Developing warm-started randomized HOSVD algorithms optimized for GPUs to efficiently conduct lossy compression of streaming tensor data.
Available materials: GitHub Presentations (Slides)
This project develops warm-started randomized higher-order SVD algorithms optimized for GPUs to efficiently compress and decompress streaming tensor data, for example the kind produced in massive scientific simulations (e.g. weather simulations). The goal is to produce error-controlled Tucker decompositions of streaming high-dimensional tensors ("snapshots") efficiently and independently to prepare for a compression pipeline, leveraging prior solutions to accelerate SVD on new data arrivals using stochastic methods.
Note: Current results are restricted to CPU implementations.
Compression fidelity on Hurricane Isabel (slides)
Compression fidelity analysis measuring PSNR and reconstruction quality across all 13 physical variables.
Preliminary experiments (slides)
Began experimental validation on synthetic and real tensor streams. Initial benchmark comparisons with baseline methods and exploration of GPU kernel optimizations.
Ideation phase and presentation (slides)
Initial conceptualization of the warm-started rHOSVD algorithm for streaming tensor decomposition.