Seung Eun Lee
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Seung Eun Lee has a total of 152 co-authors. Here are the top co-authors by number of shared publications:
- Nader Bagherzadeh
University of California, Irvine
14 shared publications Sang Don Kim
Dept. of Electronic Engineering, Seoul National University of Science and Technology, Seoul, Korea
12 shared publications- Joungmin Park
Department of Electronic Engineering, Seoul National University of Science and Technology, Seoul 01811, Republic of Korea
11 shared publications Jung Hwan Oh
Department of Electronic Engineering, Seoul National University of Science and Technology, Seoul, South Korea
11 shared publicationsSang Muk Lee
Dept. of Electronic Engineering, Seoul National University of Science and Technology
10 shared publications- Hyun Woo Oh
University of California, Irvine, Irvine, California, USA
9 shared publications
Latest Publications
- 1 October 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Embedded Systems Letters
- Vol. 18  (5) , 349-352
Why Are Differentiable Weightless Neural Networks Robust to Sensor Noise? Three Mechanisms for Edge HAR
- 1 October 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Embedded Systems Letters
- Vol. 18  (5) , 349-352
Why Are Differentiable Weightless Neural Networks Robust to Sensor Noise? Three Mechanisms for Edge HAR
- Open Access
- 22 August 2026
- Published by MDPI AG in Electronics
- Vol. 15  (17) , 3762
A Memory-Efficient Depthwise Separable Convolution Accelerator Using Run-Length Coding
On edge devices, convolutional neural network (CNN) inference is bottlenecked mainly by memory bandwidth, owing to the frequent memory accesses to feature maps and parameters. To address this challenge, we propose a memory-efficient hardware accelerator for depthwise separable convolution that minimizes off-chip memory traffic and parameter storage. The proposed architecture employs three key techniques: (1) a 64-bit run-length coding (RLC) packet compression that exploits feature-map sparsity after ReLU, (2) a mixed-precision scheme that represents feature maps and weights at different precisions, and (3) separated depthwise and pointwise convolution units. In particular, feature maps are transferred in RLC-compressed form, which reduces the amount of data exchanged with the host. The compressed data are decoded row by row, so the on-chip buffers hold only the rows required for computation rather than a complete feature map. In software simulation on ImageNet, the mixed-precision scheme reduced the parameter storage by 49.22% at a cost of 6.24 percentage points (pp) in Top-1 accuracy, and the RLC reduced the data by up to 56.07% in the deeper layers. Implemented on a Xilinx ZCU-104 FPGA, the proposed accelerator performs the depthwise separable convolution with a small number of logic resources and on-chip memory, confirming its feasibility for resource-constrained edge devices.
- 22 August 2026
- Published by MDPI AG in Electronics
- Vol. 15  (17) , 3762
Open AccessA Memory-Efficient Depthwise Separable Convolution Accelerator Using Run-Length Coding
On edge devices, convolutional neural network (CNN) inference is bottlenecked mainly by memory bandwidth, owing to the frequent memory accesses to feature maps and parameters. To address this challenge, we propose a memory-efficient hardware accelerator for depthwise separable convolution that minimizes off-chip memory traffic and parameter storage. The proposed architecture employs three key techniques: (1) a 64-bit run-length coding (RLC) packet compression that exploits feature-map sparsity after ReLU, (2) a mixed-precision scheme that represents feature maps and weights at different precisions, and (3) separated depthwise and pointwise convolution units. In particular, feature maps are transferred in RLC-compressed form, which reduces the amount of data exchanged with the host. The compressed data are decoded row by row, so the on-chip buffers hold only the rows required for computation rather than a complete feature map. In software simulation on ImageNet, the mixed-precision scheme reduced the parameter storage by 49.22% at a cost of 6.24 percentage points (pp) in Top-1 accuracy, and the RLC reduced the data by up to 56.07% in the deeper layers. Implemented on a Xilinx ZCU-104 FPGA, the proposed accelerator performs the depthwise separable convolution with a small number of logic resources and on-chip memory, confirming its feasibility for resource-constrained edge devices.
- Open Access
- 30 September 2025
- Published by MDPI AG in Electronics
- Vol. 14  (19) , 3911
Real-Time True Wireless Stereo Wearing Detection Using a PPG Sensor with Edge AI
True wireless stereo (TWS) earbuds are evolving into multifunctional wearable devices, offering opportunities not only for audio streaming but also for health-related applications. A fundamental requirement for such devices is the ability to accurately detect whether they are being worn, yet conventional proximity sensors remain limited in both reliability and functionality. This work explores the use of photoplethysmography (PPG) sensors, which are widely applied in heart rate and blood oxygen monitoring, as an alternative solution for wearing detection. A PPG sensor was embedded into a TWS prototype to capture blood flow changes, and the wearing status was classified in real time using a lightweight k-nearest neighbor (k-NN) algorithm on an edge AI processor. Experimental evaluation showed that incorporating a validity check enhanced classification performance, achieving F1 scores above 0.95 across all wearing conditions. These results indicate that PPG-based sensing can serve as a robust alternative to proximity sensors and expand the capabilities of TWS devices.
- 30 September 2025
- Published by MDPI AG in Electronics
- Vol. 14  (19) , 3911
Open AccessReal-Time True Wireless Stereo Wearing Detection Using a PPG Sensor with Edge AI
True wireless stereo (TWS) earbuds are evolving into multifunctional wearable devices, offering opportunities not only for audio streaming but also for health-related applications. A fundamental requirement for such devices is the ability to accurately detect whether they are being worn, yet conventional proximity sensors remain limited in both reliability and functionality. This work explores the use of photoplethysmography (PPG) sensors, which are widely applied in heart rate and blood oxygen monitoring, as an alternative solution for wearing detection. A PPG sensor was embedded into a TWS prototype to capture blood flow changes, and the wearing status was classified in real time using a lightweight k-nearest neighbor (k-NN) algorithm on an edge AI processor. Experimental evaluation showed that incorporating a validity check enhanced classification performance, achieving F1 scores above 0.95 across all wearing conditions. These results indicate that PPG-based sensing can serve as a robust alternative to proximity sensors and expand the capabilities of TWS devices.
- 8 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems I: Regular Papers
- Vol. 72  (12) , 7516-7528
An Accelerated Block Searching Approach in A* for Autonomous Mobile Robots
Path planning is crucial to ensure the safe navigation of autonomous mobile robots (AMRs). However, as the scale of maps and paths increases, achieving faster and more efficient computations becomes challenging due to constraints in memory and computational resources. In this paper, we present a hardware architecture for global path planning utilizing block searching A* (BSA*) for AMRs. The BSA* designates surrounding nodes as blocks and searching nodes for expansion with collision detection based on block-based jump point search (JPS(B)). This reduces the amount of stored data, enables faster map scanning, and provides advantages in parallelized architecture. The BSA* accelerator, consisting of a memory controller based on heap sorting and a parallelized collision detector, was implemented on a field-programmable gate array (FPGA). In the benchmarks for grid-based pathfinding, experimental results showed that BSA* stored 83.2% less data compared to A* and BSA* accelerator demonstrated real-time performance, ranging from 0.528 ms to 2.983 ms, successfully verifying the feasibility of real-time path planning with a block searching approach.
- 8 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems I: Regular Papers
- Vol. 72  (12) , 7516-7528
An Accelerated Block Searching Approach in A* for Autonomous Mobile Robots
Path planning is crucial to ensure the safe navigation of autonomous mobile robots (AMRs). However, as the scale of maps and paths increases, achieving faster and more efficient computations becomes challenging due to constraints in memory and computational resources. In this paper, we present a hardware architecture for global path planning utilizing block searching A* (BSA*) for AMRs. The BSA* designates surrounding nodes as blocks and searching nodes for expansion with collision detection based on block-based jump point search (JPS(B)). This reduces the amount of stored data, enables faster map scanning, and provides advantages in parallelized architecture. The BSA* accelerator, consisting of a memory controller based on heap sorting and a parallelized collision detector, was implemented on a field-programmable gate array (FPGA). In the benchmarks for grid-based pathfinding, experimental results showed that BSA* stored 83.2% less data compared to A* and BSA* accelerator demonstrated real-time performance, ranging from 0.528 ms to 2.983 ms, successfully verifying the feasibility of real-time path planning with a block searching approach.
- Open Access
- 11 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 102859-102868
A CRC Comparison-Based Screen Data Management for Energy Efficient Virtual Desktop Infrastructure
Cloud computing exploits virtualized resources from a server with minimum user specifications. These virtualized resources are called a virtual machine (VM), and the VM enables a user to perform high performance computing. As the VM enables users to utilize high performance computing, the demand for virtual desktop infrastructure (VDI) which applies the VM as a desktop is increased. To use the VDI, the transmission of screen data from the VM to the user is essential. Recently, the workload of screen data transmission has increased due to the high screen resolution. In order to reduce the workload, several methods were proposed that transmit only the changed parts of the screen. However, comparing the screen data requires memory access to detect the changes in the screen. In this paper, we propose an efficient screen data management system for the VDI with the accelerated cyclic redundancy check (CRC) generation. By comparing the CRC of the screen blocks, the system reduces the number of memory access and power consumption. The accelerated CRC generation also reduces the workloads in high resolution screen data. We simulated 90,000 frames of 1920 by 1080 resolution in different types of programs. As a result, the proposed system reduced the data traffic up to 98% with an error rate of under 0.35% and reduced the number of memory access.
- 11 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 102859-102868
Open AccessA CRC Comparison-Based Screen Data Management for Energy Efficient Virtual Desktop Infrastructure
Cloud computing exploits virtualized resources from a server with minimum user specifications. These virtualized resources are called a virtual machine (VM), and the VM enables a user to perform high performance computing. As the VM enables users to utilize high performance computing, the demand for virtual desktop infrastructure (VDI) which applies the VM as a desktop is increased. To use the VDI, the transmission of screen data from the VM to the user is essential. Recently, the workload of screen data transmission has increased due to the high screen resolution. In order to reduce the workload, several methods were proposed that transmit only the changed parts of the screen. However, comparing the screen data requires memory access to detect the changes in the screen. In this paper, we propose an efficient screen data management system for the VDI with the accelerated cyclic redundancy check (CRC) generation. By comparing the CRC of the screen blocks, the system reduces the number of memory access and power consumption. The accelerated CRC generation also reduces the workloads in high resolution screen data. We simulated 90,000 frames of 1920 by 1080 resolution in different types of programs. As a result, the proposed system reduced the data traffic up to 98% with an error rate of under 0.35% and reduced the number of memory access.
- Open Access
- 7 June 2025
- Published by MDPI AG in Electronics
- Vol. 14  (12) , 2337
Hardware Accelerator for Approximation-Based Softmax and Layer Normalization in Transformers
Transformer-based models have achieved remarkable success across various AI tasks, but their growing complexity has led to significant computational and memory demands. While most optimization efforts have focused on linear operations such as matrix multiplications, non-linear functions like Softmax and layer normalization (LayerNorm) are increasingly dominating inference latency, especially for long sequences and high-dimensional inputs. To address this emerging bottleneck, we present a hardware accelerator that jointly approximates these non-linear functions using piecewise linear approximation for the exponential in Softmax and Newton–Raphson iteration for the square root in LayerNorm. The proposed unified architecture dynamically switches operation modes while reusing hardware resources. The proposed accelerator was implemented on a Xilinx VU37P FPGA and evaluated with BERT and GPT-2 models. Experimental results demonstrate speedups of up to 7.6× for Softmax and 2.0× for LayerNorm, while maintaining less than 1% accuracy degradation on classification tasks with conservative approximation settings. However, generation tasks showed greater sensitivity to approximation, underscoring the need for task-specific tuning.
- 7 June 2025
- Published by MDPI AG in Electronics
- Vol. 14  (12) , 2337
Open AccessHardware Accelerator for Approximation-Based Softmax and Layer Normalization in Transformers
Transformer-based models have achieved remarkable success across various AI tasks, but their growing complexity has led to significant computational and memory demands. While most optimization efforts have focused on linear operations such as matrix multiplications, non-linear functions like Softmax and layer normalization (LayerNorm) are increasingly dominating inference latency, especially for long sequences and high-dimensional inputs. To address this emerging bottleneck, we present a hardware accelerator that jointly approximates these non-linear functions using piecewise linear approximation for the exponential in Softmax and Newton–Raphson iteration for the square root in LayerNorm. The proposed unified architecture dynamically switches operation modes while reusing hardware resources. The proposed accelerator was implemented on a Xilinx VU37P FPGA and evaluated with BERT and GPT-2 models. Experimental results demonstrate speedups of up to 7.6× for Softmax and 2.0× for LayerNorm, while maintaining less than 1% accuracy degradation on classification tasks with conservative approximation settings. However, generation tasks showed greater sensitivity to approximation, underscoring the need for task-specific tuning.
- 19 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-3
FPGA Based Approximate Vector Operation Accelerator for VLMs
As artificial intelligence continues to advance, the importance of autonomous robotic systems and human-robot in-teractions is growing, particularly in the areas of language-based control and behavior prediction. Vision-language models (VLMs), which simultaneously process images and texts, are essential for interpreting commands and analyzing environments. However, their high computational demands, particularly for vector op-erations like softmax and layer normalization, pose significant challenges. These operations, involving complex functions such as exponentials and square roots, consume substantial resources as model sizes grow. This paper proposes a vector operation accelerator for VLMs through approximation techniques, specifically Newton-Raphson and piecewise linear approximations, to improve speed while reducing resource usage. The optimized architecture reuses resources by targeting redundant operations. Implemented on an FPGA, the accelerator achieved up to 54% faster performance compared to an RTX 3070 GPU, with minimal aproximation error.
- 19 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-3
FPGA Based Approximate Vector Operation Accelerator for VLMs
As artificial intelligence continues to advance, the importance of autonomous robotic systems and human-robot in-teractions is growing, particularly in the areas of language-based control and behavior prediction. Vision-language models (VLMs), which simultaneously process images and texts, are essential for interpreting commands and analyzing environments. However, their high computational demands, particularly for vector op-erations like softmax and layer normalization, pose significant challenges. These operations, involving complex functions such as exponentials and square roots, consume substantial resources as model sizes grow. This paper proposes a vector operation accelerator for VLMs through approximation techniques, specifically Newton-Raphson and piecewise linear approximations, to improve speed while reducing resource usage. The optimized architecture reuses resources by targeting redundant operations. Implemented on an FPGA, the accelerator achieved up to 54% faster performance compared to an RTX 3070 GPU, with minimal aproximation error.
- 11 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-3
Flexible NTT/iNTT Accelerator Based on Montgomery Reduction
As the importance of data grows, so does the significance of homomorphic encryption(HE), which enables direct computations on encrypted data. However, HE presents a major challenge, with computational overhead increasing compared to plaintext operations. To address this, we developed a hardware accelerator for NTT and iNTT, which are commonly used in HE and contribute to this overhead. The accelerator implements modular arithmetic using Montgomery reduction, and pipelining is applied to improve throughput. Additionally, a mode-switching feature is introduced, allowing both NTT and iNTT operations to be performed on a single circuit. The accelerator is also designed to operate flexibly across various cryptographic strength parameters ( $\boldsymbol{N}$ -values). The proposed accelerator achieved an average speedup of approximately 4.24 times compared to software, demonstrating its effectiveness in reducing computation time in HE systems.- 11 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-3
Flexible NTT/iNTT Accelerator Based on Montgomery Reduction
As the importance of data grows, so does the significance of homomorphic encryption(HE), which enables direct computations on encrypted data. However, HE presents a major challenge, with computational overhead increasing compared to plaintext operations. To address this, we developed a hardware accelerator for NTT and iNTT, which are commonly used in HE and contribute to this overhead. The accelerator implements modular arithmetic using Montgomery reduction, and pipelining is applied to improve throughput. Additionally, a mode-switching feature is introduced, allowing both NTT and iNTT operations to be performed on a single circuit. The accelerator is also designed to operate flexibly across various cryptographic strength parameters ( $\boldsymbol{N}$ -values). The proposed accelerator achieved an average speedup of approximately 4.24 times compared to software, demonstrating its effectiveness in reducing computation time in HE systems.- 30 December 2024
- Published by Elsevier in Integration
- Vol. 101
SEAM: A synergetic energy-efficient approximate multiplier for application demanding substantial computational resources
- 30 December 2024
- Published by Elsevier in Integration
- Vol. 101
SEAM: A synergetic energy-efficient approximate multiplier for application demanding substantial computational resources
- Open Access
- 5 December 2024
- Published by MDPI AG in Electronics
- Vol. 13  (23) , 4795
PIMCoSim: Hardware/Software Co-Simulator for Exploring Processing-in-Memory Architectures
As the scope of artificial intelligence (AI) expands and the structure becomes more complex, the amount of data for inference and training has increased. In traditional computer architectures, the memory bandwidth limitations have intensified bottlenecks in AI systems, and processing-in-memory (PIM) architectures have been proposed to overcome this issue. PIM is an architecture that performs computations within memory, thereby reducing data movement between the CPU and memory. However, since PIM is difficult to optimize as a general-purpose architecture, it is essential to adopt an architecture suitable for the target application. While various simulators and emulators have been introduced for the design space exploration (DSE) of different PIM architectures, simulators are limited in debugging hardware operations, and emulators face challenges in flexibly modifying the system configuration, as emulators implement the entire architecture in hardware. Therefore, this paper introduces PIMCoSim, a comprehensive hardware–software co-simulator for the DSE of DRAM-PIM systems. This co-simulator partially emulates simplified hardware-implemented processing elements (PEs) and integrates software models for memory operations, facilitating the DSE of PIM systems. To validate PIMCoSim, we analyzed results for different computational workloads by varying PIM structures and operational policies, demonstrating the efficiency of DRAM-PIM systems. The co-simulation approach in PIMCoSim aims to contribute to analyzing DRAM-PIM configurations and adopting optimized structures.
- 5 December 2024
- Published by MDPI AG in Electronics
- Vol. 13  (23) , 4795
Open AccessPIMCoSim: Hardware/Software Co-Simulator for Exploring Processing-in-Memory Architectures
As the scope of artificial intelligence (AI) expands and the structure becomes more complex, the amount of data for inference and training has increased. In traditional computer architectures, the memory bandwidth limitations have intensified bottlenecks in AI systems, and processing-in-memory (PIM) architectures have been proposed to overcome this issue. PIM is an architecture that performs computations within memory, thereby reducing data movement between the CPU and memory. However, since PIM is difficult to optimize as a general-purpose architecture, it is essential to adopt an architecture suitable for the target application. While various simulators and emulators have been introduced for the design space exploration (DSE) of different PIM architectures, simulators are limited in debugging hardware operations, and emulators face challenges in flexibly modifying the system configuration, as emulators implement the entire architecture in hardware. Therefore, this paper introduces PIMCoSim, a comprehensive hardware–software co-simulator for the DSE of DRAM-PIM systems. This co-simulator partially emulates simplified hardware-implemented processing elements (PEs) and integrates software models for memory operations, facilitating the DSE of PIM systems. To validate PIMCoSim, we analyzed results for different computational workloads by varying PIM structures and operational policies, demonstrating the efficiency of DRAM-PIM systems. The co-simulation approach in PIMCoSim aims to contribute to analyzing DRAM-PIM configurations and adopting optimized structures.