Zachary Bamberger

Software Engineer

Google•  July 2019 - October 2021

* Implemented an on-device transformer for language modeling tasks (e.g., next word and emoji predictions for GBoard) that is 10x smaller than BERT-Small. * Developed an offline reinforcement learning library for Google News and YouTube recommender systems. I implemented offline DQN and Policy Gradient (REINFORCE) estimators (tested via Markov Decision Processes). * Built a distributed computing pipeline to introduce realistic synthetic typos to existing datasets. I used this pipeline to create datasets which test language model robustness to typo noise. * Implemented a distributed computing pipeline to unbalance federated datasets, thereby enabling more realistic simulations of federated language models. * Created a dashboard to display metrics from live experiments with language models. This dashboard refreshes in fixed sub-daily intervals and saves a team of 40+ engineers from running daily operations that take at least 10 minutes to run.

Technion - Israel Institute of Technology, Haifa

Computer Science, MS•  October 2021 - Present

Research: Semi-supervised pre-training in NLP, Discourse, Persuasive Argumentation, Interpretability and Robustness of NLP systems. Coursework: Seminar in NLP, Advanced Topics: Machine Learning for Human Behavior, Advanced Topics: Advanced Deep Learning (Vision and Graphs). Teaching: Introduction to Machine Learning, Introduction to Natural Language Processing.