Bing-Min(Ben) Wang

Currently, I work as a software engineer on the AdInsight team at Microsoft where I handle petabyte-scale data, maintain stable services, and utilize both deterministic and machine learning techniques to recommend keywords. In addition, I have experience in applying machine learning in other areas, such as healthcare. I am an open-minded and fast learner, with the ability to quickly adapt to new processes, systems, and technologies. I excel in time management, multitasking, and thrive under pressure. I am passionate about tackling tough technical challenges and collaborating with team members to solve difficult problems. I take pride in bringing my ideas to life through real-world applications.

E-mail: [email protected]

Skills

Have practical experience in developing and managing distributed systems, as well as handling large-scale data.
Combine deterministic and ML-based techniques to generate keyword recommendations that meet latency requirements while maintaining high quality.
Design state-of-the-art deep neural networks to solve imaging problems (2D/3D), sequential/time series and tabular data

object detection such as Faster R-CNN, SSD and YOLO, image segmentation such as U-Net
attention model such as Transformer

Demonstrate a solid understanding of generative models, such as GPT and diffusion models, showcasing proficiency in leveraging their capabilities for various applications
Test and evaluate algorithms to prove robustness
Have great communication, planning skills and profound experience in cooperating with experts in other fields

Work Experience

Microsoft，Sep 2021 - Present

Software Engineer, AdInsight (STCA)

Conduct Ads Globalization:

Enable ES/IT/NL recommendations in daily services pipeline(new keyword recommendation). About 2~3% increase in the revenue.
Support markets expansion in real-time services. The real-time product(K2K, keyword to keyword) can support extra 64 markets by leveraging table generated by partner team.
Improve the quality(coverage and depth) of the suggestions by introducing the INTL(international) TwinBERT trained from partner team.

Resolve language mismatch issue in daily services pipeline to decrease the dismiss and rate ultimately drive the revenue growth.
Set up a daily pipeline to monitor the quality of ES/IT/NL suggestions to prevent hurting user experience owing to globalization.
Resolve MAD(monitor, alerting, diagnosis) service doesn't send alerting emails in time for specific jobs to prevent team from getting ICM tickets
Leaverage LLM (large language model) to replace human-labeling and reduce ~30% of labeling budget.

HTC Healthcare (DeepQ)，Sep 2018 - Aug 2021 · 3 yrs

Senior Deep Learning Engineer, Deep Learning Apps

Nodule Detection: Developed the 3D detection model (Faster R-CNN, U-Net like backbone) from scratch and use Focal loss and adding hard negative example gradually to deal with the large class imbalance
Intracranial Hemorrhage Classification: Developed the ICH classification model (Efficient-Net) combined with transformer encoder to consider sequential information and deployed to hospital PACS system with heatmap visualization
Ischemic stroke segmentation: Develop the segmentation model to segment core/penumbra zone which can be used in deciding treatment in stroke patient.
Facial Landmark Detection: Developed the real-time face detection model(SSD: Single Shot MultiBox Detector) from scratch and deployed to mobile web browser(onnxjs, tfjs)
Fingertip Detection: Improved performance(~5 AP) of the real-time YOLO-like model running on mobile device(tflite) by a novel data augmentation

Competitions

RSNA Intracranial Hemorrhage Detection (Kaggle), ranked top 3% (silver)

Developed an algorithm to detect acute intracranial hemorrhage and its subtypes
Multi-label Classification

APTOS 2019 Blindness Detection (Kaggle), ranked top 4% (silver)

Developed a classifier that output the severity of diabetic retinopathy given the retina images
Ordinal Classification

KDD CUP 2017 - Task2, ranked top 2.3%

For every 20-minute time window, predict the entry and exit traffic volumes at tollgates

Projects

Stand ML Group - CheXpert

Developed a convolutional neural networks that output the probability of 14 observations given the available frontal and lateral radiographs

Credit Scoring (Sinopac), Jul. 2017 to Jul. 2018

Developed a classifier to predict a company will default or not, and extract the readable rules which are verified by the experts

Education

National Cheng Kung University (NCKU), Sep. 2016 - Jun. 2018

Master of Electrical Engineering

Thesis: Exploring neural network hyper-parameters on small datasets and hand-crafted features: take credit scoring as an example
GPA 4.15

National Cheng Kung University (NCKU), Sep. 2012 - Jun. 2016

Bachelor of Electrical Engineering

Independent Study: Transmitter Front-End Circuit Architecture