is a promising approach that can migrate threads closer to data, however prior hand-engineered rules for fine-grained hardware-level thread migration are either too slow to react to changes in data access patterns, or unable to exploit the large variety of data access patterns. In this article, we propose to use reinforcement learning (RL) to learn relatively complex data access patterns to improve on hardware-level thread migration techniques. By utilizing the recent history of memory access locations as input, each thread learns to recognize the relationship between prior access patterns and future memory access locations. This leads to the unique ability of the proposed technique to make fewer, more effective migrations to intermediate cores that minimize the distance to multiple distinct memory access locations. By allowing a low-overhead RL agent to learn a policy from real interaction with parallel programming benchmarks in a parallel simulator, we show that a migration policy which recognizes more complex data access patterns can be learned. The proposed approach reduces on-chip data movement and energy consumption by an average of 41%, while reducing execution time by 43% when compared to a simple baseline with no thread migration; furthermore, energy consumption and execution time are reduced by an additional 10% when compared to a hand-engineered fine-grained migration policy.

Hardware accelerators provide significant speedup and improve energy efficiency for several demanding deep neural network (DNN) applications. DNNs have several hidden layers that perform concurrent matrix-vector multiplications (MVMs) between the network weights and input features. As MVMs are critical to the performance of DNNs, previous research has optimized the performance and energy efficiency of MVMs at both the architecture and algorithm levels. In this paper, we propose to use emerging silicon photonics technology to improve parallelism, speed and overall efficiency with the goal of providing real-time inference and fast training of neural nets. We use microring resonators (MRRs) and Mach-Zehnder interferometers (MZIs) to design two versions (all-optical and partial-optical) of hybrid matrix multiplications for DNNs. Our results indicate that our partial optical design gave the best performance in both energy efficiency and latency, with a reduction of 33.1% for energy-delay product (EDP) with conservative estimates and a 76.4% reduction for EDP with aggressive estimates.

To avoid serious diabetic complications, people with type 1 diabetes must keep their blood glucose levels (BGLs) as close to normal as possible. Insulin dosages and carbohydrate consumption are important considerations in managing BGLs. Since the 1960s, models have been developed to forecast blood glucose levels based on the history of BGLs, insulin dosages, carbohydrate intake, and other physiological and lifestyle factors. Such predictions can be used to alert people of impending unsafe BGLs or to control insulin flow in an artificial pancreas. In past work, we have introduced an LSTM-based approach to blood glucose level prediction aimed at “what-if” scenarios, in which people could enter foods they might eat or insulin amounts they might take and then see the effect on future BGLs. In this work, we invert the “what-if” scenario and introduce a similar architecture based on chaining two LSTMs that can be trained to make either insulin or carbohydrate recommendations aimed at reaching a desired BG level in the future. Leveraging a recent state-of-the-art model for time series forecasting, we then derive a novel architecture for the same recommendation task, in which the two LSTM chain is used as a repeating block inside a deep residual architecture. Experimental evaluations using real patient data from the OhioT1DM dataset show that the new integrated architecture compares favorably with the previous LSTM-based approach, substantially outperforming the baselines. The promising results suggest that this novel approach could potentially be of practical use to people with type 1 diabetes for self-management of BGLs.

Convolutional neural networks (CNNs) are at the core of many state-of-the-art deep learning models in computer vision, speech, and text processing. Training and deploying such CNN-based architectures usually require a significant amount of computational resources. Sparsity has emerged as an effective compression approach for reducing the amount of data and computation for CNNs. However, sparsity often results in computational irregularity, which prevents accelerators from fully taking advantage of its benefits for performance and energy improvement. In this paper, we propose CSCNN, an algorithm/hardware co-design framework for CNN compression and acceleration that mitigates the effects of computational irregularity and provides better performance and energy efficiency. On the algorithmic side, CSCNN uses centrosymmetric matrices as convolutional filters. In doing so, it reduces the number of required weights by nearly 50% and enables structured computational reuse without compromising regularity and accuracy. Additionally, complementary pruning techniques are leveraged to further reduce computation by a factor of $2.8-7.2\times $ with a marginal accuracy loss. On the hardware side, we propose a CSCNN accelerator that effectively exploits the structured computational reuse enabled by centrosymmetric filters, and further eliminates zero computations for increased performance and energy efficiency. Compared against a dense accelerator, SCNN and SparTen, the proposed accelerator performs $3.7\times $, $1.6\times $ and $1.3\times $ better, and improves the EDP (Energy Delay Product) by $8.9\times $, $2.8\times $ and $2.0\times $, respectively.

Graph convolutional neural networks (GCNs) have emerged as an effective approach to extend deep learning for graph data analytics. Given that graphs are usually irregular, as nodes in a graph may have a varying number of neighbors, processing GCNs efficiently pose a significant challenge on the underlying hardware. Although specialized GCN accelerators have been proposed to deliver better performance over generic processors, prior accelerators not only under-utilize the compute engine, but also impose redundant data accesses that reduce throughput and energy efficiency. Therefore, optimizing the overall flow of data between compute engines and memory, i.e., the GCN dataflow, which maximizes utilization and minimizes data movement is crucial for achieving efficient GCN processing.In this paper, we propose a flexible and optimized dataflow for GCNs that simultaneously improves resource utilization and reduces data movement. This is realized by fully exploring the design space of GCN dataflows and evaluating the number of execution cycles and DRAM accesses through an analysis framework. Unlike prior GCN dataflows, which employ rigid loop orders and loop fusion strategies, the proposed dataflow can reconFigure the loop order and loop fusion strategy to adapt to different GCN configurations, which results in much improved efficiency. We then introduce a novel accelerator architecture called GCNAX, which tailors the compute engine, buffer structure and size based on the proposed dataflow. Evaluated on five real-world graph datasets, our simulation results show that GCNAX reduces DRAM accesses by a factor of $8.1 \times$ and $2.4 \times$, while achieving $8.9 \times, 1.6 \times$ speedup and $9.5 \times$, $2.3 \times$ energy savings on average over HyGCN and AWB-GCN, respectively.

To avoid serious diabetic complications, people with type 1 diabetes must keep their blood glucose levels (BGLs) as close to normal as possible. Insulin dosages and carbohydrate consumption are important considerations in managing BGLs. Since the 1960s, models have been developed to forecast blood glucose levels based on the history of BGLs, insulin dosages, carbohydrate intake, and other physiological and lifestyle factors. Such predictions can be used to alert people to impending unsafe BGLs or to control insulin flow in an artificial pancreas. In past work, we have introduced an LSTM-based approach to blood glucose level prediction aimed at “what if” scenarios, in which people could input the grams of carbohydrates in foods they might eat or units of insulin they might take and then see the effect on future BGLs. Building on these neural models for “what-if” predictions, in this work we derive a novel LSTM-based architecture that can be trained to make either insulin or carbohydrate recommendations to ensure that future BGLs attain the desired level. Experimental evaluations using data from the OhioT1DM dataset show that the neural architecture substantially outperforms the baselines. The promising results suggest that this novel approach could potentially be of practical use to people with type 1 diabetes for self-management of BGLs.

We investigate the large-scale clustering of the final spectroscopic sample of quasars from the recently completed extended Baryon Oscillation Spectroscopic Survey (eBOSS). The sample contains $343708$ objects in the redshift range $0.8, Comment: 17 pages, 13 figures, 2 tables. Accepted for publication in MNRAS. For the associated code and value-added catalogs see https://github.com/mehdirezaie/sysnetdev and https://github.com/mehdirezaie/eBOSSDR16QSOE

We introduce the task of changing the narrative point of view, where characters are assigned a narrative perspective that is different from the one originally used by the writer. The resulting shift in the narrative point of view alters the reading experience and can be used as a tool in fiction writing or to generate types of text ranging from educational to self-help and self-diagnosis. We introduce a benchmark dataset containing a wide range of types of narratives annotated with changes in point of view from deictic (first or second person) to anaphoric (third person) and describe a pipeline for processing raw text that relies on a neural architecture for mention selection. Evaluations on the new benchmark dataset show that the proposed architecture substantially outperforms the baselines by generating mentions that are less ambiguous and more natural., Comment: To appear in Information Processing & Management, Special Issue on Creative Language Processing

We propose two optical hybrid matrix multipliers for deep neural networks. Our results indicate our all-optical design achieved the best performance in energy efficiency and latency, with an energy-delay product reduction of 33.1% and 76.4% for conservative and aggressive estimates, respectively.

Network-on-Chips (NoCs) are the de facto choice for designing the interconnect fabric in multicore chips due to their regularity, efficiency, simplicity, and scalability. However, NoC suffers from excessive static power and dynamic energy due to transistor leakage current and data movement between the cores and caches. Power consumption issues are only exacerbated by ever decreasing technology sizes. Dynamic Voltage and Frequency Scaling (DVFS) is one technique that seeks to reduce dynamic energy; however this often occurs at the expense of performance. In this paper, we propose LEAD Learning-enabled Energy-Aware Dynamic voltage/frequency scaling for multicore architectures using both supervised learning and reinforcement learning approaches. LEAD groups the router and its outgoing links into the same V/F domain and implements proactive DVFS mode management strategies that rely on offline trained machine learning models in order to provide optimal V/F mode selection between different voltage/frequency pairs. We present three supervised learning versions of LEAD that are based on buffer utilization, change in buffer utilization and change in energy/throughput, which allow proactive mode selection based on accurate prediction of future network parameters. We then describe a reinforcement learning approach to LEAD that optimizes the DVFS mode selection directly, obviating the need for label and threshold engineering. Simulation results using PARSEC and Splash-2 benchmarks on a 4 × 4 concentrated mesh architecture show that by using supervised learning LEAD can achieve an average dynamic energy savings of 15.4 percent for a loss in throughput of 0.8 percent with no significant impact on latency. When reinforcement learning is used, LEAD increases average dynamic energy savings to 20.3 percent at the cost of a 1.5 percent decrease in throughput and a 1.7 percent increase in latency. Overall, the more flexible reinforcement learning approach enables learning an optimal behavior for a wider range of load environments under any desired energy versus throughput tradeoff.

Development of truly useful wearable physiologic monitoring devices for use in diabetes management is still in its infancy. From wearable activity monitors such as fitness trackers and smart watches to contact lenses measuring glucose levels in tears, we are just at the threshold of their coming use in medicine. Ultimately, such devices could help to improve the performance of sense-and-respond insulin pumps, illuminate the impact of physical activity on blood glucose levels, and improve patient safety. This is a summary of our experience attempting to use such devices to enhance continuous glucose monitoring–augmented insulin pump therapy. We discuss the current status and present difficulties with available devices, and review the potential for future use.

Figures are human-friendly but difficult for computers to process automatically. In this work, we investigate the problem of figure captioning. The goal is to automatically generate a natural language description of a given figure. We create a new dataset for figure captioning, FigCAP. To achieve accurate generation of labels in figures, we propose the Label Maps Attention Model. Extensive experiments show that our method outperforms the baselines. A successful solution to this task allows figure content to be accessible to those with visual impairment by providing input to a text-to-speech system; and enables automatic parsing of vast repositories of documents where figures are pervasive.

We have shown in previous work that LSTM networks are effective at predicting blood glucose levels in patients with type I diabetes, outperforming human experts and an SVR model trained with features computed by manually engineered physiological models. In this paper we present the results of a much larger set of experiments on real and synthetic datasets in what-if, agnostic, and inertial scenarios. Experiments on a more recent real-patient dataset, which we are releasing to the research community, demonstrate that LSTMs are robust to noise and can easily incorporate additional features, such as skin temperature, heart rate and skin conductance, without any change in the architecture. A neural attention module that we designed specifically for time series prediction improves prediction performance on synthetic data; however, the improvements do not transfer to real data. Conversely, using time of day as an additional input feature consistently improves the LSTM performance on real data but not on synthetic data. These and other differences show that behavior on synthetic data cannot be assumed to always transfer to real data, highlighting the importance of evaluating physiological models on data from real patients.

As technology scales, Network-on-Chips (NoCs), currently being used for on-chip communication in manycore architectures, face several problems including high network latency, excessive power consumption, and low reliability. Simultaneously addressing these problems is proving to be difficult due to the explosion of the design space and the complexity of handling many trade-offs. In this paper, we propose IntelliNoC, an intelligent NoC design framework which introduces architectural innovations and uses reinforcement learning to manage the design complexity and simultaneously optimize performance, energy-efficiency, and reliability in a holistic manner. IntelliNoC integrates three NoC architectural techniques: (1) multifunction adaptive channels (MFACs) to improve energy-efficiency; (2) adaptive error detection/correction and re-transmission control to enhance reliability; and (3) a stress-relaxing bypass feature which dynamically powers off NoC components to prevent overheating and fatigue. To handle the complex dynamic interactions induced by these techniques, we train a dynamic control policy using Q-learning, with the goal of providing improved fault-tolerance and performance while reducing power consumption and area overhead. Simulation using PARSEC benchmarks shows that our proposed IntelliNoC design improves energy-efficiency by 67% and mean-time-to-failure (MTTF) by 77%, and decreases end-to-end packet latency by 32% and area requirements by 25% over baseline NoC architecture.

When a new bug report is received, developers usually need to reproduce the bug and perform code reviews to find the cause, a process that can be tedious and time consuming. A tool for ranking all the source files with respect to how likely they are to contain the cause of the bug would enable developers to narrow down their search and improve productivity. This paper introduces an adaptive ranking approach that leverages project knowledge through functional decomposition of source code, API descriptions of library components, the bug-fixing history, the code change history, and the file dependency graph. Given a bug report, the ranking score of each source file is computed as a weighted combination of an array of features, where the weights are trained automatically on previously solved bug reports using a learning-to-rank technique. We evaluate the ranking system on six large scale open source Java projects, using the before-fix version of the project for every bug report. The experimental results show that the learning-to-rank approach outperforms three recent state-of-the-art methods. In particular, our method makes correct recommendations within the top 10 ranked source files for over 70 percent of the bug reports in the Eclipse Platform and Tomcat projects.

Network-on-Chips (NoCs) are becoming the standard communication fabric for multi-core and system on a chip (SoC) architectures. As technology continues to scale, transistors and wires on the chip are becoming increasingly vulnerable to various fault mechanisms, especially timing errors, resulting in exacerbation of energy efficiency and performance for NoCs. Typical techniques for handling timing errors are reactive in nature, responding to the faults after their occurrence. They rely on error detection/correction techniques which have resulted in excessive power consumption and degraded performance, since the error detection/correction hardware is constantly enabled. On the other hand, indiscriminately disabling error handling hardware can induce more errors and intrusive retransmission traffic. Therefore, the challenge is to balance the trade-offs among error rate, packet retransmission, performance, and energy. In this paper, we propose a proactive fault-tolerant mechanism to optimize energy efficiency and performance with reinforcement learning (RL). First, we propose a new proactive error handling technique comprised of a dynamic scheme for enabling per-router error detection/correction hardware and an effective retransmission mechanism. Second, we propose the use of RL to train the dynamic control policy with the goals of providing increased fault-tolerance, reduced power consumption and improved performance as compared to conventional techniques. Our evaluation indicates that, on average, end-to-end packet latency is lowered by 55%, energy efficiency is improved by 64%, and retransmission caused by faults is reduced by 48% over the reactive error correction techniques.

We introduce a new question answering paradigm in which users can interact with the system using natural language questions or direct actions within a graphical user interface (GUI). The system displays multiple time series characterizing the behavior of a patient, and a physician interacts with the system through GUI actions and questions, where answers may depend on previous interactions. To find the answers automatically, we propose parsing the questions into logical forms for execution by an inference engine over the underlying database. The semantic parser is implemented as an LSTM-based encoder-decoder that models dependencies between consecutive answers through multiple attention and copying mechanisms. To train and evaluate the model, we created a dataset of semantic parses of real interactions with the system, augmented with a larger dataset of artificial interactions. The proposed architecture obtains promising results, substantially outperforming standard sequence generation baselines.

Robust measurements of cosmological parameters from galaxy surveys rely on our understanding of systematic effects that impact the observed galaxy density field. In this paper we present, validate, and implement the idea of adopting the systematics mitigation method of Artificial Neural Networks for modeling the relationship between the target galaxy density field and various observational realities including but not limited to Galactic extinction, seeing, and stellar density. Our method by construction allows a wide class of models and alleviates over-training by performing k-fold cross-validation and dimensionality reduction via backward feature elimination. By permuting the choice of the training, validation, and test sets, we construct a selection mask for the entire footprint. We apply our method on the extended Baryon Oscillation Spectroscopic Survey (eBOSS) Emission Line Galaxies (ELGs) selection from the Dark Energy Camera Legacy Survey (DECaLS) Data Release 7 and show that the spurious large-scale contamination due to imaging systematics can be significantly reduced by up-weighting the observed galaxy density using the selection mask from the neural network and that our method is more effective than the conventional linear and quadratic polynomial functions. We perform extensive analyses on simulated mock datasets with and without systematic effects. Our analyses indicate that our methodology is more robust to overfitting compared to the conventional methods. This method can be utilized in the catalog generation of future spectroscopic galaxy surveys such as eBOSS and Dark Energy Spectroscopic Instrument (DESI) to better mitigate observational systematics., Comment: 31 pages, 25 figures, accepted for publication in MNRAS. Moderate revision throughout the paper. The new version includes a quantitative evaluation of the remaining systematic effects in the DECaLS DR7 data in Summary and Discussion and in Conclusion. Pipeline available at https://github.com/mehdirezaie/SYSNet

Recently many information retrieval (IR)-based approaches have been proposed to help locate software defects automatically by using information from bug report contents. However, some bug reports that do not semantically related to the relevant code are not helpful to IR-based systems. Running an IR-based system on these reports may produce false positives. In this paper, we propose a classification model for classifying a bug report as either helpful or unhelpful using a LSTM-network. By filtering our unhelpful reports before running an IR-based bug locating system, our approach helps reduce false positives and improve the ranking performance. We test our model over 9,000 bug reports from three software projects. The evaluation result shows that our model helps improve a state-of-the-art IR-based system's ranking performance under a trade-off between the precision and the recall. Our comparison experiments show that the LSTM-network achieves the best trade-off between precision and recall than other classification models including CNN, multilayer perceptron, and a simple baseline approach that classifies a bug report based its length. In the situation that precision is more important than recall, our classification model helps for bug locating.

The rapid growth of mobile devices has resulted in the generation of a large number of user behavior logs that contain latent intentions and user interests. However, exploiting such data in real-world applications is still difficult for service providers due to the complexities of user behavior over a sheer number of possible actions that can vary according to time. In this work, a time-aware RNN model, TRNN, is proposed for predictive analysis from user behavior data. First, our approach predicts next user action more accurately than the baselines including the n-gram models as well as two recently introduced time-aware RNN approaches. Second, we use TRNN to learn user embeddings from sequences of user actions and show that overall the TRNN embeddings outperform conventional RNN embeddings. Similar to how word embeddings benefit a wide range of task in natural language processing, the learned user embeddings are general and could be used in a variety of tasks in the digital marketing area. This claim is supported empirically by evaluating their utility in user conversion prediction, and preferred application prediction. According to the evaluation results, TRNN embeddings perform better than the baselines including Bag of Words (BoW), TFIDF and Doc2Vec. We believe that TRNN embeddings provide an effective representation for solving practical tasks such as recommendation, user segmentation and predictive analysis of business metrics.

Motivation De novo motif discovery algorithms find statistically over-represented sequence motifs that may function as transcription factor binding sites. Current methods often report large numbers of motifs, making it difficult to perform further analyses and experimental validation. The motif selection problem seeks to identify a minimal set of putative regulatory motifs that characterize sequences of interest (e.g. ChIP-Seq binding regions). Results In this study, the motif selection problem is mapped to variants of the set cover problem that are solved via tabu search and by relaxed integer linear programing (RILP). The algorithms are employed to analyze 349 ChIP-Seq experiments from the ENCODE project, yielding a small number of high-quality motifs that represent putative binding sites of primary factors and cofactors. Specifically, when compared with the motifs reported by Kheradpour and Kellis, the set cover-based algorithms produced motif sets covering 35% more peaks for 11 TFs and identified 4 more putative cofactors for 6 TFs. Moreover, a systematic evaluation using nested cross-validation revealed that the RILP algorithm selected fewer motifs and was able to cover 6% more peaks and 3% fewer background regions, which reduced the error rate by 7%. Availability and implementation The source code of the algorithms and all the datasets are available at https://github.com/YichaoOU/Set_cover_tools. Supplementary information Supplementary data are available at Bioinformatics online.

Network on Chips (NoCs) are the interconnect fabric of choice for multicore processors due to their superiority over traditional buses and crossbars in terms of scalability. While NoC’s offer several advantages, they still suffer from high static and dynamic power consumption. Dynamic Voltage and Frequency Scaling (DVFS) is a popular technique that allows dynamic energy to be saved, but it can potentially lead to loss in throughput. In this paper, we propose LEAD - Learning-enabled Energy-Aware Dynamic voltage/frequency scaling for NoC architectures wherein we use machine learning techniques to enable energy-performance trade-offs at reduced overhead cost. LEAD enables a proactive energy management strategy that relies on an offline trained regression model and provides a wide variety of voltage/frequency pairs (modes). LEAD groups each router and the router’s outgoing links locally into the same V/F domain, allowing energy management at a finer granularity without additional timing complications and overhead. Our simulation results using PARSEC and Splash-2 benchmarks on a 4 × 4 concentrated mesh architecture show an average dynamic energy savings of 17% with a minimal loss of 4% in throughput and no latency increase.

Network-on-Chips (NoCs) are quickly becoming the standard communication paradigm for the growing number of cores on the chip. While NoCs can deliver sufficient bandwidth and enhance scalability, NoCs suffer from high power consumption due to the router microarchitecture and communication channels that facilitate inter-core communication. As technology keeps scaling down in the nanometer regime, unpredictable device behavior due to aging, infant mortality, design defects, soft errors, aggressive design, and process-voltage-temperature variations, will increase and will result in a significant increase in faults (both permanent and transient) and hardware failures. In this paper, we propose QORE —a fault tolerant NoC architecture with Multi-Function Channel (MFC) buffers. The use of MFC buffers and their associated control (link and fault controllers) enhance fault-tolerance by allowing the NoC to dynamically adapt to faults at the link level and reverse propagation direction to avoid faulty links. Additionally, MFC buffers reduce router power and improve performance by eliminating in-router buffering. We utilize a machine learning technique in our link controllers to predict the direction of traffic flow in order to more efficiently reverse links. Our simulation results using real benchmarks and synthetic traffic mixes show that QORE improves speedup by 1.3 $\times$ and throughput by 2.3 $\times$ when compared to state-of-the art fault tolerant NoCs designs such as Ariadne and Vicis. Moreover, using Synopsys Design Compiler, we also show that network power in QORE is reduced by 21 percent with minimal control overhead.

As communication energy exceeds computation energy in future technologies, traditional on-chip electrical interconnects face fundamental challenges in the many-core era. Photonic interconnects have been proposed as a disruptive technology solution due to superior performance per Watt, distance independent energy consumption and CMOS compatibility for on-chip interconnects. Static power due to the laser being always switched on, varying link utilization due to spatial and temporal traffic fluctuations and thermal sensitivity are some of the critical challenges facing photonics interconnects. In this paper, we propose photonic interconnects for heterogeneous multicores using a checkerboard pattern that clusters CPU-GPU cores together and implements bandwidth reconfiguration using local router information without global coordination. To reduce the static power, we also propose a dynamic laser scaling technique that predicts the power level for the next epoch using the buffer occupancy of previous epoch. To further improve power-performance trade-offs, we also propose a regression-based machine learning technique for scaling the power of the photonic link. Our simulation results demonstrate a 34% performance improvement over a baseline electrical CMESH while consuming 25% less energy per bit when dynamically reallocating bandwidth. When dynamically scaling laser power, our buffer-based reactive and ML-based proactive prediction techniques show 40 - 65% in power savings with 0 - 14% in throughput loss depending on the reservation window size.

Understanding morphological types of galaxies is a key parameter for studying their formation and evolution. Neural networks that have been used previously for galaxy morphology classification have some disadvantages, such as not being invariant under rotation. In this work, we studied the performance of Capsule Network, a recently introduced neural network architecture that is rotationally invariant and spatially aware, on the task of galaxy morphology classification. We designed two evaluation scenarios based on the answers from the question tree in the Galaxy Zoo project. In the first scenario, we used Capsule Network for regression and predicted probabilities for all of the questions. In the second scenario, we chose the answer to the first morphology question that had the highest user agreement as the class of the object and trained a Capsule Network classifier, where we also reconstructed galaxy images. We achieved promising results in both of these scenarios. Automated approaches such as the one introduced here will greatly decrease the workload of astronomers and will play a critical role in the upcoming large sky surveys., Comment: 9 pages, 6 figures

Hand gestures provide a natural, non-verbal form of communication that can augment or replace other communication modalities such as speech or writing. Along with voice commands, hand gestures are becoming the primary means of interaction in games, augmented reality, and virtual reality platforms. Recognition accuracy, flexibility, and computational cost are some of the primary factors that can impact the incorporation of hand gestures in these new technologies, as well as their subsequent retrieval from multimodal corpora. In this paper, we present fast and highly accurate gesture recognition systems based on long short-term memory (LSTM) and convolutional neural networks (CNN) that are trained to process input sequences of 3D hand positions and velocities acquired from infrared sensors. When evaluated on real time recognition of six types of hand gestures, the proposed architectures obtain 97% F-measure, demonstrating a significant potential for practical applications in novel human-computer interfaces.

For people with type 1 diabetes, good blood glucose control is essential to keeping serious disease complications at bay. This entails carefully monitoring blood glucose levels and taking corrective steps whenever they are too high or too low. If blood glucose levels could be accurately predicted, patients could take proactive steps to prevent blood glucose excursions from occurring. However, accurate predictions require complex physiological models of blood glucose behavior. Factors such as insulin boluses, carbohydrate intake, and exercise influence blood glucose in ways that are difficult to capture through manually engineered equations. In this paper, we describe a recursive neural network (RNN) approach that uses long short-term memory (LSTM) units to learn a physiological model of blood glucose. When trained on raw data from real patients, the LSTM networks (LSTMs) obtain results that are competitive with a previous state-of-the-art model based on manually engineered physiological equations. The RNN approach can incorporate arbitrary physiological parameters without the need for sophisticated manual engineering, thus holding the promise of further improvements in prediction accuracy.

The application of information retrieval techniques to search tasks in software engineering is made difficult by the lexical gap between search queries, usually expressed in natural language (e.g. English), and retrieved documents, usually expressed in code (e.g. programming languages). This is often the case in bug and feature location, community question answering, or more generally the communication between technical personnel and non-technical stake holders in a software project. In this paper, we propose bridging the lexical gap by projecting natural language statements and code snippets as meaning vectors in a shared representation space. In the proposed architecture, word embeddings are first trained on API documents, tutorials, and reference documents, and then aggregated in order to estimate semantic similarities between documents. Empirical evaluations show that the learned vector space embeddings lead to improvements in a previously explored bug localization task and a newly defined task of linking API documents to computer programming questions.

An important approach to text mining involves the use of natural-language information extraction. Information extraction (IE) distills structured data or knowledge from unstructured text by identifying references to named entities as well as stated relationships between such entities. IE systems can be used to directly extricate abstract knowledge from a text corpus, or to extract concrete data from a set of documents which can then be further analyzed with traditional data-mining techniques to discover more general patterns. We discuss methods and implemented systems for both of these approaches and summarize results on mining real text corpora of biomedical abstracts, job announcements, and product descriptions. We also discuss challenges that arise when employing current information extraction technology to discover knowledge in text.

Objective:: Automatically extracting information from biomedical text holds the promise of easily consolidating large amounts of biological knowledge in computer-accessible form. This strategy is particularly attractive for extracting data relevant to genes of the human genome from the 11 million abstracts in Medline. However, extraction efforts have been frustrated by the lack of conventions for describing human genes and proteins. We have developed and evaluated a variety of learned information extraction systems for identifying human protein names in Medline abstracts and subsequently extracting information on interactions between the proteins. Methods and Material:: We used a variety of machine learning methods to automatically develop information extraction systems for extracting information on gene/protein name, function and interactions from Medline abstracts. We present cross-validated results on identifying human proteins and their interactions by training and testing on a set of approximately 1000 manually-annotated Medline abstracts that discuss human genes/proteins. Results:: We demonstrate that machine learning approaches using support vector machines and maximum entropy are able to identify human proteins with higher accuracy than several previous approaches. We also demonstrate that various rule induction methods are able to identify protein interactions with higher precision than manually-developed rules. Conclusion:: Our results show that it is promising to use machine learning to automatically build systems for extracting information from biomedical text. The results also give a broad picture of the relative strengths of a wide variety of methods when tested on a reasonably large human-annotated corpus.

When a new bug report is received, developers usually need to reproduce the bug and perform code reviews to find the cause, a process that can be tedious and time consuming. A tool for ranking all the source files of a project with respect to how likely they are to contain the cause of the bug would enable developers to narrow down their search and potentially could lead to a substantial increase in productivity. This paper introduces an adaptive ranking approach that leverages domain knowledge through functional decompositions of source code files into methods, API descriptions of library components used in the code, the bug-fixing history, and the code change history. Given a bug report, the ranking score of each source file is computed as a weighted combination of an array of features encoding domain knowledge, where the weights are trained automatically on previously solved bug reports using a learning-to-rank technique. We evaluated our system on six large scale open source Java projects, using the before-fix version of the project for every bug report. The experimental results show that the newly introduced learning-to-rank approach significantly outperforms two recent state-of-the-art methods in recommending relevant files for bug reports. In particular, our method makes correct recommendations within the top 10 ranked source files for over 70% of the bug reports in the Eclipse Platform and Tomcat projects.

Patients with diabetes must continually monitor their blood glucose levels and adjust insulin doses, striving to keep blood glucose levels as close to normal as possible. Blood glucose levels that deviate from the normal range can lead to serious short-term and long-term complications. An automatic prediction model that warned people of imminent changes in their blood glucose levels would enable them to take preventive action. Modeling inter-patient differences and the combined effects of insulin and life events on blood glucose have been particularly challenging in the design of accurate blood glucose forecasting systems. In this paper, we describe a solution that uses a generic physiological model of blood glucose dynamics to generate informative features for a Support Vector Regression model that is trained on patient specific data. Experimental results show that the new prediction model outperforms all three diabetes experts involved in the study, thus demonstrating the utility of using the generic physiological features in machine learning models that are individually trained for every patient.

Objective: Glycemic variability (GV) is an important component of overall glycemic control for patients with diabetes mellitus. Physicians are able to recognize excessive GV from continuous glucose monitoring (CGM) plots; however, there is currently no universally agreed upon GV metric. The objective of this study was to develop a consensus perceived glycemic variability (CPGV) metric that could be routinely applied to CGM data to assess diabetes mellitus control. Methods: Twelve physicians actively managing patients with type 1 diabetes mellitus rated a total of 250 24 h CGM plots as exhibiting low, borderline, high, or extremely high GV. Ratings were averaged to obtain a consensus and then input into two machine learning algorithms: multilayer perceptrons (MPs) and support vector machines for regression (SVR). In silica experiments were run using each algorithm with different combinations of 12 descriptive input features. Ten-fold cross validation was used to evaluate the performance of each model. Results: The SVR models approximated the physician consensus ratings of unseen CGM plots better than the MP models. When judged by the root mean square error, the best SVR model performed comparably to individual physicians at matching consensus ratings. When applied to 262 different CGM plots as a screen for excessive GV, this model had accuracy, sensitivity, and specificity of 90.1%, 97.0%, and 74.1%, respectively. It significantly outperformed mean amplitude of glycemic excursion, standard deviation, distance traveled, and excursion frequency. Conclusions: This new CPGV metric could be used as a routine measure of overall glucose control to supplement glycosylated hemoglobin in clinical practice.

This paper describes explorations in word sense disambiguation using Wikipedia as a source of sense annotations. Through experiments on four different languages, we show that the Wikipedia-based sense annotations are reliable and can be used to construct accurate sense classifiers.

In tone languages, such as Mandarin Chinese, a syllable with different tones conveys different meanings. Previously, Mandarin tone recognition based on Mel-frequency cepstral coefficients (MFCCs) and Convolutional Neural Networks (CNN) was examined and the results outperformed the model of conventional neural network using manually edited F0 data. In the present study, Mandarin tone recognition based on spectrograms, instead of MFCCs, was explored. Unsupervised feature learning was applied to the unlabeled spectrograms directly with a denoising autoencoder (dAE). Then, the model convolved the labeled spectrograms with the learnt “sound features” and produced a set of feature maps. A dataset that consisted of 4500 monosyllabic words collected from 125 children was used to evaluate the recognition performance. Compared with methods based on MFCCs, there are more parameters to train in the new approach based on spectrograms. As a result, the new model might better capture the statistical distribution in the ...

Glycemic variability, or fluctuation in blood glucose levels, is a significant factor in diabetes management. Excessive glycemic variability contributes to oxidative stress, which has been linked to the development of long-term diabetic complications. An automated screen for excessive glycemic variability, based on the readings from continuous glucose monitoring (CGM) systems, would enable early identification of at risk patients. In this paper, we present an automatic approach for learning variability models that can routinely detect excessive glycemic variability when applied to CGM data. Naive Bayes (NB), Multilayer Perceptron (MP), and Support Vector Machine (SVM) models are trained and evaluated on a dataset of CGM plots that have been manually annotated with respect to glycemic variability by two diabetes experts. In order to alleviate the impact of noise, the CGM plots are smoothed using cubic splines. Automatic feature selection is then performed on a rich set of pattern recognition features. Empirical evaluation shows that the top performing model obtains a state of the art accuracy of 93.8%, substantially outperforming a previous NB model.

This paper presents the 4 Diabetes Support SystemTM (4DSS) project as a case study in case-based reasoning (CBR) research and development. This project aims to help patients with type 1 diabetes on insulin pump therapy achieve and maintain good blood glucose control. Over the course of seven years and three clinical research studies, a series of defining cases altered the research and development path. Each of these cases suggested a new, unanticipated research direction or clinical application. New AI technologies, including naive Bayes classification and support vector regression, were incorporated. New medical research into glycemic variability and blood glucose prediction was undertaken. The CBR research paradigm has provided a strong framework for medical research as well as for artificial intelligence (AI) research. This new work has the potential to positively impact the health and wellbeing of patients with diabetes. This paper shares the 4DSS project experience.

We present a novel learning framework for pipeline models aimed at improving the communication between consecutive stages in a pipeline. Our method exploits the confidence scores associated with outputs at any given stage in a pipeline in order to compute probabilistic features used at other stages downstream. We describe a simple method of integrating probabilistic features into the linear scoring functions used by state of the art machine learning algorithms. Experimental evaluation on dependency parsing and named entity recognition demonstrate the superiority of our approach over the baseline pipeline models, especially when upstream stages in the pipeline exhibit low accuracy.

We present a new approach to multiple instance learning (MIL) that is particularly effective when the positive bags are sparse (i.e. contain few positive instances). Unlike other SVM-based MIL methods, our approach more directly enforces the desired constraint that at least one of the instances in a positive bag is positive. Using both artificial and real-world data, we experimentally demonstrate that our approach achieves greater accuracy than state-of-the-art MIL methods when positive bags are sparse, and performs competitively when they are not. In particular, our approach is the best performing method for image region classification.

Extracting semantic relationships between entities mentioned in text documents is an important task in natural language processing. The various types of relationships that are discovered between mentions of entities can provide useful structured information to a text mining system [1]. Traditionally, the task specifies a predefined set of entity types and relation types that are deemed to be relevant to a potential user and that are likely to occur in a particular text collection. For example, information extraction from newspaper articles is usually concerned with identifying mentions of people, organizations, locations, and extracting useful relations between them. Relevant relation types range from social relationships, to roles that people hold inside an organization, to relations between organizations, to physical locations of people and organizations. Scientific publications in the biomedical domain offer a type of narrative that is very different from the newspaper discourse. A significant effort is currently spent on automatically extracting relevant pieces of information from Medline, an online collection of biomedical abstracts. Proteins, genes, and cells are examples of relevant entities in this task, whereas subcellular localizations and protein-protein interactions are two of the relation types that have received significant attention recently. The inherent difficulty of the relation extraction task is further compounded in the biomedical domain by the relative scarcity of tools able to analyze the corresponding type of narrative. Most existing natural language processing tools, such as tokenizers, sentence segmenters, part-of-speech (POS) taggers, shallow or full parsers are trained on newspaper corpora, and consequently they inccur a loss in accuracy when applied to biomedical literature. Therefore, information extraction systems developed for biological corpora need to be robust to POS or parsing errors, or to give reasonable performance using shallower but more reliable information, such as chunking instead of full parsing.

This paper presents the results of a large-scale effort to construct a comprehensive database of known human protein interactions by combining and linking known interactions from existing databases and then adding to them by automatically mining additional interactions from 750,000 Medline abstracts. The end result is a network of 31,609 interactions amongst 7,748 proteins. The text mining system first identifies protein names in the text using a trained Conditional Random Field (CRF) and then identifies interactions through a filtered co-citation analysis. We also report two new strategies for mining interactions, either by finding explicit statements of interactions in the text using learned pattern-based rules or a Support-Vector Machine using a string kernel. Using information in existing ontologies, the automatically extracted data is shown to be of equivalent accuracy to manually curated data sets.

We present a novel approach to relation extraction, based on the observation that the information required to assert a relationship between two named entities in the same sentence is typically captured by the shortest path between the two entities in the dependency graph. Experiments on extracting top-level relations from the ACE (Automated Content Extraction) newspaper corpus show that the new shortest path dependency kernel outperforms a recent approach based on dependency tree kernels.

We present sentiment analyzer (SA) that extracts sentiment (or opinion) about a subject from online text documents. Instead of classifying the sentiment of an entire document about a subject, SA detects all references to the given subject, and determines sentiment in each of the references using natural language processing (NLP) techniques. Our sentiment analysis consists of 1) a topic specific feature term extraction, 2) sentiment extraction, and 3) (subject, sentiment) association by relationship analysis. SA utilizes two linguistic resources for the analysis: the sentiment lexicon and the sentiment pattern database. The performance of the algorithms was verified on online product review articles ("digital camera" and "music" reviews), and more general documents including general Webpages and news articles.