- Succeeded in getting VLMs (LLaVA, Deepseek, Qwen2.5, mPLUG-Owl3) to generate 3D hand geometry description by fine-tuning them on data generated from syntactic data generation pipeline. - Co-authored two papers (one under review) that train deep-learning-based framework to perform 3D hand-based downstream tasks. - Adapted Scene-Representation-Transformer to be compatible with the multiview 3D hand geometry representation. - Succeeded in adapting the deep-learning-based 3D human-avatar generation pipeline “gDNA” towards posed generative 3D hand reconstruction.

更多

Achieved a 66x runtime speedup on a pitch-segmentation algorithm for audio frequency data by optimizing the bottlenecks in code and restructuring the code to be able work with the Numba package

更多

- Implemented efficient versions of the contextual bandit algorithms LinUCB and Thompson Sampling on a news article recommendation setting - Gained a computation speed up of 2x and 3.4x on performing recommendation task on the Yahoo R6A dataset using parallel versions of Thompson Sampling and LinUCB algorithm respectively - Demonstrated that parallelizing LinUCB and Thompson sampling algorithms lead to insignificant loss in rewards in exchange for a massive speed gain - Presented the research results in a research fair organized by UNIST

更多