Khanh-Binh Nguyen CV

About Me

Currently working as a Ph.D. candidate of Sungkyunkwan University (South Korea) and a member of DATES Lab at Yonsei University (South Korea), a research group led by Professor Joon-Sung Yang specializing in the research and development of architecture, design issues for intelligent systems, memory architectures, dependable/fault-tolerant computing, H/W security, and computer-aided design for VLSI/SoC systems and emerging technologies.
My Ph.D. in Electrical and Computer Engineering from Sungkyunkwan University (South Korea) is expected to finish in August 2023. My study focuses on efficient computer vision training techniques; architectures for image recognition, object detection, and segmentation; self-, semi-, meta-, unsupervised learning; and representation learning.
Aside from research, I am a recurrent contestant in ML/DL contests organized by Kaggle, Zindi, and Dacon.

+82-10-5947-8134

Seoul, South Korea

DATES Lab (Yonsei University)

Specialized in Deep Neural Networks via supervised, self, and semi-supervised learning. Fully capable of the most up-to-date architectures in Image Recognition such as LeNet, VGG, Inception, Deep Residual Networks, and Vision Transformers, ...
Optimized model efficiency via regularization techniques such as dropout, data augmentations, and batch normalization,...
Researched for efficient semi-supervised learning pipeline based on popular algorithms such as FixMatch, FlexMatch, ... Diagnosed the drawbacks of conventional SSL methods such as confirmation bias, long-tailed, and data utilization; which resulted in new methods to mitigate those issues.
Assessed data model memory usage and proposed improvements by utilizing memory-friendly DNN architectures, quantization, and pruning methods.
Fully capable to deploy state-of-the-art frameworks in AI technology such as YOLO, Detectron, timm, scikit-learn, XGBoost, CatBoost, and LightGBM, ...

KEY OUTCOME

Developed a new dropout algorithm called Checkerboard Dropout. Checkerboard Dropout targets improving the conventional structured dropout (DropBlock), which inefficiently drops a large region of features in the feature map and the features of the last layer. Checkerboard Dropout helps improve the generalization of the networks, thus improving their robustness and generalization. The strengths of Checkerboard Dropout are proven through extensive experiments on a wide range of appliances such as image recognition, object detection, and segmentation.
Developed a variant of a convolutional layer, EUNConv, and a ResNet-like network using this layer, EUNNet. EUNNet allows training deep residual networks without batch normalization efficiently and stably. Extensive experiments show that EUNNet not only mitigates the gradient vanishing/exploding issue but also requires less parameters overhead and tuning; resulting in faster and more efficient models training. EUNNet surpasses previous methods on a wide range of tasks such as image recognition, object detection.
Developed a more efficient pipeline for semi-supervised learning which mitigates confirmation bias, inefficient data utilization, and long-tailed issues, ...
Supervised and instructed undergraduate students on Deep Learning fundamentals and Pytorch fundamentals for AI research.
Served as reviewer for DAC, ICCAD, DATE from 2019 to 2023.

Exgo Technologies - GreenSys

Developed the core system for a Smart Logistic Application - a platform that maximizes the truck drivers productivity & truck owners profit (like LOGIVAN).
Participated in the planning, design, and implementation of the Back-end for the system from scratch using Golang to support concurrency and consistency between two applications: Shipper and Carrier. Carrier is the application to monitor trucks for the truck owners, and Shipper is the application for the truck drivers.
Planned, designed, and implemented Front-end Web service for internal ERP to monitor the transactions between shippers and carriers. The module also supports extracting multiple indicators such as revenue, number of orders, and number of canceled and finished orders, ...
Designed, developed and monitored the business model between shippers, carriers, and solution providers.

NTT Data Vietnam

Optimized the ETL pipeline for Auchan Retail, a retailer with multiple databases and more than 1M daily transactions.
Deployed memory-friendly SQL scripts to extract data from the database to BI tools daily automatically.
Built business performance dashboards using Qlik View and Qlik Sense; which received outstanding ratings and feedback from the client’s Board of Directors.
Optimized the business operations regarding Business intelligence and Business data, which reduced workloads for the IT team by 70%.

Ph.D. of Engineering

Sungkyunkwan University, South Korea

2019-2023

Bachelor of Computer Science

Ho Chi Minh City University of Technology

2014-2018

English

■■■■■■□□□□ Advance

Vietnamese ****

■■■■■■■■■ Native

Nguyen, Khanh-Binh, Jaehyuk Choi, and Joon-Sung Yang. ”Checkerboard Dropout: A Structured Dropout With Checkerboard Pattern for Convolutional Neural Networks.” IEEE Access 10 (2022): 76044-76054.
Nguyen, Khanh-Binh, Jaehyuk Choi, and Joon-Sung Yang. ”EUNNet: Efficient UN-normalized Convolution layer for stable training of Deep Residual Networks without Batch Normalization layer.” IEEE Access 10 (2023).