Seung-Kyu Choi

A scalable deep-learning accelerator supporting the training process is implemented for device personalization of deep convolutional neural networks (CNNs). It consists of three processor cores operating with distinct energy-efficient dataflow for different types of computation in CNN training. Unlike the previous works where they implement design techniques to exploit the same characteristics from the inference, we analyze major issues that occurred from training in a resource-constrained syste

Training convolutional neural network on device has become essential where it allows applications to consider user's individual environment. Meanwhile, the weight update operation from the training process is the primary factor of high energy consumption due to its substantial memory accesses. We propose a dedicated weight update architecture with two key features: (1) a specialized local buffer for the DRAM access deduction (2) a novel dataflow and its suitable processing element array structur

Personalization by incremental learning has become essential for IoT devices to enhance the performance of the deep learning models trained with global datasets. To avoid massive transmission traffic in the network, exploiting on-device learning is necessary. We propose a software/hardware co-design technique that builds an energy-efficient low-bit trainable system: (1) software optimizations by local low-bit quantization and computation freezing to minimize the on-chip storage requirement and c

Among ultra-high temperature ceramics (UHTCs), zirconium diboride (ZrB2) ceramics have high electrical/thermal conductivity, excellent corrosion resistance, and a chemically stable crystal structure. These properties make ZrB2 suitable for extreme chemical and thermal environments, such as atmospheric re-entry and hypersonic aircrafts. ZrB2 has low intrinsic sinterability due to its strong covalent bonds and low self-diffusion coefficients. In the ZrB2 sintering process, decreased ZrB2 particle

As deep learning applications often encounter accuracy degradation due to the distorted inputs from a variety of environmental conditions, training with personal data has become essential for the edge devices. Hence, ‘training on edge’ by supporting a trainable deep learning accelerator has been actively studied. Nevertheless, previous research does not consider the fundamental datapath for training and the importance of retaining the high performance for inference tasks. In this w

This study investigated the correlation between the surge energy capability and Bi2O3 volatilization in ZnO varistors by analyzing the volatilization phenomenon of Bi2O3. A ZnO varistor is composed mainly of Sb2O3, Bi2O3. Bi2O3 has the lowest vapor pressure among these components. Bi2O3 play an important role, moreover, in forming it grain boundaries. These grain boundaries have low conductivity in the leakage current region, but, becomes higher in the region. It also determines the surge energy

COVID-19 sparked a public health crisis and created a series of public policy challenges. This article examines how COVID-19 interventions played out at the state level given the absence of guidance and coordinated national response. We focus on how the level of policy rigidness and enforcement of behavioral interventions helps us understand the success and failures of reducing the number of positive test rates over a 20-week period (March–July 2020). Specifically, we examine how four specific i

DNN training has become a major workload in on-device situations to execute various vision tasks with high performance. Accordingly, training architectures accompanying approximate computing have been steadily studied for efficient acceleration. However, most of the works examine their scheme on from-the-scratch training where inaccurate computing is not tolerable. Moreover, previous solutions are mostly provided as an extended version of the inference works, e.g., sparsity/pruning, quantization

A scalable deep learning accelerator supporting both inference and training is implemented for device personalization of deep convolutional neural networks. It consists of three processor cores operating with distinct energy-efficient dataflow for different types of computation in CNN training. Two cores conduct forward and backward propagation in convolutional layers and utilize a masking scheme to reduce 88.3% of intermediate data to store for training. The third core executes weight update pr

With the growing demand for processing deep learning applications on edge devices, on-device DNN training has become a major workload to execute a variety of vision tasks suited for users. Therefore, architectures employed with algorithm co-design to accelerate the training process have been steadily studied. However, previous solutions are mostly supported by extended versions of the inference studies, such as sparsity, data flow, quantization, etc. Moreover, most works examine their schemes on

본 논문에서는 pulsed-CW를 사용하는 위협 레이더 신호를 상쇄시키기 위한 능동 상쇄 신호에 임의의 파라미터 추정 오차가 존재하는 경우 능동 상쇄 신호의 영향을 분석하고, 이를 기반으로 능동 상쇄 측면에서 maximum-likelihood estimation (MLE) 기법 및 2차 보간 기법이 포함하는 추정 오차의 영향을 분석한다. 그리고 능동 상쇄 신호의 성능을 향상시키기 위해MLE 기법의 파라미터 추정 오차를 보정하는 correlation 기반의 오차 보정 기법을 개선하여 적용하고, 이 기법이 능동 상쇄관점에서 다른 기법보다 우수함을 모의실험을 통해 검증한다.

Various diagnostic models for combustion instability based on statistical techniques and machine learning using measured or monitored data are being developed to suppress instabilities in power generation gas turbines. In this paper, we proposed a combustion state diagnosis model using an auto-encoder, one of machine learning approaches. This model has been trained using two main input components; magnitude and waveform pattern of time series dynamic pressure which is measured from a real-scale

Various statistical approaches have been recently proposed to detect combustion instabilities in a short time through a fast data processing on dynamic pressure of time domain measured in a gas turbine combustor. In the current study, four methods, PE(permutation entropy), TK(temporal kurtosis), ZCR(zero crossing rate), and STFT(short time Fourier transform), were applied and compared with the conventional instability onset-detection method, RMS(root mean square), using the same dynamic pressure

본 논문에서는 다중 사용자 다중 안테나 하향링크 채널에서 에너지 효율을 향상시키기 위하여 zero-forcing beamforming (ZFBF) 전처리 필터 기반의 기지국 활성 안테나 수 최적화 기법을 제안한다. 제안하는 기법에서는 최적의 안테나 수를 찾는 연산과정의 복잡도를 줄이기 위하여 사용자의 순시 채널 이득 대신 ZFBF의 평균 채널 이득을 사용한다. 그리고 분수함수 형태의 목적함수를 차를 이용한 목적함수로 변환하여 최적의 안테나 수와 최대 에너지 효율을 찾는 과정을 반복 수행하여 문제를 해결한다. 모의실험을 통해 제안하는 기법의 에너지 효율은 exhaustive search 방법으로 찾은 최대 에너지 효율과 거의 동일함을 확인한다.