Members of the Western Michigan Transformative Interdisciplinary Human+AI Research Group have been engaged in two consecutive NSF-funded projects to promote AI readiness in diverse STEM disciplines. Putting equal emphasis on theory and practice, our goal is to instill knowledge and competency in safe, secure, and reliable AI across a wide range of learners from high school students through to university students and practitioners who wish to upskill. The second project that is currently underway has a specific focus on machine-assisted processing of massive data. This presentation focuses on the development of immersive learning experiences.

Summary Target‐Decoy database is currently the method of choice to assess the quality of Proteins' search engines. Decoy versions of real peptides are generated and injected to the same database of real ones with different labels. Quality of search engines results is assessed based on the number of decoys retrieved as hits. In Crux‐Tide search engine, which is one of the fastest search engines currently available, the process of indexing and generating decoys is computationally expensive. In this paper, we analyze the serial algorithm in detail and show improvement possibilities, and then describe a parallel‐shared memory solution using OpenMP. To completely break up the dependency in the serial algorithms, a clever hashing technique is utilized to localize the process. The parallel solution and the hashing technique together are able to reduce the computation cost by approximately 70‐80% using few threads. Besides the parallelization, we redesign part of the serial code so that the memory consumption becomes more efficient. The parallel version can index the same files using around two‐third of the memory space that the serial version consumes. This solution could impact and support future distributed developments of Crux‐Tide searching phase, where each parallel unit could rank the observed spectra independently.

several peptides search engines have been developed in the recent decades. Most of the time and for the same inputs, different search engines’ result in different peptides were identified, which can confuse the stakeholders in the field of proteomics. The massive amount of generated spectra by high throughput spectrometers adds another challenge which handicaps the current search engines. This motivates the researchers to evaluate the combination of several search engines. Several studies provided ensemble solutions over shared and distributed computing environments for reliable results. However, the massive amount of MS/MS spectra is a cumbersome traffic over the systems’ networks. This issue directly impacts the searching performance and also adds unnecessary extra costs (computing, storage, network traffic) if cloud cluster is being used. The main question of this paper is: Can we build a central MS/MS spectra preprocessing for semantically different protein search engines? We evaluate different statistical reduction techniques using four popular protein search engines. In order to fairly evaluate the results, we build ground truth unanimous-based datasets for two different species; yeast and human. Our techniques result in significant peak reduction, where only around 30% of the spectra peaks are enough to report reliable identifications from the used search engines in this study.

The results of protein search engines depend mainly upon a set of parameters that adjust the searching space. One of the most effective parameters is the peptide mass window tolerance (w). Most of the current search engines use a constant user-defined value for this parameter. As an alternative option, Comet search engine designers proposed a statistical technique to estimate the best tolerance window for an input spectra file. However, this technique sometimes fails in picking a value, may set the parameter to a value that results in a loss of many correct matches, and is available only for one type of mass; namely ppm. In this paper, we propose to use particle swarm optimization (PSO) to improve the coverage of search engines by picking the optimal value for this influential parameter to maximize PSMs. Our results show that this biologically-inspired algorithm can be utilized to find peptide mass window tolerance values that facilitate Comet to increase peptide spectra matches, resulting in improved peptide identification. We also show experimental evidence that an open search (i.e., wide tolerance window) does not always optimize spectra matching using the current search engines and that narrow tolerance windows improve the coverage of protein search engines.

The diversity of the available protein search engines with respect to the utilized matching algorithms, the low overlap ratios among their results and the disparity of their coverage encourage the community of proteomics to utilize ensemble solutions of different search engines. The advancing in cloud computing technology and the availability of distributed processing clusters can also provide support to this task. However, data transferring and results' combining, in this case, could be the major bottleneck. The flood of billions of observed mass spectra, hundreds of Gigabytes or potentially Terabytes of data, could easily cause the congestions, increase the risk of failure, poor performance, add more computations' cost, and waste available resources. Therefore, in this study, we propose a deep learning model in order to mitigate the traffic over cloud network and, thus reduce the cost of cloud computing. The model, which depends on the top 50 intensities and their m/z values of each spectrum, removes any spectrum which is predicted not to pass the majority voting of the participated search engines. Our results using three search engines namely: pFind, Comet and X!Tandem, and four different datasets are promising and promote the investment in deep learning to solve such type of Big data problems.

Despite the linear relation between the number of observed spectra and the searching time, the current protein search engines, even the parallel versions, could take several hours to search a large amount of MSMS spectra, which can be generated in a short time. After a laborious searching process, some (and at times, majority) of the observed spectra are labeled as non-identifiable. We evaluate the role of machine learning in building an efficient MSMS filter to remove non-identifiable spectra. We compare and evaluate the deep learning algorithm using 9 shallow learning algorithms with different configurations. Using 10 different datasets generated from two different search engines, different instruments, different sizes and from different species, we experimentally show that deep learning models are powerful in filtering MSMS spectra. We also show that our simple features list is significant where other shallow learning algorithms showed encouraging results in filtering the MSMS spectra. Our deep learning model can exclude around 50% of the non-identifiable spectra while losing, on average, only 9% of the identifiable ones. As for shallow learning, algorithms of: Random Forest, Support Vector Machine and Neural Networks showed encouraging results, eliminating, on average, 70% of the non-identifiable spectra while losing around 25% of the identifiable ones. The deep learning algorithm may be especially more useful in instances where the protein(s) of interest are in lower cellular or tissue concentration, while the other algorithms may be more useful for concentrated or more highly expressed proteins.

Deep learning has been successfully applied in several fields such as machine translation, manufacturing, and pattern recognition. However, successful application of deep learning depends upon appropriately setting its parameters to achieve high-quality results. The number of hidden layers and the number of neurons in each layer of a deep machine learning network are two key parameters, which have main influence on the performance of the algorithm. Manual parameter setting and grid search approaches somewhat ease the users' tasks in setting these important parameters. Nonetheless, these two techniques can be very time-consuming. In this paper, we show that the Particle swarm optimization (PSO) technique holds great potential to optimize parameter settings and thus saves valuable computational resources during the tuning process of deep learning models. Specifically, we use a dataset collected from a Wi-Fi campus network to train deep learning models to predict the number of occupants and their locations. Our preliminary experiments indicate that PSO provides an efficient approach for tuning the optimal number of hidden layers and the number of neurons in each layer of the deep learning algorithm when compared to the grid search method. Our experiments illustrate that the exploration process of the landscape of configurations to find the optimal parameters is decreased by 77 % - 85%. In fact, the PSO yields even better accuracy results.

Indexing research articles in scientific publications can be arduous. The authors tag their articles by the topics or domains relevant to their research. A publication's organizers may tag them by the broad topics of the specific publication. A third-party may index or tag these articles based on their subject knowledge. Hence indexing of articles can be uneven due to inconsistencies in area knowledge by third-parties or the niche topic representation by the authors. Publications may have schemes in place for indexing or tagging the articles but such schemes cannot keep up with the continuously changing landscape of research. These schemes may need to be updated with newer topics or domains being churned out by the state of the art research. Our technique endeavors to address this problem. We present a methodology to find similarity among domains extracted from the content of research papers, and cluster related domains. Analysis of these clusters provides insights into how the existing indexing schemes may be enhanced by adding newer domains.

The purpose of this study is to develop a localized muscle fatigue detection system to assist patients during isometric exercise. A mobile device is fastened to the forehand of the subject to receive electromyography (EMG) signals sent wirelessly from BITalino board. Our proposed system then uses surface electromyography (sEMG) technique to capture electromyography (EMG) signals that are processed in time-frequency domain using short-term Fourier transform. The vibration speed received through accelerometer sensor caused by the muscle fatigue at biceps brachii of a subject is calculated in parallel. The downward shift in Median Frequency and increase in vibration speed are taken as parameters to determine the localized muscle fatigue. Results indicate that localized muscle fatigue can be observed effectively with these two parameters combined together.

Studies related to public transportation systems help the commuting public by increasing road safety and circulation. These result in optimized traffic flow, shorter origin-destination travel time and reduced incident rate. Vehicular Ad-Hoc Network uses a number of sensors to gather data on the road. Intelligent Transportation Systems (ITS) draw inference from the gathered data. In this paper we discuss our experience of using Auto Regressive Integrated Moving Average (ARIMA) based techniques emphasizing on the integration of short-range and long-range dependencies of the historical traffic volume. We also analyze traffic data for patterns across different types of roads and derive computational complexity of ARIMA. Finally, improvements are identified for better prediction. We empirically show that SARIMA and ARIMA-GARCH exhibit similar road traffic prediction. ARIMA-GARCH is better than ARIMA and SARIMA for prediction, with stable model order across different historical traffic volumes. We further analyzes model orders across different types of roads and historical traffic volume; and its implications for practical applicability in ITS.

These keynote speeches discuss the following: opportunistic resource utilization networks; error free computing? where are we now?; and software development for u-healthcare services and its application.

Abstract Current trends in computing indicate that there is a great potential for service‐oriented computing and similar technologies, such as application‐oriented networks (AONs), where services can relocate to adapt to the conditions of the underlying network. In such environments, providing and consuming services and establishing a relationship between consumers (users of services) and producers (providers of services) are still challenging and vastly researched aspects. Bearing this in mind, we define a service location and planning (SLP) problem that uniquely matches producers to consumers and accounts for realistic parameters such as, quality of service (QoS) constraints of throughput and delay, and network constraints of underlying link layer bandwidth capacities, and cost of meeting consumer requests. Our contribution lies in the mathematical formulation of the SLP problem as an integer linear programming (ILP) problem that can be solved optimally for small‐scale networks and extending this work using Lagrangean relaxation (LR) approximation techniques to solve the SLP problem for large‐scale networks. Copyright © 2009 John Wiley & Sons, Ltd.

Current trends in computing indicate that there is a great potential for service oriented computing and similar technologies, such as Cisco's Application Oriented Networks, where services provide a higher-level of abstraction to traditional applications. In such cases, providing and consuming services and establishing a relationship between consumers (users of services) and producers (providers of services) are still challenging and vastly researched aspects. We define a service location planning problem (SLP) that is novel in the approach it takes to match consumers to producers, such that the cost of providing services is minimized while taking into account quality of service constraints of throughput and delay. In previous work, we presented an integer linear programming (ILP) model to formulate the service location planning problem. In this paper, we present Lagrangean Relaxation of the model for large-scale networks. We use the approach of Trick [3] to decrease computational overheads.

In tracking applications of wireless sensor networks (WSNs), the classification of an object or event of interest is envisaged to be one of the most computationally intensive tasks that recur frequently over the lifetime of the network. It is imperative that the implementations of such tasks be power efficient and computationally feasible for resource-constrained WSNs. Existing implementations of the best known classifiers such as maximum Aposterior (MAP) classifier are computationally infeasible for WSN environments. The focus of this paper is to investigate computational techniques to realize power efficient distributed implementation of the MAP classifier. In the MAP classifier, one of the most computationally challenging steps is the computation of the inverse of the covariance matrices. In this paper, we study computationally efficient methods for realizing the inverse of a matrix. We present a detailed discussion of one-sided Jacobi Iterations and LU Decomposition for approximating and computing the inverse of the covariance matrices. For LU Decomposition-based solutions, we also apply folding techniques to ensure equal power dissipation among the sensor network nodes. Our contribution lies in the power consumption analysis of executing the MAP classifier using LU decomposition and one-sided Jacobi Iterations to substitute the inverse of the covariance matrices. We show that MAP with one-sided Jacobi Iterations greatly simplifies classification process and makes it more feasible and efficient choice for sensor network applications.

Class 2 opportunistic networks (oppnets) are a new paradigm for Collaborative Computing that aims at integrating communication, computation, sensing, actuation, storage, and other resources and services. Oppnets achieve global tasks and goals through the collaboration and coordination of their nodes (some of which join an oppnet dynamically). We describe the concept of oppnets, discuss related work, and present the standard API framework for oppnets name Oppnet Virtual Machine (OVM). We present the design and implementation details of a small-scale proof-of-concept system, named MicroOppnet, in terms of the OVM primitives. We describe both the design and implementation of MicroOppnet, which not only is a proof of concept but also constitutes our tool for experiments in collaborative computing with oppnets. We are currently working on extending MicroOppnet into a larger oppnet prototype and an oppnet testbed.

We propose two lightweight techniques to detect masquerade attacks on wireless sensor networks (WSN). Our solutions take into consideration, important WSN properties like coverage, connectivity, data aggregation and specific communication patterns. The two proposed techniques complement each other when used concurrently. The mutual guarding (MG) technique does not work when nodes are not completely covered by their neighbors or when adversary has shorter transmission range than the sensor nodes. It also does not protect nodes near the boundary. The SRP technique does not have these drawbacks. In this paper, we present our proposed techniques and analyze their performance in terms of successful masquerade detection rate and traffic and computational overhead.

We present a novel paradigm of opportunistic networks or oppnets in the context of emergency preparedness and response (EPR). Oppnets constitute the category of ad hoc networks where diverse systems, not employed originally as nodes of an oppnet, join it dynamically in order to perform certain tasks they have been called to participate in. After describing the oppnets and their operation, we discuss the oppnet virtual machine (OVM) - a standard implementation framework for oppnet applications. Oppnets can significantly improve effectiveness and efficiency of EPR one of the six mission areas within the national strategy for homeland security. They can also improve other diverse applications, including agriculture, environment, healthcare, manufacturing, surveillance, and transportation. Oppnets should create new application niches as yet hard to imagine. To the best of our knowledge we have been the first to work on oppnets.

In this paper, we present and discuss a novel service location problem that satisfies user demand for services, by installing services on nodes in the network, which can also meet quality of service (QoS) constraints of throughput and delay defined by the user. Service location problems can arise in various different networks, we specifically consider service location in Cisco's Application Oriented Networks (AON). Our goal is to formulate and present the service location problem as an Integer Linear Programming (ILP) problem and solve it optimally for networks with small number of nodes. Our ILP and results meet "real-world" demands of traffic splitting (due to link layer capacity) and service federation (to incur minimal service installation cost). This will enable service providers to offer services in a cost-effective manner while meeting consumer's demand for quality of service for throughput and delay.

Collaborative signal processing is one of the most promising applications that are currently being investigated for sensor networks. In this paper, we use FFT computation as a vehicle to highlight the issues involved in realizing distributed computations over sensor networks that have global and local communication and synchronization characteristics. We present a power efficient algorithm for computing 1-D Fast Fourier Transform (FFT) over single and multi-hop wireless sensor networks. The proposed algorithm reduces the number of transmissions, eliminates typical redundant computations in a distributed FFT algorithm and uniformly maps complex multiplications over all the sensors nodes by introducing an extra bit-complement permutation stage after first (log2N)/2 iterations. We show that the proposed algorithm improves energy consumption by 36 % on the average on multi-hop sensor networks. This saving in energy consumption significantly improves the battery life of the sensor nodes thereby increasing lifetime of the sensor network. 1.

With the emergence of wireless sensor networks and pervasive computing, innovative location-aware technologies and services are being investigated. Several iterative approaches employing distributed computations over sensors have been proposed in the literature for locating all the sensor nodes in the network. Due to their iterative nature these techniques are inefficient in terms of power, a very precious resource in sensor networks. This paper presents a novel power efficient approach aimed at identifying the locations of all the nodes in a sensor network given the location of a small subset of nodes. The technique, using evolution strategies, is independent of the ranging method used to estimate distances between nodes and involves sink nodes in the computation. The proposed approach provides substantial energy savings over existing techniques while providing comparable accuracy, and requires the presence of at least one neighbor for each sensor node compared to at least 3 neighbors for most of the existing techniques

No abstract available.

The paper addresses the design of distributed methods which incorporate numerical extrapolation into adaptive multivariate integration, in order to increase the functionality of the integration algorithms. When attempting to deal with singularities, adaptive integration algorithms need a very fine subdivision in the proximity of these "hot spots". This is not practical in higher dimensions where a vast number of subregions result. These problems may be alleviated through the use of a suitable extrapolation strategy. We present a strategy which, incorporated as a global extrapolation method in distributed adaptive integration, allows for load balancing.

We describe a coarse grain parallel algorithm for multivariate adaptive integration using MPI. The algorithm is asynchronous in nature and allows for load balancing. Timing results show good speedups obtained on a network of workstations for a class of integrals from Bayesian statistics.

We present strategies for the parallel computation of the integrals typically arising in finite element problems. In order to deal efficiently with difficulties in the integrands affecting the domain in areas of which the exact locations may be unknown, a parallel adaptive strategy is applied to the entire domain. The elements comprising the domain form the initial pool of subregions in the adaptive strategy and are refined further where needed, thereby allowing load sharing or a load balanced global priority queue within each participating process group.

Multivariate integration problems arising in the real world often lead to computationally intensive numerical solutions. If the singularities and/or peaks in the integrand are not known a priori, the use of adaptive methods is recommended. The efficiency of adaptive methods depends heavily on focusing on the sub-regions that contain singularities or peaks in the integrands. In this paper, we present techniques based on evolutionary strategies that can be used to identify such subregions. Adaptive integration algorithms and evolutionary strategies can be parallelized easily and hence combining the parallel implementations of these result in efficient parallel adaptive integration algorithms.

The authors present embeddings of complete binary trees into butterfly networks with or without wrap-around connections. Let m be an even integer and q=m+(log m)-1. The authors show how to embed a 2/sup q+1/-1-node complete binary tree T(q) into a (m+1)2/sup m+1/-node wrap-around butterfly B/sub w/(m+1) with a dilation of 4, and how to embed T(q) into a (m+2)2/sup m+2/-node wrap-around butterfly B/sub w/(m+2) with an optimal dilation of 2. They also present an embedding of a wrap-around butterfly B/sub w/(m) into a (m+1)2/sup m/-node no-wrap-around butterfly B(m) with a dilation of 3. Using this embedding it is shown that T(q) can be embedded into a no-wrap butterfly B(m+1) (resp. B(m+2)) with a dilation of 8 (resp. 5).>