Designed multi-task learning neural network model for click-through rate (CTR) prediction for ads displayed on various third-party platforms, leveraging the similarity among tasks for each platform, and the larger combined training dataset.

Improved CTR prediction AUC by 1%-2%, compared to per-platform single-task production models, across all platforms in LinkedIn Audience Network, LinkedIn's third-party ads delivery system.

Developed data pipelines in Spark that processes billions of samples for production-scale model training and evaluation, model training in Tensorflow, and model evaluation in pandas.

Led the research of end-to-end natural language understanding (NLU) deep neural network (DNN) models for Alexa, simplifying the NLU system by directly predicting user intent and keywords from audio signal. On offline test sets, it reduced error rate by 2% for all input, and by 8% for noisy input, compared to two-stage production baseline.

Launched context-aware NLU DNN models to help Alexa understand multi-turn dialog better. Built representation learning models for encoding user dialog history, and data pipelines for fetching contextual data. It reduced user dissatisfaction rate by up to 10% in online A/B tests, and was deployed into more than 5 Alexa production models.

Developed prototype DNN models in MXNet and PyTorch, run-time model serving and data pipelines in C++, and offline data pipelines for training/testing in Python and Spark.

11-785 Introduction to Deep Learning

Developed and tested homework programs in NumPy and PyTorch.