Chan Kha Vu

A curious human being

About

I’m currently a software engineer at Microsoft in Redmond, United States. Working on customer-centric products of such scales is sooooo exciting!

Previously, I worked as a computer vision engineer at Samsung Research in Ukraine, specializing in Computer Vision and optimizing Neural Networks for resource-constrained Mobile Devices.

On a side for fun, I travel to new places, blog about stuffs I find exciting, paint, read (a lot of) Manga, and play Go (feel free to challenge me). Oh, I also started to compete on Kaggle lately out of boredom.

Timeline

Oct 2019 — Now
Microsoft, Redmond, United States
Software Developer Engineer
Feb 2017 — Oct 2019
Samsung Research, Ukraine
Research Computer Vision/AI Engineer
2017 — 2019
Taras Shevchenko National University
Master's degree in Applied Mathematics.
2013 — 2017
Taras Shevchenko National University
Bachelor's degree in Applied Mathematics

Research (kind of)

Adaptive Multi-Head Network Branching using Feature Similarity (2019)

This work eliminates the limitations of previous Multi-Task Branching Scheme Search methods: it doesn’t require each sample to have labels for all tasks, is not limited only to classification, and demands less resources. The core idea is to apply Network Slimming regularization after each block, then estimate affinity of branches as the sum of Jensen–Shannon Divergence between activations, weighted by pruning importance. This is part of my Master’s thesis. Work in progress.

Breast Cancer Screening by analyzing the Interphase Nuclei of the Buccal Epithelium (BS thesis, 2017)

Under the supervision of Prof. Klyushin Dmitrii, we provided significant improvements to the non-invasive, radiation-free approach on diagnosing Breast Cancer by Boroday et al. that analyses the optical density of the Interphase Nuclei of Buccal Epithelium. The proposed method of locally fusing CNN features with with hand-crafted features (i.e. local fractal dimension) allowed us to achieve good results with limited amount of data.

Open Source, Kaggle, and more

hav4ik/Hydra — a Multi-Task Learning Framework on PyTorch. State-of-the-art methods are implemented to effectively train models on multiple tasks.

My 22nd Place Solution (out of 736 teams) for Google Landmarks Challenge 2020. I used EffNetB6, GemPool, and ArcMargin with dynamic margin scheduling.

hav4ik/eyesight — A python framework for designing high-performance Computer Vision pipelines at the Edge. Supports Coral Edge TPU & Raspberry Pi Camera.

Check out my survey on Deep Multi-Task Learning! I did some in-depth review of the most reliable methods in this area, and how I’m using them personally.