How can generative models create/hallucinate visual represenation useful for sovling task? Can generative models generalize well?

Research Scientist, MIT.

2020 - present.

An application of generative models for creating an interactive design by putting human in the loop.

Research Scientist, MIT.

2020 - present.

Understanding capabilities and biases of generative models: What are the capabilities of generative models, what biases they encode and how these biases speak to the underlying data distributions? Can generative models generalize well? How about extrapolationg to out of distribution?

Research Scientist, MIT.

2019 - 2020.

Instance semantic segmenation across domains leveraged by generative models : We develp computer vision algrithms that can be used across different domains, e.g., recycling of cellpones, cancer, objects and people, and autonomous driving, all with one general algorithm. We desire an AI framework that can be trained without any ad-hod settings per data domain.

Postdoc Associate, Research Scientist, MechE MIT.

June 2018 - 2020.

Understanding Virtual Identities Across Platforms: We develop computational tools for understanding visual aestheics and analytics of how users in different cultures represent themselves in virtual platforms, e.g. social networks and games.

Postdoc Associate, CSAIL MIT.

June 2016 - 2017.

Mining Visual Evolution of Web Design: We trained the Neural Networks on a large-scale web page dataset to disver design evolution over 21 years of web design.

Postdoc Associate, CSAIL MIT.

March 2016 - present.

Web Page Gist: Web pages are designed interfaces that differ from natural scenes in interesting and potentially important ways. They usually consist of a complex mixture of multiple images, organized text, and graphical elements. Here, we ask an important and new question: What can we see at a glance on a web page? We provide intuitions about what stimulus cues might be driving performance and discuss the implications of our findings for vision science and for human-computer interaction.

Postdoc Associate, CSAIL MIT.

Janurary 2015 - present.

Color Semantics: developed a probabilistic generative modeling framework to associate colors and linguistic concepts.

Postdoc Associate, CSAIL MIT.

Research Assistant, Purdue University.

December 2012 – 2016.

Learning Visual Balance: developed a probabilistic modeling framework to learn visual balance from a large-scale dataset of aestheically highly-rated images.

Research Assistant, Purdue University, 2014.

Autonomous Color Theme Extraction: developed a suite of algorithms for extracting color themes the way people perform the task by utilizting human vision saliency models.

Research Assistant, Purdue University, 2014.

Document Aesthetics: performed research on Automatic Design of Magazine Covers and Quantifying Aesthetics of Design.

Research Assistant, Purdue University.

Sponsored by HP Labs, Palo-Alto, June 2010 – 2013.

Image Analysis: developed a software tool for Autonomous Inspection of Indigo Press Digital Front-End GUI Screen Shots.

Research Assistant, Purdue University.

Sponsored by Indigo, Boise, January – May 2010.

Visualization Techniques: implemented and investigated a number of clustering techniques on large scale data sets for 3D Flow Volume Visualization.

Research Assistant, Purdue University.

Funded by NSF, May – December 2009.

Game Design and Development: developed a 3D game to enhance the English skills of international students.

Research Assistant, Purdue University.

Funded by Purdue Graduate School, August – December 2011.

Game Design and Development: developed a 2D/3D game to teach American Sign Language to students with deafness.

Research Assistant, Purdue University.

Funded by NSF, 2008 – 2009.