Computing Seminar Schedule (2021 Fall)


Meeting time: 2:00-3:00 pm


The Computing Seminar series is run by the Computer Division of EE at KAIST (homepage)
Subscription to Google Groups

List of semesters
2022 Spring | 2021 Fall | 2021 Spring | 2020 | 2019 | 2018 |

Category
Date
Title
Presenter
Location
Ph.D. Defense
6/3 (Fri) 4:00 PM
국문: 거대 다중안테나 시스템에서 기계학습 기반 적응형 채널예측 기법 연구
English: Massive MIMO Channel Prediction Via Adaptive Machine Learning
김환진 (Hwanjin Kim)
IT융합빌딩(N1) 709호 & Online:
Abstract
In this paper, we propose the adaptive channel prediction for massive multiple-input multiple-output (MIMO) systems. Accurate channel state information is crucial to fully exploit massive MIMO systems. However, wireless channels vary in time due to the mobility of user equipment, resulting in performance degradation in massive MIMO. Thus, we develop and compare a vector Kalman filter (VKF)-based chan- nel predictor and a machine learning (ML)-based channel predictor. The VKF-based channel predictor developed in this paper exploits the autoregressive parameters based on the Yule-Walker equations. Then, the ML-based channel predictor using the linear minimum mean square error (LMMSE)-based noise pre-processed data is developed. Numerical results reveal that both channel predictors have sub- stantial gain over the outdated channel in terms of the channel prediction accuracy and data rate. However, both channel prediction techniques require large time overhead to obtain accurate prediction, which cannot effectively adapt to new environments. To accurately predict wireless channels for new environments with reduced training overhead, we propose a fast adaptive channel prediction technique based on a meta-learning algorithm for massive MIMO systems. We exploit the model-agnostic meta- learning (MAML) algorithm to achieve quick adaptation with small amount of labeled data. Also, to improve the prediction accuracy, we adopt the denoising process for the training data by using deep im- age prior (DIP). Numerical results show that the proposed MAML-based channel predictor can improve the prediction accuracy with only a small number of ne-tuning samples.
Refer to TA announcement email for zoom link
Ph.D. Defense
6/3 (Fri) 2:00 PM
국문: 확률론적 세기 성형 기술을 이용한 세기 변조/직접 수신 광전송 시스템
English: Intensity-modulation/direct-detection optical transmission system using probabilistic intensity shaping
김대호 (Daeho Kim)
정보전자공학동(E3-2) 4217호
Abstract
Probabilistic constellation shaping (PCS) has recently attracted considerable interest in optical fiber transmission systems. PCS adjusts the probability of occurrence of the symbol constellation points to maximize the information rate meanwhile maintaining their geometric location. This technique shows a remarkable diminution of the shaping gap in the modulation-constraint signaling by approximating Gaussian signaling. The intensity modulation (IM) and direct detection (DD) optical transmission systems are the simplest of all-optical transmission schemes. Since IM-DD systems offer lower power consumption and lower system cost, this system is mainly used in short- and intermediate-reach applications. To meet the growing demand for data traffic, the significant challenge of the IM-DD systems is to increase transmission rate while ensuring their characteristics in cost and power consumption. In the IM-DD systems, the PCS may be used as a method of increasing the transmission rate while keeping low power consumption and low implementation cost. However, the IM-DD system which deliver the signal inposing it onto the optical intensity has differnt system constraints from the optical coherent transmission system in which the PCS is successfully activated. In this dissertation, the system constraints of various IM-DD system have been investigated for applying the PCS technique and the probability distribution of the IM signal have benn studied to achieve channel capacity in the IM-DD systems.
Ph.D. Defense
6/2 (Thu) 3:00 PM
국문: 레이다 및 통신 시스템을 위한 60 GHz 4 채널 송신단 빔포밍 회로
English: 60 GHz 4 Channel Transmitter Beamforming Front-end IC for Radar and Communication
소철 (Cheol So)
나노종합기술원 설악산세미나실(E19, S-311호)
Abstract
The 60 GHz frequency band with a wide bandwidth may have a high data rate in communication and an increased distance resolution in radar. However, the 60 GHz frequency band has a large attenuation characteristic in the atmosphere, and accordingly, the SNR decreases. To increase the SNR, beamforming technology is being studied. In this thesis, a study on the transmitter front-end IC for beamforming is conducted. Phase shifter adapts a pair of RF-DAC structures for each turn on and off complementarily. This makes the input and output impedances invariant to control states, which are significant characteristics of reducing the phase shifter’s RMS gain and phase errors. The RF-DAC cell is implemented with body floated FETs to reduce the Cds at on and off cells for high gain. The proposed I/Q signal generator is implemented by inserting an R-L polyphase filter into the inter-stage node of a cascode amplifier, which provides good inter-stage matching and results in a finite gain in the I/Q generator. In a power amplifier, linearization techniques are essential to get high linear output power. Linearization techniques that dynamically neutralize the parasitic capacitor and dynamically controlled matching network at the cascode internode are proposed. A 1-channel transmitter beamforming front-end IC composed of a phase shifter and a power amplifier is implemented, and a 4-channel transmitter beamforming front-end IC is implemented. A test module is implemented to verify the 4-channel transmitter beamforming front-end IC, and its performance is verified through free space measurements.
Ph.D. Defense
6/2 (Thu) 10:00 AM
국문: 산화 아연 기반 복합 재료의 생체역학적 에너지 수확 능력에 관한 연구
English: A Study on Biomechanical Energy Harvesting Capability of ZnO-based Composite Materials
윤종세 (Chongsei Yoon)
미래융합소자동(E3-3) 2301호 세미나실
Abstract
This dissertation presents the studies on zinc oxide (ZnO)-based piezoelectric nanogenerators such as the highly cost-effective design for an aluminum (Al) foil-based ZnO/Ag/ZnO-stacked piezoelectric nanogenerator (ZAZ-NG), a symmetrically stacked highly durable and energy-efficient sandwich-type ZnO/carbon tape/ZnO piezoelectric nanogenerator (ZCZ-NG), and an eco-friendly stretchable flexoelectricity-enhanced piezoelectric nanogenerator (F-PENG) based on zinc-aluminum layered double hydroxide nanosheets (ZnAl:LDH Ns)-ZnO heterostructure. Both Al foil sheets and a silver (Ag) paste layer were utilized to make a ZAZ-NG composed of an Ag paste layer sandwiched between two ZnO layers. The output voltages of the ZAZ-NGs with various ZnO thicknesses are measured for three different bending strains. As a result, the devices could generate a relatively high peak-to-peak output voltage (Vpp) of up to 2.5 V, which is 28 times higher than that of the single ZnO layered device. In addition, the device performance shows a strong dependence on the thickness of the ZnO layer. Moreover, the ZAZ-NG is structurally stable and can be fabricated using cost-effective methods. The sandwich-type ZnO/carbon tape/ZnO nanogenerators (ZCZ-NGs) were fabricated in a cost-effective way by depositing ZnO layers on indium tin oxide (ITO)-coated polyethylene naphthalate (PEN) substrates in a radio frequency (RF) magnetron sputtering system to form ZnO/ITO/PEN-stacked blocks as well as using a conductive double-sided adhesive carbon tape to bond two ZnO/ITO/PEN blocks together, appreciably reducing the overall fabrication time and processing steps. The output voltage and current of the fabricated ZCZ-NG devices were measured for various bending strain rates, device sizes, and thicknesses of ZnO layers, generating up to about 30 V in terms of the peak-to-peak output voltage, which is much higher than those of other similar sandwich-type piezoelectric nanogenerators. Moreover, the output voltage variations of various ZCZ-NG devices due to their ZnO thicknesses, bending strain rates, and device sizes were investigated through the observation of their output voltages. An eco-friendly and stretchable flexoelectricity-enhanced piezoelectric nanogenerator (F-PENG) based on zinc-aluminum layered double hydroxide nanosheets (ZnAl:LDH Ns)-ZnO heterostructure is demonstrated on the stretchable polydimethylsiloxane (PDMS) substrates as the fabrication of a high-performance piezoelectric nanogenerator (PENG) with high stretchability and durability is desirable for the next-generation of stretchable and wearable electronics. The vertically-oriented eco-friendly ZnAl:LDH Ns were facilely synthesized by dipping the 10 wt% aluminum-doped zinc oxide (AZO) thin films in deionized (DI) water at room temperature. The enhanced output performance of the F-PENG is demonstrated under tapping, bending, and stretching modes, and is attributed to the synergistic flexoelectric and piezoelectric effects. The achieved maximum output power density of F-PENG under tapping is ~2.7 μW/cm2. The pressure-sensing capability of the F-PENG is demonstrated by the generated outputs under the three applied modes. In addition, the biomechanical energy harvesting capability of the F-PENG is demonstrated by subjecting it to various biomechanical motions. The F-PENG exhibits excellent mechanical durability in all three modes of operation. The present study not only paves the way toward the facile fabrication of stretchable and high-performance F-PENG with combined flexoelectric and piezoelectric effects but also validates a wide range of applications in the next generation of stretchable and wearable electronics.
Ph.D. Defense
5/31 (Tue) 3:30 PM
국문: 보안성 높은 무선통신을 위한 적대적 기계학습 기반의 도청자 회피 기법
English: Machine Learning Algorithms for Sparse Supervision
서중하 (Jungha Seo)
IT융합빌딩(N1) 709호 세미나실 & Online:
Abstract
Acquiring secure communication link between legitimate receiver and transmitter has been an important issue for decades. Especially for wireless communication systems, physical layer security (PLS), which utilizes channel state information to achieve information-theoretic secrecy, is applied along with cryptography which takes place in the upper layer by encrypting the data with secret key. Recently, adversarial signal design, one of the machine learning techniques that fools the pretrained deep learning model, has been applied for deep learning-aided wireless communication systems to ensure the security, referred to as secure, yet adversarial signal. The objective of this secure, yet adversarial signal is to design an adversarial signal that lets the legitimate receiver classify accurately while the eavesdropper misclassifies. While such adversarial signal design can be thought of as joint-PLS-cryptography approach assuming coherent DL-based receiver, huge computational complexity for generating the adversarial signal prevents this coherence, i.e., channel becomes outdated while crafting the adversarial signal. We investigate to reduce this computational complexity via two schemes: (i) meta perturbation; (ii) universal surrogate model. Key idea of (i) the meta perturbation is to find (meta-learn) a good initialization of the perturbation signal for the adversarial signal to ensure coherent, secure communication between the legitimate receiver and the transmitter and (ii) the universal surrogate model is to learn parameters of a generative surrogate model in lieu of target models for crafting transferable secure, yet adversarial signal. Numerical results verify that the proposed approaches reduce computational complexity of the secure, yet adversarial signal design with a reasonable performance degradation.
Refer to TA announcement email for zoom link
Ph.D. Defense
5/30 (Mon) 3:30 PM
국문: 극한의 모션을 갖는 영상에서의 비디오 프레임 보간법 연구
English: Extreme Video Frame Interpolation: Extending VFI to video with extremely large motion
심현준 (Hyeonjun Sim)
Online:
Abstract
Video Frame Interpolation (VFI) temporally interpolates one or more intermediate frames between every two consecutive frames with temporal coherence so that smoothly rendered fast motion is more visually pleasing. However, VFI is a long-standing and challenging task in computer vision, which is attributed to several factors such as non-linear motion, deformable object motion, large motions and motion blur in the video sequences. In this dissertation, we focus on the challenging issue of fast motions with extremely large pixel displacements which often occur in ultra-high-definition (UHD) video sequences and thus result in severe motion judder. To study VFI with large motion, we present an extreme VFI network, called XVFI-Net, that first handles the VFI for 4K videos with large motion. Even after training of the XVFI-Net, input frames can be flexibly down-scaled into any smaller size to cope with the spatial resolution and the degree of motion magnitudes of input frames during inference time, unlike the previous pyramid structures of a fixed number of scale levels. The recursive multiscale shared structure of XVFI-Net allows for the large motion to be effectively captured in flexibly small-scale levels. In order to interpolate middle frames at any intermediate time instances, the bilateral optical flows are stably approximated by a novel complementary flow reversal (CFR) technique using the bidirectional motion. We also propose a coarse-direction-and-fine-attention (CDFA) module to extend the XVFI-Net, where multi-hypothesis flow vectors are utilized to deal with complex motion boundaries of fast-moving objects. The coarse direction vector is supervised by a self-guided training scheme. Then the fine attention vectors attend to the local details based on the coarse direction vector in a residual learning manner. Extensive experimental results show that our algorithms can successfully capture the essential information of objects with extremely large motions and complex textures while state-of-the-art methods exhibit poor performance. Furthermore, our XVFI-Net framework also performs comparably on the previous lower resolution benchmark dataset, which shows the robustness of our algorithm as well.
Refer to TA announcement email for zoom link
Ph.D. Defense
5/30 (Mon) 3:00 PM
국문: 비디오 모델을 위한 주파수 선택적 정규화 기법 연구
English: Frequency selective regularization methods for video model
김진형 (Jinhyung Kim)
Online:
Abstract
Recently, various regularization techniques for deep neural network based video models have been studied. From the perspective of supervised learning, it is known that deep neural network models for video action recognition are easily prone to overfitting to training data due to a large amount of parameters, and regularization techniques can be one solution to alleviate the problem. From the perspective of unsupervised learning (or self-supervised learning), a group of data augmentation method, which is a type of regularization method, is used in various ways as an essential element of the contrastive learning method that is being actively studied. In this study, we propose frequency-selective regularization techniques for supervised/unsupervised learning of video models. First, in order to solve the overfitting problem of the video action recognition model, we propose a regularization technique in which small random changes are made to low-frequency signals in the feature stage. Second, we propose a data augmentation method in which the video model arbitrarily filters the spatiotemporal low-frequency signal from the input video signal in order to learn a better representation through contrastive learning. Through these frequency-selective regularization techniques, it can be confirmed that the video model improves the action recognition performance in the target task without using additional training data.
Refer to TA announcement email for zoom link
Ph.D. Defense
5/30 (Mon) 2:00 PM
국문: 저지도 상황에서의 기계학습 알고리즘 연구
English: Machine Learning Algorithms for Sparse Supervision
서준 (Jun Seo)
Online:
Abstract
The recent advance in artificial intelligence technology has been attributed to the exponential increase in the amount of data. On the other hand, however, it is becoming more difficult to categorize or label the data for training of artificial intelligence model. In this sense, machine learning algorithms in low-supervised situation with little or no labeled data are becoming increasingly important. This thesis proposes sparse supervision machine learning algorithms using only a few labeled data or utilizing unlabeled data. In the first part of thesis, we focus on few-shot learning method that trains the model to classify new classes with only a few labeled data. The proposed few-shot learning model has the ability to quickly adapt to new tasks through a linear projection of the feature space. In the second part, we focus on few-shot segmentation, which aims to conduct semantic segmentation for a new class with only a small amount of data. We propose a few-shot segmentation method utilizing existing semantic segmentation model by transforming the unfamiliar novel feature into more comprehensible form based on few labeled data. Finally in the third part, we focus on self-supervised learning that trains the useful representation using unlabeled data. We propose an contrastive self-supervised learning methods utilizing nonlinear feature transformation via the self-attention technique to overcome the limitation of existing contrastive self-supervised learning.
Refer to TA announcement email for zoom link
Ph.D. Defense
5/30 (Mon) 2:00 PM
국문: 신뢰적인 나노 공정에 기반한 고선형 고속 동작하는 팔라듐 수소 센서
English: Highly Linear and Ultra-fast Pd Hydrogen Gas Sensor Exploiting Reliable Nanowire Fabrication
조민승 (Min-Seung Jo)
나노종합기술원(E19) 517호
Abstract
Hydrogen has been attracted as an eco-friendly energy source, however, careful management is required because of its high explosiveness. In particular, palladium is widely used as a sensing material for hydrogen sensors thanks to the advantage of selectively reacting with hydrogen. When reacting with high concentration of hydrogen, phase transition occurs in palladium hydride along with volume expansion and the saturation of hydrogen in palladium, resulting in low linearity and durability. In this study, a high-linearity, high-speed operation, and high-durability hydrogen sensor was proposed and verified through engineering a palladium nanostructure.
Ph.D. Defense
5/24 (Tue) 5:00 PM
국문: 학습된 로컬 커널을 활용한 단일 이미지 및 양방향 참조 기반 비디오 초해상화
English: Single Image and Bidirectional Reference-based Video Super-Resolution with Learned Local Kernels
김수예 (Soo Ye Kim)
Online:
Abstract
Upscaling image and video resolutions, also commonly termed as super-resolution, has various applications, including increasing the resolution of legacy images and videos for viewing on modern displays, or pre-processing various images and videos (e.g., natural images, medical images, surveillance videos) before performing other computer vision tasks such as detection or classification. Similar to other computer vision problems, unprecedented image/video quality has been achieved in super-resolution by employing deep learning models and training them on abundant paired datasets. However, three fundamental challenges still exist for deep learning-based super-resolution models: (i) exploiting spatially variant characteristics in the low-resolution input image/video, (ii) generalization of deep learning models for super-resolution, and (iii) overcoming the highly ill-posed nature in super-resolution. This dissertation aims to tackle these issues, presenting two super-resolution frameworks: a blind single image superresolution method and a bidirectional reference-based video super-resolution method. The blind super-resolution model addresses generalization of super-resolution models by jointly learning the spatially variant degradation kernels and restoration kernels. The bidirectional reference-based video super-resolution model exploits the spatial redundancy in videos by assuming bidirectional (past and future) reference frame inputs when performing video super-resolution. This allows to narrow down the high-resolution solution space for the low-resolution frames in the same shot since video frames are likely to be highly redundant within the shot. Both the proposed single image and video super-resolution models incorporate local kernels in the model design so that spatially variant operations can be applied on images, video frames, and feature maps. The proposed models are analyzed and validated with extensive experiments including comparison to state-of-the-art methods, various kernel visualizations, and ablation studies.
Refer to TA announcement email for zoom link
Ph.D. Defense
5/24 (Tue) 4:00 PM
국문: 영률이 낮은 응률 저감층 기반 신축 플랫폼을 활용한 신축 유기 발광 다이오드의 연구
English: Study of stretchable organic light-emitting diodes with a stretchable platform based on low Young’s modulus strain relief layers
김태현 (Taehyun Kim)
Online:
Abstract
In this thesis, a stretchable organic light-emitting diode (OLED) was realized using a stretchable platform based on a strain relief layer with low Young's modulus. The stretchable platform, on which OLEDs are stacked, consists of rigid islands connected electrically and mechanically with serpentine-shaped interconnectors. When a serpentine-shaped interconnector is stretched, out-of-plane buckling is occurred to reduce the strain applied to the interconnector. However, when interconnectors attached to elastomeric substrates are stretched, the vertical movement is restricted, and the applied strain tends to increase. To solve this problem, a strain relief layer having a very low Young's modulus is inserted between the elastomeric substrate and the rigid islands. Due to the very low Young's modulus of the strain relief layer, the serpentine interconnector has more freedom to move in a vertical direction, enabling the overall system to withstand a higher degree of stretching. Stretchable OLEDs are realized on this platform, and it was confirmed that stable operation is possible even when it is stretched to a significant degree. First-generation stretchable OLEDs are fabricated with a photo-patternable photoresist, SU-8, enabling stretchability is relatively low due to the limitation of the material property. Therefore, polyimide (PI), a widely used plastic substrate in flexible OLED displays, is used to fabricate a more stable stretchable platform. To pattern this, commercially accessible laser patterning was used, and to secure a surface at a level where OLEDs can be deposited, a photoresist was coated to protect the surface during patterning. After this, poly(glycidyl acrylate-co-2-(dimethylamino)ethyl methacrylate (pGAD) was deposited to further increase the degree of planarization of the surface, enabling the fabrication of an OLED that has very stable properties even on the patterned PI. In addition, a multi-layer thin-film encapsulation was applied to increase stability in the external environment, leading to the realization of a passive matrix stretchable OLED that can be operable in ambient air.
Refer to TA announcement email for zoom link
Ph.D. Defense
5/24 (Tue) 10:30 AM
국문: 고효율 및 고유연 양자점-유기 고분자 하이브리드 태양전지 제작에 관한 연구
English: A study on the fabrication of highly efficient and flexible quantum dot-organic polymer hybrid solar cells
김창조 (Changjo Kim)
KI빌딩 5F Lecture room(E4 B501호)
Abstract
Colloidal quantum dots (CQDs) are promising materials for next-generation flexible photovoltaic devices because of near-infrared absorption, facile bandgap tunability, and superior air stability. However, their flexibility has not been developed yet to be used in wearable devices due to the brittle CQD film. Various attempts have been proceeded to form a composite film with a polymer to overcome the weak QD-QD bonding, but power conversion efficiency (PCE) has inevitably decreased. In this thesis, we propose a method to fabricate highly efficient PbS CQD solar cells via two-step post annealing process. Also, we propose a method to improve the mechanical properties while maintaining the high PCE of PbS CQD solar cells. We introduced 3-Aminopropyltriethoxysilane (APTS) on CQD thin films to strengthen the dot-to-dot bonding. As a result, the PCE of 11.04% and the high mechanical reliability of 88% under 12,000 cycles at bending radius of 8.3 mm was achieved, which are the highest in flexible PbS CQD solar cells. In this thesis, we also propose a novel AgBiS2 NC/organic hybrid concept to fabricate highly efficient device structure. PCE of 9.1% was achieved by introducing NC/BHJ (PBDB-T-2F:BTP-4Cl) hybrid structure via complementary absorption.
Ph.D. Defense
5/23 (Mon) 10:30 AM
국문: 에지 컴퓨팅 환경에서 IoT 기기를 위한 딥러닝 기반 에너지 관리 프레임워크
English: Deep Learning-based Energy Management Framework for IoT Devices in Edge Computing Environment
한재섭 (Jaeseob Han)
정보전자공학동(E3-2) 3222호
Abstract
Recently, various cutting edge Internet of Things (IoT) services and solutions markets such as smart home, energy, and manufacturing rapidly increase in size. In most cases, continuous monitoring application is expect to be the main compone application is expect to be the main component which these IoT solutions and services are being nt which these IoT solutions and services are being provided to users. In order to provide a continuous monitoring service stably in such an environment, provided to users. In order to provide a continuous monitoring service stably in such an environment, multiple IoT devices should be deployed, and a huge amounts of data are generated and collected multiple IoT devices should be deployed, and a huge amounts of data are generated and collected from tfrom these IoT devices. Once these IoT devices are deployed in the environment, it is difficult to hese IoT devices. Once these IoT devices are deployed in the environment, it is difficult to change the configuration of them. Also, it is difficult to supply power to all the IoT devices, in most change the configuration of them. Also, it is difficult to supply power to all the IoT devices, in most cases. Therefore, typical IoT devices are generally suppliecases. Therefore, typical IoT devices are generally supplied energy from a capacityd energy from a capacity--constrained constrained battery. Furthermore, IoT devices transmit measurement data in f uniform period fashion in order battery. Furthermore, IoT devices transmit measurement data in f uniform period fashion in order to continuously provide a monitoring service. Therefore, energy saving of IoT devices has lately to continuously provide a monitoring service. Therefore, energy saving of IoT devices has lately become a critical issue ibecome a critical issue in diverse IoT applications due to their limited battery capacity. Particularly, n diverse IoT applications due to their limited battery capacity. Particularly, unnecessary energy cost for redundant data transmission such as transmitting duplicate or similar unnecessary energy cost for redundant data transmission such as transmitting duplicate or similar data generally happens. Therefore, based on these mentioned problems, an deep data generally happens. Therefore, based on these mentioned problems, an deep learninglearning--based based energy management framework method for IoT devices in an edge computing system environment energy management framework method for IoT devices in an edge computing system environment is newly proposed. To reduce energy consumption of multiple IoT devices, the transmission period is newly proposed. To reduce energy consumption of multiple IoT devices, the transmission period of data transmitted from IoT devices is dynamically aof data transmitted from IoT devices is dynamically adjusted by considering both prediction for djusted by considering both prediction for imputation error of unimputation error of un--transmitted data and energy consumption of IoT devices, simultaneously. transmitted data and energy consumption of IoT devices, simultaneously. The proposed imputation error prediction module is designed using various deep learning model The proposed imputation error prediction module is designed using various deep learning model structures and the optistructures and the optimal transmission period value is obtained through an optimization problem mal transmission period value is obtained through an optimization problem considering both the prediction value of imputation error and the numerically modeled energy considering both the prediction value of imputation error and the numerically modeled energy consumption of the IoT device. Furthermore, the proposed transmission period control framconsumption of the IoT device. Furthermore, the proposed transmission period control framework ework is implemented in the configured testbed environment and checks whether the proposed framework is implemented in the configured testbed environment and checks whether the proposed framework can be developed in the practical environment. In addition, radio frequencycan be developed in the practical environment. In addition, radio frequency--based energy harvesting based energy harvesting technology is considered for the energy supply of IoT dtechnology is considered for the energy supply of IoT devices. When a large number of IoT devices evices. When a large number of IoT devices are distributed in the OFDMA environment, the minimum harvesting energy of IoT devices should are distributed in the OFDMA environment, the minimum harvesting energy of IoT devices should be optimized through efficient wireless channel and transmission power resource allocation to be optimized through efficient wireless channel and transmission power resource allocation to prolong the lifetime of Iprolong the lifetime of IoT devices. Therefore, two resource allocation methodologies that optimizes oT devices. Therefore, two resource allocation methodologies that optimizes the harvesting power of IoT devices has been studied. First one is a method to maximize the the harvesting power of IoT devices has been studied. First one is a method to maximize the minimum harvested power of an IoT device, and the other one maximizes the harvested poweminimum harvested power of an IoT device, and the other one maximizes the harvested power of r of each IoT device while considering the minimum harvested power constraint. In the both resource each IoT device while considering the minimum harvested power constraint. In the both resource allocation methodologies, the amount of DC power converted from RF power is considered as a allocation methodologies, the amount of DC power converted from RF power is considered as a nonnon--linear model for practical energy harvesting result.linear model for practical energy harvesting result.
Ph.D. Defense
5/20 (Fri) 1:00 PM
국문: 저전력, 저휘도의 수직 적층된 유기발광다이오드로 제작된 유기 산소포화도센서 구현에 관한 연구
English: A Study on Realization of Pulse Oximetry Sensors Consist of Vertically-stacked OLED and OPD with Low Power and Low Luminance
이승희 (Seunghee Lee)
Online:
Abstract
In the past, it has been a challenge to adequately provide a personal medical service to human beings worldwide due to the lack of advanced technologies. Nowadays, however, along with the drastically increased number of personal mobile devices, such as the above 95% penetration rate in 2020 in Korea, people can easily access the wearable devices in a wide range, allowing the combination with real-life establishing an active medical culture. This circumstance enables the scientist to provide advanced personal medical service with mobile and wearable devices. Among them, a non-invasive SpO2 sensor can be one of the promising candidates for advanced personal medical devices because of its simple, painless, and cost-effective measurement method. In addition, it has a great potential to keep the consumer away from severe diseases, especially heart disease and cerebrovascular disease, which are among the leading causes of death, by simply monitoring the amount of oxygen-saturated hemoglobin relative to total hemoglobin in the blood. But, it still needs to improve several weaknesses, such as accuracy, skin irritation, and efficiency, to use these non-invasive biosensors commercially. This study develops a miniaturized reflective heart rate and pulse oximetry sensor with a new form factor using organic materials as light-emitting and light-receiving parts. Unlike the semiconductor light-emitting diodes and silicon-based photodiodes commercial oximetry sensors, our newly developed organic flexible thin-film sensor helps minimize the mechanical stress and enhance the signal-to-noise ratio compared to rigid electronics providing conformal contact with the skin surface. Based on the skin optical modeling method, the building structure on the skin surface is constructed to have the optimal light receiving efficiency for two wavelengths. At the same time, it enables to maximize the area of organic light-emitting diodes (OLEDs) and minimize the area of the organic photodiode by implementing vertically-stacked OLEDs using the same common electrode in the light-emitting part. Furthermore, a ring-shaped circular structure light-emitting part is designed to minimize the power reduction based on the distance and the light spreading considering in all directions. With an efficient structural designing, the accurate signals are detected even with a small current consumption of a few microampere, and in terms of luminance, photoplethysmogram signals are also extracted with a very small number of photons per area (13~ cd/m2). In conclusion, through a lower driving current and a large area design with the tandem structure of OLED, we successfully decrease the luminance by more than 100 times compared to the previous reports. Given that the luminance is the biggest determining-factor of device lifetime, which is a very important part of wearable health monitoring, a low-power, low-brightness organic pulse oximeter, what we fabricate, provides the light information efficiently collected through the sensor and contributes to developing a next-generation of the healthcare monitoring system that requires continuous observation.
Refer to TA announcement email for zoom link
Ph.D. Defense
5/17 (Tue) 2:00 PM
국문: 고대역폭 메모리 전력분배망 최적화를 위한 트랜스포머 네트워크 기반 강화학습 방법론
English: Transformer Network-based Reinforcement Learning Method for Optimization of Power Distribution Network (PDN) in High Bandwidth Memory (HBM)
박현욱 (Hyunwook Park)
Online:
Abstract
In this article, for the first time, we propose a transformer network-based reinforcement learning (RL) method for power distribution network (PDN) optimization of high bandwidth memory (HBM). The proposed method can provide an optimal decoupling capacitor (decap) design to maximize the reduction of PDN self- and transfer impedance seen at multiple ports. An attention-based transformer network is implemented to directly parameterize decap optimization policy. The optimality performance is significantly improved since the attention mechanism has powerful expression to explore massive combinatorial space for decap assignments. Moreover, it can capture sequential relationships between the decap assignments. The computing time for optimization is dramatically reduced due to the reusable network on positions of probing ports and decap assignment candidates. This is because the transformer network has a context embedding process to capture meta-features including probing ports positions. In addition, the network is trained with randomly generated data sets. Therefore, without additional training, the trained network can solve new decap optimization problems. The computing time for training and data cost are critically decreased due to the scalability of the network. Thanks to its shared weight property, the network can adapt to a larger scale of problems without additional training. For verification, we compare the results with conventional genetic algorithm (GA), random search (RS), and all the previous RL-based methods. As a result, the proposed method outperforms in all the following aspects: optimality performance, computing time, and data efficiency.
Refer to TA announcement email for zoom link
Ph.D. Defense
5/10 (Tue) 10:30 AM
국문: 광통신 도달 거리 연장을 위한 프레임 기반 펄스 폭 변조 고속 송수신기
English: High-Speed Framed-Pulsewidth Modulation Transceiver for Extended Reach Optical Links
권우현 (Woohyun Kwon)
나노종합기술원 백두산세미나실(E19, S-321호)
Abstract
This paper proposes a high-speed framed-pulsewidth modulation (FPWM) transceiver that reduces signal loss due to optical fiber dispersion as a method to improve frequency spectrum efficiency in optical communication. An additional 75% of coding gain is obtained through proposed time-domain modulation. It increases the minimum pulsewidth by 1.75 times compared to NRZ, lowering the Nyquist frequency to 4/7. The FPWM signaling not only reduces losses in the metal channel, but also suppresses the penalty due to chromatic dispersion in the optical fiber proportional to the square of bandwidth. Low power and latency modulator and de-modulator are implemented based on successive approximation and weighted sum algorithm each. Measurements under back-to-back condition using an externally modulated laser and avalanche photo diode show that the FPWM signal has 6dB better receiver sensitivity at a bit error rate of 2e-5 than PAM-4 at 26Gbps. FPWM outperforms NRZ by 8dB over 15km of singlemode fiber and PAM-4 by 4dB at 20km of SMF. Thus, proposed IC is possible to send a longer reach compared to the existing method under the same optic conditions. The full chip occupies 2.2x2.0mm2 with bi-directional two lanes and two PLLs. It consumes 262mW per lane from a 0.9V supply. The random jitter of the PLL measured at Tx clock pattern is 265fs, rms. The testchip prototype is fabricated in a 28-nm CMOS process and packaged in a flip-chip chip scale package.
Ph.D. Defense
4/29 (Fri) 1:00 PM
국문: 광위상배열 칩 기반의 고속 무선전송모듈 개발 및 상용화
English: Commercialization of high-speed wireless optical transmitter module based on optical phased array chip
이현우 (Hyunwoo Rhee)
정보전자공학동(E3-2) 2233호 & Online:
Abstract
Optical wireless communication (OWC) is an alternative technology to overcome RF communication, which is reaching the limit of bandwidth due to the continuous growth of communication traffic. OWC technology, which has a much higher carrier frequency than RF or mm-wave, can be a fundamental solution to meet the recent large bandwidth requirements such as 8K TV, mobility, metaverse and 6G. By using optical phased array (OPA), it is possible to implement a beam-forming OWC that steers a laser signal with high-speed modulated data in a desired direction. In particular, silicon-based OPA has many advantages in terms of size, power consumption, and steering range. In this research, we propose an optical wireless transmission system using OPA. We demonstrate 32 Gbps pseudo-random bit sequence (PRBS) data transmission based on silicon OPA by combining an electrooptic phase shifter and a thermo-optic emitter for two-dimensional beam steering by selecting suitable components for the transmitter and receiver. In order to commercialize high-speed data transmission using OPA, we further research on the optical packaging method that can modularize the developed OPA chip. We developed a new technology called direct optical wire (DOW) bonding which uses polymer to create an optical waveguide that can be fabricated in free space. This technology interconnects optical chips in a very similar way to metal wire bonding, and can solve the problems in existing optical packaging methods which is low yield, large volume, and high cost. The wireless transmission module is developed by analyzing the advantages and disadvantages of other optical packaging technologies applicable to the OPA chip. Finally, 8K video data transmission using the OPA chip based wireless optical transmission module was demonstrated for proving that it can be applied to commercial systems.
Refer to TA announcement email for zoom link
Ph.D. Defense
4/26 (Thu) 9:00 AM
국문: LTE 제어 평면 프로토콜에서의 부채널을 통한 프라이버시 침해 공격 및 대응기술 연구
English: A study of privacy leakage using unwanted side channel in LTE control plane
배상욱 (Sangwook Bae)
Online:
Abstract
Cellular networks have become a dominant means of serving a variety of applications, including, not limited to the call service, but also our professional and private activities. This trend gives rise to the exchanged data through the cellular network, which is private information that people are reluctant to share as it can reveal their political, financial, and personal interests. As this motivates a wide range of attacks, conducting a security analysis of the cellular network design and its implementation is essential. In this dissertation, we analyze the unwanted side channels in the three-domain of LTE networks; downlink channel, uplink channel, and implementation of baseband processors in mobile devices. As a result, although the security mechanisms in LTE preserve the subscriber’s privacy, we argue that a)unwanted side channels do exist in the cellular network due to the unprotected LTE control plane protocol’s lower layers and diversities in the UE implementation, and b) those side channels can be utilized to the privacy threatening attacks. Especially, this dissertation focuses on discussing side-channel attacks using leaked information in unprotected layers in three different domains with practical considerations. First, we investigate the information leakage in the downlink channel in LTE and show that an unprivileged third party can exploit the information leakages in the downlink channel for identifying a video title that a victim is watching. Second, we conduct a security analysis on the uplink channel in LTE. We show that an adversary who has no access to victims’ devices or cell towers can locate the target victim by using the side channel in the design of the uplink channel. Third, we analyze the differences in the baseband implementation. We then present the UE identification attack which infers the name of the device and conduct the large-scale evaluation using 80 smartphones and 16 IoT devices to show its practicality in the real world. Lastly, we conduct a case study of utilizing the side channels in LTE for defeating voice phishing eco-system. In summary, although those security mechanisms in LTE preserve the subscriber’s privacy, we argue that a)unwanted side channels do exist in the cellular network due to the unprotected LTE control plane protocol’s lower layers and diversities in the UE implementation, and b) those side channels can be utilized to the privacy threatening attacks. Thus, this work concretely enhances the security of the current and future mobile network generations.
Refer to TA announcement email for zoom link
Computing Seminar
4/22 (Fri) 2:00 PM
MyDJ: Sensing Food Intakes with an Attachable on Your Eyeglass Frame (ACM CHI'22)
신재민 (Jaemin Shin)
Online:
Abstract
Various automated eating detection wearables have been proposed to monitor food intakes. While these systems overcome the forgetfulness of manual user journaling, they typically show low accuracy at outside-the-lab environments or have intrusive form-factors (e.g., headgear). Eyeglasses are emerging as a socially-acceptable eating detection wearable, but existing approaches require custom-built frames and consume large power. We propose MyDJ, an eating detection system that could be attached to any eyeglass frame. MyDJ achieves accurate and energy-efficient eating detection by capturing complementary chewing signals on a piezoelectric sensor and an accelerometer. We evaluated the accuracy and wearability of MyDJ with 30 subjects in uncontrolled environments, where six subjects attached MyDJ on their own eyeglasses for a week. Our study shows that MyDJ achieves 0.919 F1-score in eating episode coverage, with 4.03× battery time over the state-of-the-art systems. In addition, participants reported wearing MyDJ was almost as comfortable (94.95%) as wearing regular eyeglasses.
Refer to TA announcement email for zoom link
Ph.D. Defense
4/14 (Thu) 4:00 PM
국문: 클라우드 서비스 시스템의 보안성 강화를 위한 하드웨어 기반 격리 기술 활용에 대한 연구
English: Enhancing Security of Cloud Service Systems by Leveraging Hardware-based Isolation Technology
한주형 (Juhyeng Han)
IT융합빌딩(N1) 816세미나실 & Online:
Abstract
Hardware-based isolation technologies have been utilized within a cloud environment to satisfy the demands for secure and scalable security services. However, a naïve adoption of hardware-based isolation fails to protect on-cloud services against emerging security threats arising from the untrusted nature of a cloud environment. Also, it is challenging to apply hardware-based isolation to legacy on-cloud services while considering security, flexibility, scalability, and performance, as this requires expertise in multiple domains. In this dissertation, we argue that providing high-level abstractions of hardware-based isolation encapsulating security-sensitive components helps in implementing secure and scalable on-cloud services. To substantiate our claim, we present the design and implementation of two security service systems. First, we show EVE, a secure middlebox framework that enables visibility on encrypted traffic over multiple encryption protocols. EVE securely processes encrypted traffic leveraging a combination of trusted execution environment (TEE) and software security technology, and provides high-level abstractions for the secure middlebox processing. EVE abstractions relieve engineering efforts while supporting diverse use cases with multiple encryption protocols. Next, we propose ScaleTrust, a scalable and secure key management system that virtualizes cloud-backed hardware security modules (HSMs) using a TEE. The virtual HSM partition abstracts away the details of secure key management such as secure channel establishment and the verification of isolated key usages. By supporting the isolation of each virtual HSM partition, ScaleTrust provides multi-tenancy scalability and security against insider threats in an untrusted cloud environment.
Refer to TA announcement email for zoom link
Computing Seminar
3/25 (Fri) 2:00 PM
Towards Large-Scale Machine Learning
손지용 (Jy-yong Sohn)
Online:
Abstract
In this talk, I share my recent works for enabling large-scale machine learning (ML). In the first part, I discuss the path towards enabling “trustworthy ML” by making algorithms robust and fair, which is challenging for large-scale models. To be specific, I present GenLabel, a mixup relabeling scheme that improves generalization and robustness by making use of the underlying data distribution learned by generative models. In the second part, I discuss the path towards enabling “efficient ML” by pruning/compressing neural networks, which will take a significant role for deploying large-scale ML algorithms. Specifically, I present Gem-Miner, the first algorithm that finds lottery tickets at initialization and passes existing sanity checks. The two papers presented in this talk are on arxiv:
[GenLabel] https://arxiv.org/pdf/2201.02354.pdf
[Gem-Miner] https://arxiv.org/pdf/2202.12002.pdf
Refer to TA announcement email for zoom link