CGSW 9.0 Student Research Presentation Abstracts

		Session 1:  Systems and Networks
10:30am	10:45am		ALBADross: Active Learning Based Anomaly Diagnosis for Production HPC Systems Diagnosing causes of performance anomalies in High-Performance Computing (HPC) systems is a daunting challenge due to the systems’ scale and complexity.  Variations in application performance result in premature job termination, lower energy efficiency, or wasted computing resources. This paper proposes a novel active learning-based anomaly diagnosis framework that achieves a 0.95 F1-score using 28x fewer labeled samples compared to a state-of-the-art supervised approach, even when there are previously unseen applications and application inputs in the test dataset.
10:45am	11:00am		A Distributed Optimization Framework to Control the Overall Load Consumption of a Residential Neighborhood This work formulates a mixed integer linear programming problem to control the overall load consumption of a residential neighborhood by maintaining the participants’ comfort. By using a distributed optimization framework based on Dantzig-Wolfe decomposition technique, we obtain an approximate solution in reasonable time.
11:00am	11:15am		Coevolving Latent Space Networks with Attractors The recently developed coevolving latent space networks with attractors (CLSNA) is a broadly applicable class of models for temporal networks. We develop inference methods to significantly scale up the model and extend the model to networks with members join/leave, which allow for a more powerful modeling of social interactions such as polarizing forces in American politics over the past decade. In addition, we provide a mathematical characterization of the CLSNA process from the perspective of stochastic differential equations, providing greater insight into the types of phenomena that can be captured by these models.
11:15am	11:30am		Site-Wide HPC Data Center Demand Response Data centers can save energy costs by offering services in a smart grid, but such offerings may result in negative impacts to their users’ quality of service (QoS). This work shows that we can loosely couple a data center’s site-wide power model with a QoS-aware server power manager to increase the cost-saving opportunities. Through simulations, we demonstrate up to 1.3x cost savings by adding a site-wide model to a QoS-aware server demand response policy.
		Session 2:  Control of Autonomous Agents
11:45am	12:00pm		Optimal Merging Control of an Autonomous Vehicle in Mixed Traffic: an Optimal Index Policy A single autonomous vehicle merging with a group of human-driven vehicles with the goal of minimizing both travel time and energy consumption of the entire group of vehicles. As a next step, the results will be used to generalize the problem formulation by incorporating more AVs and establishing connectivity between them.
12:00pm	12:15pm		Cooperative Energy and Time-Optimal Lane Change Maneuvers with Minimal Highway Traffic Disruption
12:15pm	12:30pm		Information processing in autonomous systems in uncertain environments An information-theoretic control system design principle is to choose control actions that maximize the amount of information gained about a state or parameter. This talk considers the design of controllers that process the maximum amount of information required to achieve a given minimum admissible mean square error objective. We illustrate the concept for steering a stochastic linear discrete-time system to a desired state in finite time, and briefly discuss potential applications involving vision-based information processing of an uncertain scene.
12:30pm	12:45pm		Heterogeneous Coverage and Multi-Resource Allocation in Supply-Constrained Teams We consider a team of heterogeneous robots, each equipped with various types and quantities of resources, and tasked with supplying these resources to multiple areas of demand. We propose a Voronoi-based coverage control approach to deploy robots to areas of demand by minimizing the locational cost, and allowing robots to prioritize the various demand locations in a continuous, distributed fashion.
		Session 3:  Bio/Chem Inspired Methods and Applications
2:15pm	2:30pm	Arincheyan Gerald	Multimodal Sensing and Haptic Feedback to Improve Minimally Invasive Surgery
2:30pm	2:45pm		Improved Predictions of MHC-peptide Binding using Protein Language Models Predicting the interaction between major histocompatibility complex molecules (MHC) with peptides (from exogenous antigens) is an important step in cellular immune system studies. In this work, using protein language modes and deep learning techniques, a computational model is developed to predict the binding between MHC-I and peptides. This model is capable to outperform the current stat-of-the-art methods.
2:45pm	3:00pm	Mehdi Kermanshah	Robust Filtering based on Complex Cell Networks from the Visual Cortex
3:00pm	3:15pm		Particle-based Stochastic Reaction-Drift-Diffusion Models Volume reactivity particle-based stochastic reaction-drift-diffusion (PBSRDD) models provide a meso-scale perspective into the spatial dynamics of chemical and biological systems. In this work, we prove the rigorous large-population limit of the weak measure-valued stochastic process (MVSP) representation for our new PBSRDD model, with the resulting mean field limit corresponding to a system of partial integro-differential equations (PIDEs).
		Session 4:  Learning from Data
3:30pm	3:45pm		Implicit Parameter-free Online Learning with Truncated Linear Models Propose new parameter-free algorithms that can take advantage of truncated linear models through a new update that has an “implicit” flavor. Based on a novel decomposition of the regret, the new update is efficient, requires only one gradient at each step, never overshoots the minimum of the truncated model, and retains the favorable parameter-free properties.
3:45pm	4:00pm		Optimal Comparator Adaptive Online Learning with Switching Cost I will introduce a framework to achieve problem-dependent acceleration for sequential decision making problems with long term effects. The framework is designed by transforming the discrete time problem into continuous time, which makes the key problem structure easier to analyze.
4:00pm	4:15pm		AugCSE: Contrastive Sentence Embedding with Diverse Augmentations Data augmentation techniques can often introduce noise and are not applicable to domain general application. In our work, we leverage a diverse set of data augmentations to improve generic sentence embedding models. Our method is simple, and effective at improving general representational space, with applications outside of natural language processing.
		Session 5:  Privacy and Security
4:30pm	4:45pm		Correlation Detection of Databases In this work, we consider the database correlation detection problem: given two databases that each contains a random Gaussian feature vector per user, can we determine whether they are correlated by an unknown user permutation?  We provide upper and lower bounds on the risk of this statistical problem which are tight up to constants when the feature length is larger than the number of users. Moreover, we are able to show that, in certain regimes, correlation detection is easier than permutation recovery in this setting.
4:45pm	5:00pm		Privacy-preserving for network telemetry systems Network telemetry systems face challenges such as user privacy protection, data retention cost, and query accuracy. We develop an autoencoder-based system to find a solution for these tradeoffs. We aim to preserve user privacy while minimizing the data storage size and keeping high query accuracy.
5:00pm	5:15pm		PyReT – Python Real-Time Vulnerability Detection We developed a novel tool to detect vulnerabilities while a developer is writing code. We leverage the power of BERT and GNNs to understand the semantic and structural properties of vulnerabilities in source code. We aim to reduce the human effort of designing rules for the rule-based vulnerability detection tools.