arXiv subjects

Minseok Choi

Publications and source records attributed to Minseok Choi.

At least 37 records · Page 2Linked to original sources

arXiv2024

PairEval: Open-domain Dialogue Evaluation with Pairwise Comparison

Building a reliable and automated evaluation metric is a necessary but challenging problem for open-domain dialogue systems. Recent studies proposed evaluation metrics that assess generated responses by considering their relevance to previous dialogue histories. Although effective, these metrics evaluate individual responses directly rather than considering their relative quality compared to other responses. To handle this, we propose PairEval, a novel dialogue evaluation metric for assessing responses by comparing their quality against responses in different conversations. PairEval is built on top of open-sourced and moderate-size language models, and we make them specialized in pairwise comparison between dialogue responses. Extensive experiments on multiple benchmarks demonstrate that our metric exhibits a higher correlation with human judgments than baseline metrics. We also find that the proposed comparative metric is more robust in detecting common failures from open-domain dialogue systems, including repetition and speaker insensitivity.

cs.CL↗

arXiv2024

Protecting Privacy Through Approximating Optimal Parameters for Sequence Unlearning in Language Models

Although language models (LMs) demonstrate exceptional capabilities on various tasks, they are potentially vulnerable to extraction attacks, which represent a significant privacy risk. To mitigate the privacy concerns of LMs, machine unlearning has emerged as an important research area, which is utilized to induce the LM to selectively forget about some of its training data. While completely retraining the model will guarantee successful unlearning and privacy assurance, it is impractical for LMs, as it would be time-consuming and resource-intensive. Prior works efficiently unlearn the target token sequences, but upon subsequent iterations, the LM displays significant degradation in performance. In this work, we propose Privacy Protection via Optimal Parameters (POP), a novel unlearning method that effectively forgets the target token sequences from the pretrained LM by applying optimal gradient updates to the parameters. Inspired by the gradient derivation of complete retraining, we approximate the optimal training objective that successfully unlearns the target sequence while retaining the knowledge from the rest of the training data. Experimental results demonstrate that POP exhibits remarkable retention performance post-unlearning across 9 classification and 4 dialogue benchmarks, outperforming the state-of-the-art by a large margin. Furthermore, we introduce Remnant Memorization Accuracy that quantifies privacy risks based on token likelihood and validate its effectiveness through both qualitative and quantitative analyses.

cs.CL↗

arXiv2024

Crystallographic defects in Weyl semimetal LaAlGe

Crystallographic defects in a topological semimetal can result in charge doping, and the scattering due to the defects may mask its exotic transport properties. Here, we investigate the possible crystallographic defects including vacancy and antisite in Weyl semimetal LaAlGe using hybrid-density-functional theory calculations. We show that a considerable concentration of Al- and Ge-related defects naturally form during growth due to their low formation enthalpy. Specifically, Al can be easily replaced by Ge in the $I4_1md$ phase of LaAlGe, forming the Ge-on-Al antisite, Ge$_{\rm Al}$. The counterpart, Al-on-Ge (Al$_{\rm Ge}$), is also probable. The most abundant defect Ge$_{\rm Al}$ is donor-like, effectively electron-doping, and these defects are therefore not only scattering centers in the electronic transport process but may also induce the substantial vertical shift of the chemical potential. The results imply that the naturally occurring defects hinder both spectroscopic and transport features arising from the Weyl physics in LaAlGe. Our work can be applied to the $R$AlGe family ($R$=rare earth) and help improve the quality of single-crystal magnetic Weyl semimetal.

cond-mat.mtrl-sci↗

arXiv2023

SimCKP: Simple Contrastive Learning of Keyphrase Representations

Keyphrase generation (KG) aims to generate a set of summarizing words or phrases given a source document, while keyphrase extraction (KE) aims to identify them from the text. Because the search space is much smaller in KE, it is often combined with KG to predict keyphrases that may or may not exist in the corresponding document. However, current unified approaches adopt sequence labeling and maximization-based generation that primarily operate at a token level, falling short in observing and scoring keyphrases as a whole. In this work, we propose SimCKP, a simple contrastive learning framework that consists of two stages: 1) An extractor-generator that extracts keyphrases by learning context-aware phrase-level representations in a contrastive manner while also generating keyphrases that do not appear in the document; 2) A reranker that adapts scores for each generated phrase by likewise aligning their representations with the corresponding document. Experimental results on multiple benchmark datasets demonstrate the effectiveness of our proposed approach, which outperforms the state-of-the-art models by a significant margin.

cs.CL↗

arXiv2023

PRiSM: Enhancing Low-Resource Document-Level Relation Extraction with Relation-Aware Score Calibration

Document-level relation extraction (DocRE) aims to extract relations of all entity pairs in a document. A key challenge in DocRE is the cost of annotating such data which requires intensive human effort. Thus, we investigate the case of DocRE in a low-resource setting, and we find that existing models trained on low data overestimate the NA ("no relation") label, causing limited performance. In this work, we approach the problem from a calibration perspective and propose PRiSM, which learns to adapt logits based on relation semantic information. We evaluate our method on three DocRE datasets and demonstrate that integrating existing models with PRiSM improves performance by as much as 26.38 F1 score, while the calibration error drops as much as 36 times when trained with about 3% of data. The code is publicly available at https://github.com/brightjade/PRiSM.

cs.CL↗

arXiv2023

Two Tales of Platoon Intelligence for Autonomous Mobility Control: Enabling Deep Learning Recipes

This paper presents the deep learning-based recent achievements to resolve the problem of autonomous mobility control and efficient resource management of autonomous vehicles and UAVs, i.e., (i) multi-agent reinforcement learning (MARL), and (ii) neural Myerson auction. Representatively, communication network (CommNet), which is one of the most popular MARL algorithms, is introduced to enable multiple agents to take actions in a distributed manner for their shared goals by training all agents' states and actions in a single neural network. Moreover, the neural Myerson auction guarantees trustfulness among multiple agents as well as achieves the optimal revenue of highly dynamic systems. Therefore, we survey the recent studies on autonomous mobility control based on MARL and neural Myerson auction. Furthermore, we emphasize that integration of MARL and neural Myerson auction is expected to be critical for efficient and trustful autonomous mobility services.

cs.AI↗

arXiv2023

HistRED: A Historical Document-Level Relation Extraction Dataset

Despite the extensive applications of relation extraction (RE) tasks in various domains, little has been explored in the historical context, which contains promising data across hundreds and thousands of years. To promote the historical RE research, we present HistRED constructed from Yeonhaengnok. Yeonhaengnok is a collection of records originally written in Hanja, the classical Chinese writing, which has later been translated into Korean. HistRED provides bilingual annotations such that RE can be performed on Korean and Hanja texts. In addition, HistRED supports various self-contained subtexts with different lengths, from a sentence level to a document level, supporting diverse context settings for researchers to evaluate the robustness of their RE models. To demonstrate the usefulness of our dataset, we propose a bilingual RE model that leverages both Korean and Hanja contexts to predict relations between entities. Our model outperforms monolingual baselines on HistRED, showing that employing multiple language contexts supplements the RE predictions. The dataset is publicly available at: https://huggingface.co/datasets/Soyoung/HistRED under CC BY-NC-ND 4.0 license.

cs.CL↗

arXiv2023

SplitGP: Achieving Both Generalization and Personalization in Federated Learning

A fundamental challenge to providing edge-AI services is the need for a machine learning (ML) model that achieves personalization (i.e., to individual clients) and generalization (i.e., to unseen data) properties concurrently. Existing techniques in federated learning (FL) have encountered a steep tradeoff between these objectives and impose large computational requirements on edge devices during training and inference. In this paper, we propose SplitGP, a new split learning solution that can simultaneously capture generalization and personalization capabilities for efficient inference across resource-constrained clients (e.g., mobile/IoT devices). Our key idea is to split the full ML model into client-side and server-side components, and impose different roles to them: the client-side model is trained to have strong personalization capability optimized to each client's main task, while the server-side model is trained to have strong generalization capability for handling all clients' out-of-distribution tasks. We analytically characterize the convergence behavior of SplitGP, revealing that all client models approach stationary points asymptotically. Further, we analyze the inference time in SplitGP and provide bounds for determining model split ratios. Experimental results show that SplitGP outperforms existing baselines by wide margins in inference time and test accuracy for varying amounts of out-of-distribution samples.

cs.LG↗

arXiv2022

Bayesian deep learning framework for uncertainty quantification in high dimensions

We develop a novel deep learning method for uncertainty quantification in stochastic partial differential equations based on Bayesian neural network (BNN) and Hamiltonian Monte Carlo (HMC). A BNN efficiently learns the posterior distribution of the parameters in deep neural networks by performing Bayesian inference on the network parameters. The posterior distribution is efficiently sampled using HMC to quantify uncertainties in the system. Several numerical examples are shown for both forward and inverse problems in high dimension to demonstrate the effectiveness of the proposed method for uncertainty quantification. These also show promising results that the computational cost is almost independent of the dimension of the problem demonstrating the potential of the method for tackling the so-called curse of dimensionality.

stat.ML↗

arXiv2022

Age-of-Information Aware Contents Caching and Distribution for Connected Vehicles

To support rapid and accurate autonomous driving services, road environment information, which is difficult to obtain through vehicle sensors themselves, is collected and utilized through communication with surrounding infrastructure in connected vehicle networks. For this reason, we consider a scenario that utilizes infrastructure such as road side units (RSUs) and macro base station (MBS) in situations where caching of road environment information is required. Due to the rapidly changed road environment, a concept which represents a freshness of the road content, age of information (AoI), is important. Based on the AoI value, in the connected vehicle system, it is essential to keep appropriate content in the RSUs in advance, update it before the content is expired, and send the content to the vehicles which want to use it. However, too frequent content transmission for the minimum AoI leads to indiscriminate use of network resources. Furthermore, a transmission control, that content AoI and service delay are not properly considered adversely, affects user service. Therefore, it is important to find an appropriate compromise. For these reasons, the objective of this paper is about to reduce the system cost used for content delivery through the proposed system while minimizing the content AoI presented in MBS, RSUs and UVs. The transmission process, which is able to be divided into two states, i.e., content caching and service, is approached using Markov decision process (MDP) and Lyapunov optimization framework, respectively, which guarantee optimal solutions, as verified via data-intensive performance evaluation.

cs.DC↗

arXiv2022

Search Space Adaptation for Differentiable Neural Architecture Search in Image Classification

As deep neural networks achieve unprecedented performance in various tasks, neural architecture search (NAS), a research field for designing neural network architectures with automated processes, is actively underway. More recently, differentiable NAS has a great impact by reducing the search cost to the level of training a single network. Besides, the search space that defines candidate architectures to be searched directly affects the performance of the final architecture. In this paper, we propose an adaptation scheme of the search space by introducing a search scope. The effectiveness of proposed method is demonstrated with ProxylessNAS for the image classification task. Furthermore, we visualize the trajectory of architecture parameter updates and provide insights to improve the architecture search.

cs.LG↗

arXiv2022

AoI-Aware Markov Decision Policies for Caching

We consider a scenario that utilizes road side units (RSUs) as distributed caches in connected vehicular networks. The goal of the use of caches in our scenario is for rapidly providing contents to connected vehicles under various traffic conditions. During this operation, due to the rapidly changed road environment and user mobility, the concept of age-of-information (AoI) is considered for (1) updating the cached information as well as (2) maintaining the freshness of cached information. The frequent updates of cached information maintain the freshness of the information at the expense of network resources. Here, the frequent updates increase the number of data transmissions between RSUs and MBS; and thus, it increases system costs, consequently. Therefore, the tradeoff exists between the AoI of cached information and the system costs. Based on this observation, the proposed algorithm in this paper aims at the system cost reduction which is fundamentally required for content delivery while minimizing the content AoI, based on Markov Decision Process (MDP) and Lyapunov optimization.

cs.DC↗

arXiv2022

Quantum Distributed Deep Learning Architectures: Models, Discussions, and Applications

Although deep learning (DL) has already become a state-of-the-art technology for various data processing tasks, data security and computational overload problems often arise due to their high data and computational power dependency. To solve this problem, quantum deep learning (QDL) and distributed deep learning (DDL) has emerged to complement existing DL methods. Furthermore, a quantum distributed deep learning (QDDL) technique that combines and maximizes these advantages is getting attention. This paper compares several model structures for QDDL and discusses their possibilities and limitations to leverage QDDL for some representative application scenarios.

quant-ph↗

arXiv2022

SWIPT-enabled NOMA in Distributed Antenna System with Imperfect Channel State Information for Max-Sum-Rate and Max-Min Fairness

Motivated by the fact that the data rate of non-orthogonal multiple access (NOMA) can be greatly increased with the help of the distributed antenna system (DAS), we presents a framework in which the DAS contributes not only to the data rate but also the energy harvesting of simultaneous wireless information and power transfer (SWIPT) enabled NOMA. This study considers the sum-rate maximization problem and the max-min fairness problem for SWIPT-enabled NOMA in DAS and proposes two different schemes of power splitting and power allocation for SWIPT and NOMA, respectively, with imperfect channel state information (CSI). Numerical results validate the theoretical findings and demonstrate that the proposed framework of using SWIPT-enabled NOMA in DAS achieves the higher data rates than the existing SWIPT-enabled NOMA while guaranteeing the minimum harvested energy.

cs.IT↗

arXiv2021

Impacts of Device Caching of Content Fractions on Expected Content Quality

This paper explores caching of fractions of a video content, not caching of an entire content, to increase the expected video quality. We first show that the highest-quality content is better to be cached and propose the caching policy of video chunks having different qualities. Our caching policy utilizes the characteristics of video contents that video files can be encoded into multiple versions with different qualities, each file consists of many chunks, and chunks can have different qualities. Extensive performance evaluations are conducted to show that caching of content fractions, rather than an entire content, can improve the expected video quality especially when the channel conditions is sufficiently good to cooperate with nearby BS or helpers.

cs.MM↗

arXiv2021

Delay-Sensitive and Power-Efficient Quality Control of Dynamic Video Streaming using Adaptive Super-Resolution

In a decade, the adaptive quality control of video streaming and the super-resolution (SR) technique have been deeply explored. As edge devices improved to have exceptional processing capability than ever before, streaming users can enhance the received image quality to allow the transmitter to compress the images to save its power or pursue network efficiency. In this sense, this paper proposes a novel dynamic video streaming algorithm that adaptively compresses video chunks at the transmitter and separately enhances the quality at the receiver using SR. In order to allow transmission of video chunks with different compression levels and control of the computation burden, we present the adaptive SR network which is optimized by minimizing the weighted sum of losses extracted from different layer outputs. for dynamic video streaming. In addition, we jointly orchestrate video delivery and resource usage, and the proposed video delivery scheme balances the tradeoff well among the average video quality, the queuing delay, buffering time, transmit power, and computation power. Simulation results show that the proposed scheme pursues the quality-of-services (QoS) of the video streaming better than the adaptive quality control without the cooperation of the transmitter and the receiver and the non-adaptive SR network.

cs.MM↗

arXiv2021

Quality-Aware Deep Reinforcement Learning for Streaming in Infrastructure-Assisted Connected Vehicles

This paper proposes a deep reinforcement learning-based video streaming scheme for mobility-aware vehicular networks, e.g., vehicles on the highway. We consider infrastructure-assisted and mmWave-based scenarios in which the macro base station (MBS) cannot directly provide the streaming service to vehicles due to the short range of mmWave beams so that small mmWave base stations (mBSs) along the road deliver the desired videos to users. For a smoother streaming service, the MBS proactively pushes video chunks to mBSs. This is done to support vehicles that are currently covered and/or will be by each mBS. We formulate the dynamic video delivery scheme that adaptively determines 1) which content, 2) what quality and 3) how many chunks to be proactively delivered from the MBS to mBSs using Markov decision process (MDP). Since it is difficult for the MBS to track all the channel conditions and the network states have extensive dimensions, we adopt the deep deterministic policy gradient (DDPG) algorithm for the DRL-based video delivery scheme. This paper finally shows that the DRL agent learns a streaming policy that pursues high average quality while limiting packet drops, avoiding playback stalls, reducing quality fluctuations and saving backhaul usage.

cs.NI↗

arXiv2021

Defect engineering of magnetic ground state in EuTiO$_3$ epitaxial thin films

Atomistic defect engineering through the pulsed laser epitaxy of perovskite transition metal oxides offers facile control of their emergent opto-electromagnetic and energy properties. Among the various perovskite oxides, EuTiO3 exhibits a strong coupling between the lattice, electronic, and magnetic degrees of freedom. This coupling is highly susceptible to atomistic defects. In this study, we investigated the magnetic phase of EuTiO$_3$ epitaxial thin films via systematic defect engineering. A magnetic phase transition from an antiferromagnet to a ferromagnet was observed when the unit cell volume of EuTiO3 expanded due to the introduction of Eu-O vacancies. Optical spectroscopy and density functional theory calculations show that the change in the electronic structure as the ferromagnetic phase emerges can be attributed to the weakened Eu-Ti-Eu super-exchange interaction and the developed ferromagnetic Eu-O-Eu interaction. Facile defect engineering in EuTiO$_3$ thin films facilitates understanding and tailoring of their magnetic ground state.

cond-mat.mtrl-sci↗