Myeonghwan AHN

Ph.D candidate at CMALab, SNUCSE

About Me

I’m CSE student/Ph.D candidate at CMALab, Seoul National University.

My general topic of interest is hardware-software co-design.

Specifically, from application down to hardware:

Publication

NIPQ: Noise proxy-based Integrated Pseudo-Quantization

10.1109/CVPR52729.2023.00375

I was accidentally not included in the author list. You can find me in the Acknowledgments section.

CVPR 2023

Juncheol Shin, Junhyuk So, Sein Park, Seungyeop Kang, Sungjoo Yoo, and Eunhyeok Park

Differentiable mixed precision QAT, part of IITP project.

BlindFilter: Privacy-Preserving Spam Email Detection Using Homomorphic Encryption

10.1109/SRDS60354.2023.00014

SRDS 2023

Dongwon Lee, Myeonghwan Ahn, Hyesun Kwak, Jin B. Hong and Hyoungshick Kim

Pseudo quantization noise as a defensive measure against adversarial attack on parameters of homomorphic encryption algorithm.

Projects

Lab cluster networking and management

  • Designed and implemented networking for the cluster in our lab.
    • 9(=5+2+2) racks across 3 sites
    • management fabric
      • OSPF over WireGuard site-to-site
      • 10G ~ 1G leaf-spine for each site
    • data fabric for 5-rack site
      • OSPF + ECMP leaf-spine with L3 HW offloading
      • 100G ~ 10G with jumbo frame
  • Implemented all-flash Ceph storage cluster in data fabric.
  • Implemented management system for the whole cluster
    • monitoring via node-exporter, Prometheus and Grafana
    • automatic alert using Grafana and Discord webhook
      • anomaly temperature
      • unhealthy Ceph state
      • idle on-demand GPUs
    • on-demand resources as-user-want via canonical MAAS

Samsung MX project

This was supported by Mobile eXperience(MX) Business, Samsung Electronics Co., Ltd. // 본 연구는 삼성전자 MX사업부의 지원을 받아 수행한 결과임

I’m contributing to the project starting from 2H 2024.

IITP governmental project

Development of model compression framework for scalable on-device AI computing on Edge applications. MSIT, No.2019-0-01906, No.2021-0-00105, and No.2021-0-00310

I’m contributing to the project starting from 2H 2021. The project ended in 2024.

TinyML image classification

lightb0x/arduino_trash_classification

From SNUCSE 2020 “creative integrated design” class.

Trash classification on Arduino Nano 33 BLE and Raspberry Pi 2.

From training various models(MbV1, MbV2, ShV2) on resized imagenet(96x96, 128x128, …), fine-tuning on trashnet down to actually running on hardwares.

Coursework TA

Hardware System Design

2022, 2023

This course covers hardware-software co-design in terms of building custom NN accelerator.

Till 2022, the course covered: Xilinx FPGA (Zedboard), verilog on Vivado, C++ and MNIST classification (caffe2).

In 2023, I revised the course to cover: Amaranth simulator, simple NN compiler, and torch. Specifically:

  • GEMM accelerator
    • simple instruction set
    • reconfigurable
    • outer product matrix-matrix multiplication
    • on Amaranth
  • simple NN compiler
    • generate simple instruction set code given input pytorch GEMM

Embedded Systems and Applications

2022

lightb0x/yolov5_nipq_practice

This course covers latest topics in ML applications overall.

In 2022, I covered YOLOv5 and applying NIPQ on it.

Education

Graduate

2021.9 ~

Ph.D candidate at CMALab, Seoul National University

Undergraduate

2015.3 ~ 2017.7 // 2019.7 ~ 2021.8

Dept. of Computer Science and Engineering, Seoul National University

Experience

CMALab, Seoul National University

Internship

2020.12 ~ 2021.9

I wrote my undergraduate thesis during an internship. It focused on applying differentiable neural architecture search (NAS) with Gumbel-softmax to select [Binary|INT8] layerwise for an image classification network targeted at TinyML.

It was extension of lightb0x/arduino_trash_classification

Samsung Electronics

Internship

2020.1.6 ~ 2020.2.14

In cloud management group, objective of the team was to try out new technologies, especially Kubernetes and Terraform at that time.

I tried toy project with React, golang-gin, Docker, Kubernetes, and Terraform.