ACM Student Research Competition
This work introduces visualization tools for Neuroscience-Inspired Dynamic Architecture (NIDA) networks and for the Dynamic Adaptive Neural Network Array (DANNA) hardware implementation of NIDA. NIDA is a novel artificial neural network architecture that performs well on various tasks. The 3D NIDA visualization shows network structure and simulates activity on NIDAs. The 2D interface for DANNAs represents the physical arrangement of elements and provides interactions to configure and save networks.
People engaged in multiple primary activities at once (e.g., driving and browsing the web) cannot efficiently access web content and safely monitor the roads at the same time. Thus, researchers must explore ways to enable users to access the mobile web effectively while driving with less distraction from competing activities. In this paper, we investigate the role of using voice-controlled aural interfaces while driving.
Asthma affects an increasing fraction of the population, especially children, and a known trigger is air pollution. Researchers have tried to quantify the association between pollution levels and incidence, but the problem is intrinsically difficult and hard to solve with traditional statistical methods. With this poster, we will present the most hindering limitations of traditional methods, and propose an alternative study based on itemset mining modified for risk characterization.
Evaluating researchers' productivity is a non trivial task that depends on various aspects. Usually, such systems rely on bibliometry, which may not be always fair. Another important aspect is the properties of the relationships among researchers. We take a step forward on analyzing such
data by exploring the strength of co-authorship ties in social
networks. Our results show that different properties explain variations in the strength of ties.
Mobile phones use a push notification channel for receiving notifications from server. To maintain this connection we send Keep-Alive packets. Our project answer the question of how efficiently we can send KA packets. We have discussed four strategies. First strategy sends KA packets when the radio connection is ON. Second strategy sends packets at KA interval. Other two approaches use probability to send KA packets. Fourth approach is better relatively.
We present a game theoretic analysis of various endgames in a quantized version of chess. Our mathematical analysis examines the probabilistic outcomes of quantized endgame studies and establishes that in quantized chess, it is possible to force checkmate with two knights. This result represents a significant departure from endgame theory in classical chess, which holds that in conditions of perfect play, two knights cannot force checkmate.
To model energy consumption of nano and centralized DCs, traffic-based and time-based energy consumption models are proposed.
Our results indicate that saving energy using nano DCs depends on: type of applications, type of access network attached to nano servers, and the active time of the nano servers. Our results show that the best energy savings comes from applications that generate data continuously in end-user premises such as sensors in IoTs.
Description logic (DL-)programs are a prominent approach to combine nomonotonic rules and ontological knowledge bases. As it happens, inconsistency may arise from the interplay of the rules and the ontology. We develop approaches for resolving inconsistencies in DL-programs over lightweight Description Logics, realize the algorithms within the dlvhex system and evaluate repair computation techniques on a set of novel benchmarks. The evaluation results demonstrate effectiveness of our algorithms.
In this work we evaluate techniques for discovering virtual machines behind a load balancer configured in public IaaS clouds. We investigate four methods: IPid sequences, TCP timestamps, front-end identifiers in webpages, and sticky session cookies. We use controlled experiments in Amazon EC2, packet capture and analysis, and scraping of webpages. We find the techniques work in different scenarios, so an ideal measurement framework should use a combination of techniques.
Body Area Networks are considered as the key building blocks for the emerging application of upcoming networks of wearable technology as well as Internet of Things. The longevity and network lifetime of Wireless BANs are crucial concerns. This paper describes research-based approaches that include data aggregation, topology control and sensor placements which provide a unifying frame to investigate effect of energy consumption, network lifetime and thus, improving the network performances.
The Kolmogorov complexity of a string is the length of the shortest possible description of the string. The Normalized Information Distance measures the similarity between strings using Kolmogorov complexity. Based on the hypothesis that there is shared information between DNA-binding proteins and between genes within operons, we have determined a novel approach of using a support vector machine with a Kolmogorov complexity kernel to classify DNA-binding proteins and operons.
Motion vectors of depth-maps in multiview and free-viewpoint videos
exhibit strong spatial and inter-component clustering tendency. This paper presents a novel motion vector coding technique that first compresses the multidimensional bitmaps of macroblock mode information and then encodes only the non-zero components of motion vectors. This technique is capable of exploiting the spatial and inter-component correlations efficiently without the restriction of scanning the bitmap in any specific linear order .
The need for cost-effective and thorough web application testing is growing greater constantly as the prevalence of web applications increases, however current web application testing methods tend to be blind to subject applications’ persistent state, creating test suites that are ultimately cost-ineffective. In order to address this, we leveraged an evaluation of resource and parameter naming conventions to develop a parameter-determining algorithm that returns a persistent-state aware test suite.
We set out to explore how Social Network Analysis (SNA) can be applied to the field of security by investigating what interactions could an attacker infer if they were able to gather data on a person’s network and whether this could help predict their “hierarchical importance”. For this experiment we conducted our investigation on the Enron email dataset and then investigated the effectiveness of our techniques on a fresh dataset.
Power density has become the limiting factor for keeping the path of performance improvements across technology generations.
A promise technique to cope with this limitation is operating chips at voltages close to the threshold voltage.
However, at ultra-low voltages the stability of SRAM cells is compromised, leading to potential failures.
Taking advantage of the natural redundancy of on-chip data, we propose content management techniques for efficient operation at near-threshold voltages.
Present study provides a useful insight towards the determinant factors for preservation of ligand’s bound pose upon optimization. Binding mode analysis of protein-ligand complex entries in PDBBind where one ligand could have naturally been derived from chemical optimization of the other showed that ligand-flips in binding orientation are surprisingly common (~16%). These flips can be predicted by machine learning model based on properties of ligand with an accuracy of 89%.
Sustained periods of elevated intracranial pressure (ICP) in traumatic brain injury patients can result in death or severe disability. Detecting these crisis periods is a challenging task that currently relies on clinical expertise. I present a statistical machine learning model, based on data from 556 patients, that predicts crises 30 minutes in the future using only 30 minutes of ICP values and the time since a patient's last crisis.
We show that cache-oblivious recursive divide-and-conquer
(CORDAC) algorithms for solving a class of dynamic programming problems are more energy-efficient and have more flexibility in the runtime-power tradeoff than their corresponding iterative or tiled algorithms. Our experimental results show that CORDAC algorithms are more robust in
terms of performance, have better bandwidth-utilizations,
and more cache-adaptive in a multiprogramming environment compared to the popular iterative and tiled-loop implementations of the same problem.
We propose a new attestation protocol to detect replay-attacks on sensors that continuously perceive the physical environment and report readings to an external entity. We evaluate our approach for cameras in surveillance applications. The breakthrough of our approach is that we do not send the attestation challenge to the device; instead we modify the environment the device is observing and verify that the desired changes reflect in the sensor readings.
Metagenomic studies have primarily relied on de novo approaches for reconstructing genes and genomes from microbial mixtures. As the number of genomes present in public databases is rapidly increasing, they can be an invaluable resource in assisting and improving metagenomic assemblies. Here I describe the first indexing-based approach for reference guided metagenomic assembly that can effectively complement de novo metagenomic assembly methods, leading to substantially improved reconstructions.
This work focuses on the propagation of influence through the concept of communities, which are formed by researchers who share common interests. Our goal is to measure the degree of influence of a researcher in the communities he/she publishes. Then, we propose an indicator, called 3c-index (Cross-Community Contribution), to aggregate the score of the researcher in the base community and the transfer of influence among communities.
Why do granite countertops appear shiny while granite curbs, with the same chemical composition, appear matte? Computer graphics Microfacet Theory predicts macroscopic surface shininess from microscopic geometry. I built a digital microscope-microtome for imaging micrometer scale geometry in 2D slices and an algorithm for reconstructing the 3D microsurface. My preliminary results experimentally validate Microfacet Theory. As future work, I will compare ray tracing simulations on my data to real photographs.
WaveCluster is an important family of grid-based clustering algorithms. In this paper, we investigate techniques to perform WaveCluster while ensuring differential privacy. We show that straightforward techniques based on synthetic data generation and introduction of random noise when quantizing the data, introduce too much noise to preserve useful clusters. We then propose two optimized techniques, and conduct extensive experiments that demonstrate the superiority of the optimized techniques.
In object tracking in videos, object's appearance changes due to real-world unconstrained factors that contribute to challenges for tracking algorithms. A new single object tracking approach is proposed to solve these problems by :(1) using features from a context region having correlated (similar) motion with the object; (2) constructing a structure that represents the spatial relationship of these features (extracted from [target + context] region).
In this research, we developed a novel learning based co-clustering technique that improves predictive modeling in high dimensional and sparse data. We show with extensive empirical evaluation in diverse domain that our method generates co-clusters that improve predictive power of learning model by 10% over the original data set and have better performance than two state-of-the-art co-clustering methods. The results demonstrate the effectiveness and utility of our algorithm in practice.
CrowdCheer attempts to utilize the crowd to motivate the runners when they need it most. I am studying the effects of crowdsourced & targeted motivation for runners because I want to understand how I might utilize the existing crowd to deliver targeted motivation and thus improve the runner’s performance. I take the novel approach of situational crowdsourcing to better leverage the crowd’s ability to supply motivation for the runner.
This project focuses on solving the problem of tracking using a deep learning approach. This experiment takes cues from how the brain is modeled to create deep convolutional networks that mimic how the human brain tracks objects. By using optical flow and deep networks to implement a dual appearance and motion stream, our tracker outperforms current state of the art methods.