Kyuyoung Kim

I am an MS student with AI focus in the Computer Science Department at Stanford University, where I work in the Stanford Intelligent Systems Laboratory led by Prof. Mykel Kochenderfer.

Previously, I was a senior software engineer at Google working on developing machine learning algorithms for natural language generation used to localize the Google Assistant. I was also part of the Display Ads developing the ads backend system and improving auction algorithms. I have a BS in computer science with a minor in applied mathematics from Cornell University.

Email  /  GitHub  /  Google Scholar  /  CV  /  LinkedIn

Research

I am interested in machine learning, decision making under uncertainty, and optimization.

Adaptive Algorithms for Efficient Risk Estimation of Black-Box Systems

Adaptive algorithms for efficient risk estimation of black-box systems and their applications to evaluating autonomous vehicle policies in simulation.

A Deep Reinforcement Learning Approach to Rare Event Estimation

Deep reinforcement learning-based adaptive importance sampling algorithms for efficient estimation of the probability of rare events for sequential decision making systems.

BEHAVIOR-1K: A Benchmark for Embodied AI with 1,000 Everyday Activities and Realistic Simulation.

A human-centric benchmark for Embodied AI in simulation with 1,000 everyday activities, a diverse dataset of 5,000+ objects and 50 scenes, and a simulation environment, OmniGibson, that reaches high levels of simulation realism.

Using Machine Translation to Localize Task Oriented NLG Output

Developed a set of methods to localize at scale a task-oriented natural language application like the Google Assistant using neural machine tranlsation. The proposed approach includes development of an appropriate sequence model with an encoding scheme of structured input, fine-tuning on in-domain translations, and automatic error detection.

Taskmaster-1: Toward a realistic and diverse dialog dataset

Built a Wizard-of-Oz system to collect 13,215 task-based dialogs comprising six domains. Released the dataset for NLP research.

Finding Overlapping Communities From Subspaces

Developed spectral methods for finding overlapping community structures in networks.

Other Projects

These include coursework, side projects, and unpublished research work.

Shallow and Deep Methods for Time-Series Modeling

Studied statistical and deep learning approaches to time-series modeling and analyzed the methods in terms of the amount of data needed, forecasting accuracy, and ability to incorporate covariates into the models.

Effectiveness of Combining Deep Reinforcement Learning Algorithms

Explored the effectiveness of combining multiple deep reinforcement learning algorithms in training stability, convergence speed, and ability to handle partial-observability.

Effectiveness of Recurrent Network for Partially-Observable MDPs

Use of recurrent neural networks in deep Q-learning to handle partial-observability in MDPs.

Design and source code from Leonid Keselman's website