Publications by authors named "Joon-Hyuk Chang"
Article Synopsis
- The article discusses the significance of developing broadband photosensors, specifically focusing on indium-gallium-zinc oxide (IGZO) phototransistors integrated with lead sulfide (PbS) quantum dots (QDs).
- It highlights two strategies to enhance the reliability of these phototransistors: p-type doping of the PbS layer through oxidation, and controlling the agglomeration of PbS QDs via thermal annealing.
- The improved GaO/PbS QD/IGZO phototransistors not only demonstrated exceptional photoresponsivity and detectivity but also showed better stability, maintaining performance for over five weeks compared to just two weeks for traditional systems.
A Multi-modal Teacher-student Framework for Improved Blood Pressure Estimation.
Annu Int Conf IEEE Eng Med Biol Soc
July 2023
Article Synopsis
- This study introduces a new method for estimating blood pressure (BP) by transferring knowledge from a complex multi-modal model to a simpler uni-modal model using a teacher-student training approach.
- The multi-modal model uses a combination of inputs, including photoplethysmogram (PPG), electrocardiogram (ECG), and patient demographics, to create features and estimate BP.
- The proposed uni-modal model focuses only on the PPG signal but learns to predict BP effectively by aligning its embeddings with those from the multi-modal model, achieving accurate prediction errors within accepted standards for systolic and diastolic BP.
Article Synopsis
- Several studies have explored cuffless blood pressure measurement using finger photoplethysmogram (PPG) signals, leading to the development of a new BP estimation system that uses PPG signals under varying finger pressure.
- The new system includes a multi-channel PPG and force measurement sensor that enhances accuracy by reducing errors related to finger positioning during the measurement process.
- Utilizing a deep-learning algorithm with an attention mechanism, the system effectively selects the best PPG channel for accurate blood pressure estimation, achieving acceptable error margins for systolic and diastolic blood pressure.
High-Performance Broadband Phototransistor Based on TeO/IGTO Heterojunctions.
ACS Appl Mater Interfaces
January 2022
Article Synopsis
- Researchers developed a new heterostructure using 9 nm-thick tellurium oxide (TeO) and 8 nm-thick InGaSnO (IGTO) to enhance photodetection capabilities across a broad range of wavelengths, from ultraviolet to infrared.
- The resulting photosensor shows impressive performance metrics: detectivity of 1.6 × 10 Jones, responsivity of 84 A/W, and external quantum efficiency of 222% when exposed to blue light (450 nm), demonstrating its potential for high-performance applications.
- The device also exhibits decent infrared detection capabilities and can be produced at low temperatures, making it suitable for creating flexible and stable optoelectronic devices with broad detection capability.
Article Synopsis
- The study focuses on improving sound event detection (SED) performance in noisy environments using a combination of specialized neural networks for noise suppression and classification.
- The proposed approach utilizes a context codec method-equipped temporal convolutional network for noise suppression and a convolutional recurrent neural network for SED, achieving reduced model complexity without sacrificing performance.
- Experimental results demonstrate that this method effectively enhances classification performance in various noisy conditions.
We propose a deep-learning algorithm that directly compensates for luminance degradation because of the deterioration of organic light-emitting diode (OLED) devices to address the burn-in phenomenon of OLED displays. Conventional compensation circuits are encumbered by high cost of the development and manufacturing processes because of their complexity. However, given that deep-learning algorithms are typically mounted onto systems on chip (SoC), the complexity of the circuit design is reduced, and the circuit can be reused by only relearning the changed characteristics of the new pixel device.
View Article and Find Full Text PDFArticle Synopsis
- The paper introduces a new method called Multi-TALK that uses a multi-channel cross-tower network with attention mechanisms to reduce both echo and noise in audio recordings.
- It employs a parallel encoder to gather information from multiple microphones and incorporates spatial data through a 1D convolution process.
- The method iteratively enhances the quality of the near-end speech by using attention mechanisms to selectively eliminate unwanted noise without distorting the actual speech, achieving better results than traditional algorithms.
Article Synopsis
- - This paper presents a method that combines deep neural network (DNN) techniques for dereverberation and beamforming to improve sound event detection (SED) in environments with multiple audio channels.
- - The process involves calculating audio signal features, enhancing them using DNN-supported methods, and training three interconnected modules with a single loss function to optimize performance.
- - To tackle data imbalance in training, the authors introduce focal loss, which significantly boosts SED performance in challenging noisy and reverberant conditions, as demonstrated by experimental results.
Article Synopsis
- Researchers developed an electrical sensor that mimics the human sense of touch by using a porous graphene film to encode complex surface textures.
- The sensor analyzes electrical signals created when touching objects, allowing it to recognize tactile patterns.
- Machine learning techniques improve the sensor’s accuracy in classifying textures, outperforming human touch in recognizing delicate fabric samples.
Deep learning ensemble with asymptotic techniques for oscillometric blood pressure estimation.
Comput Methods Programs Biomed
November 2017
Article Synopsis
- This paper introduces a deep learning approach using an ensemble regression estimator to improve the accuracy of oscillometric blood pressure (BP) measurements by combining techniques like bootstrap and Monte-Carlo methods to reduce uncertainty in estimated values.
- It addresses challenges such as the difficulty in selecting the best deep belief network (DBN) and deep neural network (DNN) estimators, as well as the risk of overfitting due to a small sample size from only five measurements per subject.
- By employing a two-stage ensemble method that combines bootstrap-aggregation and AdaBoost, the study successfully reduces estimation errors, demonstrating that this technique yields more reliable blood pressure estimates compared to traditional single regression estimators.
Article Synopsis
- Polysomnography (PSG) is the standard method for sleep stage classification but is intrusive, leading to the development of noninvasive sleep stage algorithms that haven't been proven reliable yet.
- This study introduces a new approach using low-cost, noncontact multi-modal sensors that analyze radar signals and sound to classify sleep stages, specifically designed for sleep disorder patients.
- The proposed algorithm, which integrates medical insights and customized thresholds, shows promising results in comparison to single sensor methods and is validated against a commercial device, indicating potential for commercialization in sleep monitoring.
Article Synopsis
- * This hybrid approach outperforms traditional methods by producing lower mean error and narrower confidence intervals, achieving an "A" grade under specific testing standards.
- * The findings suggest that this innovative methodology allows for individualized characteristic ratios, leading to more precise blood pressure estimates.
Article Synopsis
- Current oscillometric blood pressure devices mainly provide single-point estimates for systolic (SBP) and diastolic blood pressures (DBP) without confidence ranges, while a new method aims to include confidence intervals (CIs) for these measurements.
- This method employs multiple regression to optimally combine SBP and DBP estimates from various algorithms, using a weighted bootstrap approach to generate a pseudo sample set for better accuracy despite a limited number of estimates.
- Testing on 85 patients showed that this new technique not only offers improved BP estimates but also produces smaller CIs compared to traditional methods, enhancing the reliability of single BP readings.
Article Synopsis
- - The paper introduces a new method for voice activity detection (VAD) that utilizes a kernel subspace approach to enhance the effectiveness of existing kernel-based VAD systems.
- - It employs a linear transform matrix derived from kernel principal component analysis to manage two covariance matrices simultaneously while creating the kernel subspace.
- - Experimental results indicate that this new VAD algorithm significantly improves performance compared to traditional methods, especially in noisy environments.
Article Synopsis
- - This paper introduces a new method for detecting voice activity using advanced statistical techniques in a complex feature space created through nonlinear mapping.
- - It utilizes a Gaussian density model along with kernel principal component analysis to capture the unique characteristics of speech signals.
- - The method features a decision-making process that relies on a multiple observation likelihood ratio test applied in the kernel space to improve detection accuracy.
A subspace approach based on embedded prewhitening for voice activity detection.
J Acoust Soc Am
November 2011
Article Synopsis
- - This paper introduces a new method for voice activity detection (VAD) that uses a subspace approach and involves a prewhitening scheme to analyze clean speech and noise.
- - The method applies a likelihood ratio test in the signal subspace domain to improve accuracy in distinguishing between speech and noise.
- - Experimental results indicate that this subspace-based VAD method is more effective than the traditional Gaussian model approach, especially in situations with low signal-to-noise ratios.