arXiv subjects

Seungmin Lee

Publications and source records attributed to Seungmin Lee.

25 records · Page 2Linked to original sources

arXiv2024

Very-Large-Scale GPU-Accelerated Nuclear Gradient of Time-Dependent Density Functional Theory with Tamm-Dancoff Approximation and Range-Separated Hybrid Functionals

Modern graphics processing units (GPUs) provide an unprecedented level of computing power. In this study, we present a high-performance, multi-GPU implementation of the analytical nuclear gradient for Kohn-Sham time-dependent density functional theory (TDDFT), employing the Tamm-Dancoff approximation (TDA) and Gaussian-type atomic orbitals as basis functions. We discuss GPU-efficient algorithms for the derivatives of electron repulsion integrals and exchange-correlation functionals within the range-separated scheme. As an illustrative example, we calculated the TDA-TDDFT gradient of the S1 state of a full-scale green fluorescent protein with explicit water solvent molecules, totaling 4353 atoms, at the wB97X/def2-SVP level of theory. Our algorithm demonstrates favorable parallel efficiencies on a high-speed distributed system equipped with 256 Nvidia A100 GPUs, achieving >70% with up to 64 GPUs and 31% with 256 GPUs, effectively leveraging the capabilities of modern high-performance computing systems.

physics.chem-ph↗

arXiv2020

HetPipe: Enabling Large DNN Training on (Whimpy) Heterogeneous GPU Clusters through Integration of Pipelined Model Parallelism and Data Parallelism

Deep Neural Network (DNN) models have continuously been growing in size in order to improve the accuracy and quality of the models. Moreover, for training of large DNN models, the use of heterogeneous GPUs is inevitable due to the short release cycle of new GPU architectures. In this paper, we investigate how to enable training of large DNN models on a heterogeneous GPU cluster that possibly includes whimpy GPUs that, as a standalone, could not be used for training. We present a DNN training system, HetPipe (Heterogeneous Pipeline), that integrates pipelined model parallelism (PMP) with data parallelism (DP). In HetPipe, a group of multiple GPUs, called a virtual worker, processes minibatches in a pipelined manner, and multiple such virtual workers employ data parallelism for higher performance. We also propose a novel parameter synchronization model, which we refer to as Wave Synchronous Parallel (WSP) to accommodate both PMP and DP for virtual workers, and provide convergence proof of WSP. Our experimental results on a given heterogeneous setting show that with HetPipe, DNN models converge up to 49% faster compared to the state-of-the-art DP technique.

cs.DC↗

arXiv2020

Online Graph Matching Problems with a Worst-Case Reassignment Budget

In the online bipartite matching with reassignments problem, an algorithm is initially given only one side of the vertex set of a bipartite graph; the vertices on the other side are revealed to the algorithm one by one, along with its incident edges. The algorithm is required to maintain a matching in the current graph, where the algorithm revises the matching after each vertex arrival by reassigning vertices. Bernstein, Holm, and Rotenberg showed that an online algorithm can maintain a matching of maximum cardinality by performing amortized $O(\log^2 n)$ reassignments per arrival. In this paper, we propose to consider the general question of how requiring a non-amortized hard budget $k$ on the number of reassignments affects the algorithms' performances, under various models from the literature. We show that a simple, widely-used algorithm is a best-possible deterministic algorithm for all these models. For the unweighted maximum-cardinality problem, the algorithm is a $(1-\frac{2}{k+2})$-competitive algorithm, which is the best possible for a deterministic algorithm both under vertex arrivals and edge arrivals. Applied to the load balancing problem, this yields a bifactor online algorithm. For the weighted problem, which is traditionally studied assuming the triangle inequality, we show that the power of reassignment allows us to lift this assumption and the algorithm becomes a $\frac{1}{2}$-competitive algorithm for $k=4$, improving upon the $\frac{1}{3}$ of the previous algorithm without reassignments. We show that this also is a best-possible deterministic algorithm.

cs.DS↗

arXiv2020

V2X-Based Vehicular Positioning: Opportunities, Challenges, and Future Directions

Vehicle-to-Everything (V2X) will create many new opportunities in the area of wireless communications, while its feasibility on enabling vehicular positioning has not been explored yet. Vehicular positioning is a crucial operation for autonomous driving. Its complexity and stringent safety requirement render conventional technologies like RADAR and LIDAR inadequate. This article aims at investigating whether V2X can help vehicular positioning from different perspectives. We first explain V2X's critical advantages over other approaches and suggest new scenarios of V2X-based vehicular positioning. Then we review the state-of-the-art positioning techniques discussed in the ongoing 3GPP standardization and point out their limitations. Lastly, some promising research directions for V2X-based vehicular positioning are presented, which shed light on realizing fully autonomous driving by overcoming the current barriers.

cs.IT↗

arXiv2019

Drop to Adapt: Learning Discriminative Features for Unsupervised Domain Adaptation

Recent works on domain adaptation exploit adversarial training to obtain domain-invariant feature representations from the joint learning of feature extractor and domain discriminator networks. However, domain adversarial methods render suboptimal performances since they attempt to match the distributions among the domains without considering the task at hand. We propose Drop to Adapt (DTA), which leverages adversarial dropout to learn strongly discriminative features by enforcing the cluster assumption. Accordingly, we design objective functions to support robust domain adaptation. We demonstrate efficacy of the proposed method on various experiments and achieve consistent improvements in both image classification and semantic segmentation tasks. Our source code is available at https://github.com/postBG/DTA.pytorch.

cs.CV↗

arXiv2019

Joint Frequency-and-Phase Modulation for Backscatter-Tag Assisted Vehicular Positioning

Autonomous driving (auto-driving) has been becoming a killer technology for next generation vehicles, whereas some fatal accidents grow concerns about its safety. A fundamental function for safer auto-driving is to recognize the vehicles' locations, termed vehicular positioning. The state-of-the-art vehicular positioning is to rely on anchors that are stationary objects whose locations are known, i.e. satellites for GPS and base stations for cellular positioning. It is important for reliable positioning to install anchors densely, helping find enough anchors nearby. For the deployment to be cost-effective, there are some trials to use backscatter tags as alternative anchors by deploying them on a road surface, but its gain is limited by several reasons such as short contact time and difficulties in maintenance. Instead, we propose a new backscatter-tag assisted vehicular positioning system where tags are deployed along a roadside, which enables the extension of contact duration and facilitates the maintenance. On the other hand, there is a location mismatch between the vehicle and the tag, calling for developing a new backscatter transmission to estimate their relative position. To this end, we design a novel waveform called joint frequency-and-phase modulation (JFPM) for backscatter-tag assisted vehicular positioning where a transmit frequency is modulated for the distance estimation assuming that the relevant signal is clearly differentiable from the others while the phase modulation helps the differentiation. The JFPM waveform leads to exploiting the maximum Degree-of-Freedoms (DoFs) of backscatter channel in which multiple-access and broadcasting channels coexist, leading to more accurate positioning verified by extensive simulations.

cs.IT↗

arXiv2015

Structural Analysis of Criminal Network and Predicting Hidden Links using Machine Learning

Analysis of criminal networks is inherently difficult because of the nature of the topic. Criminal networks are covert and most of the information is not publicly available. This leads to small datasets available for analysis. The available criminal network datasets consists of entities, i.e. individual or organizations, which are linked to each other. The links between entities indicates that there is a connection between these entities such as involvement in the same criminal event, having commercial ties, and/or memberships in the same criminal organization. Because of incognito criminal activities, there could be many hidden links from entities to entities, which makes the publicly available criminal networks incomplete. Revealing hidden links introduces new information, e.g. affiliation of a suspected individual with a criminal organization, which may not be known with public information. What will we be able to find if we can run analysis on a larger dataset and use link prediction to reveal the implicit connections? We plan to answer this question by using a dataset that is an order of magnitude more than what is used in most criminal networks analysis. And by using machine learning techniques, we will convert a link prediction problem to a binary classification problem. We plan to reveal hidden links and potentially hidden key attributes of the criminal network. With a more complete picture of the network, we can potentially use this data to thwart criminal organizations and/or take a Pareto approach in targeting key nodes. We conclude our analysis with an effective destruction strategy to weaken criminal networks and prove the effectiveness of revealing hidden links when attacking to criminal networks.

cs.SI↗