Research

We study the problem of 3D object generation. We propose a novel framework, 3D Generative Adversarial Network (3D-GAN), leveraging recent advances in volumetric convolutional networks and generative adversarial nets.

We develop a computational model that explains how people make causal judgments in physical scenes by mentally simulating counterfactual outcomes and comparing those to what actually happened.

Our goal is to develop a socially intelligent team coacher agent that helps humans communicate, strategize, and work together efficiently.

In order to be able to design synthetic organs that function autonomously, we will need to engineer artificial tissue homeostasis. To control the size of these artificial tissues, two major mechanisms will have to be engineered.

Automatic speech recognition (ASR) has been a grand challenge machine learning problem for decades. Our ongoing research in this area examines the use of deep learning models for distant and noisy recording conditions, multilingual, and low-resource scenarios.

Developing techniques to allow self-driving cars and other AI-driven systems to explain behaviors and failures.

We develop image analysis algorithms for whole-heart cardiac MRI in patients with severe congenital heart disease.

Integrating multi-modal clinical data and using recurrent and convolution neural networks to predict when patients will need important interventions.

We develop novel algorithms for neuroimage segmentation, registration, and joint modeling of brain structure with clinical and genetic data.

Knitting is the new 3d printing. It has become popular again with the widespread availability of patterns and templates, together with the maker movements. Lower-cost industrial knitting machines are starting to emerge, but we are still missing the corresponding design tools. Our goal is to fill this gap.

A system that can help you to design our own drone!

We use computational models of biological pathways in disease to estimate the range of possible responses to various hypothetical drugs, given patient variability and limited pathway-relevant data.

Our goal is to develop collaborative agents (software or robots) that can efficiently communicate with their human teammates. Key threads involve designing algorithms for inferring human behavior and for decision-making under uncertainty.

Existing methods for cloning and recombination of DNA enable construction of arbitrary sequences. However, the sequential nature of these techniques makes them time-consuming and expensive. Furthermore, while the transformation of an existing plasmid into a host strain can be reliable when a selection marker is used, there are many current limitations: the number of different plasmids that can be co-transformed is limited by the choice of markers and compatible origins of replication; plasmids are less stable than chromosomal DNA and are difficult to maintain indefinitely without mutation; and cistronic interactions cannot be designed since each new nucleotide sequence added is on an unconnected DNA molecule. To overcome these limitations, we are designing reconfigurable chromosomes consisting of both fixed and variable regions. While the fixed region is carefully optimized and tuned ahead of time, the variable region can be modified in the field, at the point-of-use, leading to rapid and on-demand realization of novel biocircuits with many different phenotypes.

A new approach for designs in multi-material fabrication

Developing state-of-the-art tools that process 3D surfaces and volumes

Help robots learn faster by providing demonstrations when they need help

Almost every object we use is developed with computer-aided design (CAD). While CAD programs are good for creating designs, using them to actually improve existing designs can be difficult and time-consuming.

The creation of low-power circuits capable of speech recognition and speaker verification will enable spoken interaction on a wide variety of devices in the era of Internet of Things.

Our goal is to develop better metagenomic binning by origin species of fragments of sequenced environmental DNA.

We develop algorithms for fetal MRI interpretation, to enable noninvasive fetal monitoring from MRI.

Inspired by the history of the user interface for personal computing, we are developing new interaction models for personal fabrication.

We develop, parameterize, and validate a model for tumor growth inhibition using in vivo mouse data and study the effects of modeling uncertainty and inter-individual variability on drug candidate efficacy predictions.

We develop machine learning algorithms to automatically and quantitatively assess the severity of pulmonary edema from chest x-ray images.