Collaborative Research Projects

All of the planets in our solar system orbit around the Sun. Planets that orbit around other stars are called exoplanets. Exoplanets are very hard to see directly with telescopes. Instead, astronomers can observe indirectly measuring how the brightness of the star changes during a transit. This can help them figure out the size of the planet. The Box Least Squares (BLS) periodogram is a statistical tool used for detecting transiting exoplanets and eclipsing binaries in time series photometric data. The current implementation of the BLS algorithm is rather slow for the data sets we hope to process in the future. This is due to the lack of certain optimizing transformations, e.g., locality enhancing transformations (tiling) as well as use of appropriate resources. With that we mean that the existing code cannot be executed on a parallel computer system based using CPUs and/or GPUs. Transforming the current periodogram computing software (essentially the BLS algorithm) into a modern parallel code that can exploit all levels of (nested) parallelism and balance the execution for the imbalanced and input dependent cases. We propose to develop a parallel version of the BLS algorithm using the STAPL and Charm4Py parallel programming environments and then try to combine them into a single, better performing code.

Generative modeling of media is at the forefront of artificial intelligence. Computers can now synthesize text, images, and video, at high quality, even generating curiously novel outputs at times. Models such as Open AI’s DALL·E 2 or Google Imagen have been prominent in the news today, and are producing fascinating digital artwork.

However, one area that has not seen a significant advance is that of audio generation. This is not due to lack of interest: automatic generation of audio signals is a strongly sought-after technology in the entertainment industry, and generation of synthetic acoustic data is a crucial step in the development of technologies that range from speaker and mic design, to deep sea drilling monitoring and biomedical diagnostic modeling.

This research project will tackle the necessary work to adapt modern generative models to time-series in general. This will primarily include enabling long-term temporal dependencies and addressing the ubiquitous problem of superposition.

At a high level, patterns set by systems like DALL·E 2 will be followed. A generative model that is conditioned on text and audio will be developed, and would be able to generate audio data based on textual descriptions but also based on audio inputs themselves.

This research will explore data-driven rendering -- where one makes realistic images and sequences of images by compositing real data assets in various ways.  Recent work from the project team will be adapted to produce data-driven rendering systems that can be used in Virtual Reality or even Augmented Reality applications.  This project aims to produce a scene representation pipeline that will allow a user to take a small number of pictures of a room, and make a model; then take a small number of images each of some objects to produce models of those objects; then insert the objects into the room.  At each step, the user should just need to place and move objects within the room. The final model of room and objects should be small, fast to render, accurate under arbitrary change of viewpoint, and will admit stereo rendering.

The goal is for the model to show what happens when a particular spotlight is turned on or off.  The project team’s recent work on image based reshading and relighting will be expanded to make a composite model that is consistently shaded and is relightable, allowing the AR user to perceive a more realistic scene.