A5039718583- Advanced Image Processing Techniques
- Adversarial Robustness in Machine Learning
- Advanced Memory and Neural Computing
- Image and Signal Denoising Methods
- Advanced Image Fusion Techniques
- Analog and Mixed-Signal Circuit Design
- Semiconductor materials and devices
- IoT-based Smart Home Systems
- Domain Adaptation and Few-Shot Learning
- COVID-19 diagnosis using AI
- Smart Grid Energy Management
- Explainable Artificial Intelligence (XAI)
- Advanced DC-DC Converters
- Water Systems and Optimization
- Image Processing Techniques and Applications
- Radio Frequency Integrated Circuit Design
- Anomaly Detection Techniques and Applications
- VLSI and Analog Circuit Testing
- Health, Environment, Cognitive Aging
- Advanced Neural Network Applications
- Privacy-Preserving Technologies in Data
- Photoreceptor and optogenetics research
- Neural Networks and Reservoir Computing
- Caching and Content Delivery
- E-commerce and Technology Innovations
Samsung (South Korea)
2022-2023
Sungkyunkwan University
2018-2023
Yonsei University
2018-2022
Non-intrusive load monitoring (NILM), also known as energy disaggregation, is a blind source separation problem where household’s aggregate electricity consumption broken down into usages of individual appliances. In this way, the cost and trouble installing many measurement devices over numerous household appliances can be avoided, only one device needs to installed. The has been well-known since Hart’s seminal paper in 1992, recently significant performance improvements have achieved by...
We ask whether the neural network interpretation methods can be fooled via adversarial model manipulation, which is defined as a fine-tuning step that aims to radically alter explanations without hurting accuracy of original models, e.g., VGG19, ResNet50, and DenseNet121. By incorporating results directly in penalty term objective function for fine-tuning, we show state-of-the-art saliency map based interpreters, LRP, Grad-CAM, SimpleGrad, easily with our manipulation. propose two types...
This article proposes practical design techniques to enhance performance and reliability of 1024 GB/s high-bandwidth memory-3 (HBM3). Effective data-bus methods are applied transfer data from multi-bank a with sufficient fetch margin. A symbol-based on-die error-correcting code (OD-ECC) correct 16-bit error, bounded by sub-wordline (WL), parallelized inversion (DBI) implemented. Error check scrub (ECS) mode repair capability (RCC) an internal serial interface designed support system...
A time-domain-control-based quasi-V <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> buck converter with small area, fast transient response, wide load current range, and constant switching frequency is proposed in this paper. The time-domain-based controller achieves compact core area simplified control by replacing the conventional voltage-domain comparator a phase detector having an adaptive detection window. Moreover, coupling-based...
Modeling spatiotemporal brain dynamics from high-dimensional data, such as functional Magnetic Resonance Imaging (fMRI), is a formidable task in neuroscience. Existing approaches for fMRI analysis utilize hand-crafted features, but the process of feature extraction risks losing essential information scans. To address this challenge, we present SwiFT (Swin 4D Transformer), Swin Transformer architecture that can learn directly volumes memory and computation-efficient manner. achieves by...
The convolutional neural network (CNN)-based super-resolution (SR) has shown outstanding performance in the field of computer vision. implementation inference hardware for CNN-based SR suffered from intensive computation with severely unbalanced load among layers. Various light-weighted networks have been researched little degradation. However, hardware-efficient dataflow is also required to efficiently accelerate within limited resources. In this article, we propose that reduces by...
A time-interleaved duobinary encoding scheme for the low-power high-bandwidth memory (HBM) I/O interface is proposed with a 65-nm CMOS process. To reduce power consumption in HBM using multiple through-silicon via (TSV) I/Os, transmitter (TX) that performs signaling voltage-mode driver proposed. small area encoder implemented to generate output and an edge-boosted pre-driver improve slew rate robustness against process, voltage, temperature (PVT) variations of driver. convert signal into...
Neural network interpretation methods, particularly feature attribution are known to be fragile with respect adversarial input perturbations. To address this, several methods for enhancing the local smoothness of gradient while training have been proposed attaining robust attributions. However, lack considering normalization attributions, which is essential in their visualizations, has an obstacle understanding and improving robustness methods. In this paper, we provide new insights by...
This paper proposes a convolutional neural network (CNN)-based super-resolution accelerator for up-scaling to ultra-HD (UHD) resolution in real-time edge devices. A novel error-compensated bit quantization is adopted reduce depth the SR task. Spatially independent layer fusion exploited satisfy high throughput requirements at UHD by increasing parallelism. Burst operation with write mask dual-port SRAM increases process element utilization allowing concurrent multi-access without exploiting...
Low power consumption is very important for Internet of Things (IoT) achieving long lifetime. On the other hand, non-volatile memory such as Flash or Ferroelectronic Random Access Memory (FRAM) has been commercially produced reduction in micro control unit (MCU). In this paper, MCU based on and FRAM are compared analyzed terms through IoT implementation. For practical analysis, heating, ventilation, air conditioning (HVAC) implemented measured MCU. From measurement results, time shutdown...
We propose DoPAMINE, a new neural network based multiplicative noise despeckling algorithm. Our algorithm is inspired by Neural AIDE (N-AIDE), which recently proposed adaptive image denoiser. While the original NAIDE was designed for additive case, we show that same framework, i.e., adaptively learning pixel-wise affine denoisers minimizing an unbiased estimate of MSE, can be applied to case as well. Moreover, derive double-sided masked CNN architecture control variance activation values in...
This paper presents comparative analysis of digital STDP learning circuits designed using counter and shift register for spiking neural network (SNN). In addition, it is possible the implemented to operate when two or more spikes occur at same time. The conventional Von Neumann architecture has limitations such as speed bottleneck because deep requires a lot parallel data processing. order solve this problem, researches on SNN similar human have been widely performed. paper, unsupervised...
In this paper, a highly accurate, fully digital temperature sensor with curvature correction scheme is proposed. Conventional analog sensors are complex and require large area. Digital simple small area, but they have inaccuracies due to process variations errors. particular, errors become more severe as the technology scales down. Thus, accurate method proposed that can be used even in latest nodes. The achieves lower smaller area than conventional by using correction. A error occurs up 17...
Artificial intelligence networks have been researched in many fields such as computer vision, health care, and military service. Convolutional neural network (CNN) is one of the basic that uses convolutional operations basis to train data perform desired application. However, important parameters used CNN applications suffer from security issues. Thus, need for protection increasing. The traditional approach storing training weight encrypted external memory has disadvantage large hardware...
We propose DoPAMINE, a new neural network based multiplicative noise despeckling algorithm. Our algorithm is inspired by Neural AIDE (N-AIDE), which recently proposed adaptive image denoiser. While the original N-AIDE was designed for additive case, we show that same framework, i.e., adaptively learning pixel-wise affine denoisers minimizing an unbiased estimate of MSE, can be applied to case as well. Moreover, derive double-sided masked CNN architecture control variance activation values in...
열전 변환 에너지 하베스팅을 위한 저 전력 부스트 컨버터에 사용하는 새로운 Zero Current Sensor (ZCS)를 이 논문에서 제안한다.새로 제안하는 ZCS를 Switching은 기존 방식인 아날로그 비교기를 사용한 Switching방식 보다 파워 측면에서 큰 장점을 보이고 기존의 다른 딜레이 라인을 이용하는 Switching 방식보다 면적에서 보인다. ZCS는 비교기에 고의적으로 offset을 발생시키고 offset의 양을 digital code로 calibration 하여 출력이 나오는 시간을 조절한다. 이용한 대략 10배정도 적은 파워를 사용하면서 같은 성능을 This paper presents a low power boost converter using offset controlled (ZCS) control for thermoelectric energy harvesting.[1] [5] Offset ZCS uses adjustable pre-offset...
Non-intrusive load monitoring (NILM), also known as energy disaggregation, is a blind source separation problem where household's aggregate electricity consumption broken down into usages of individual appliances. In this way, the cost and trouble installing many measurement devices over numerous household appliances can be avoided, only one device needs to installed. The has been well-known since Hart's seminal paper in 1992, recently significant performance improvements have achieved by...
Neural network interpretation methods, particularly feature attribution are known to be fragile with respect adversarial input perturbations. To address this, several methods for enhancing the local smoothness of gradient while training have been proposed attaining \textit{robust} attributions. However, lack considering normalization attributions, which is essential in their visualizations, has an obstacle understanding and improving robustness methods. In this paper, we provide new insights...