by Computer Science
Ph.D. in Computer Science student Yifu Deng will present a doctoral dissertation proposal at 10:30 a.m. on Aug. 16. The defense will be held in person in Rekhi 101 and virtually via Zoom.
Deng’s proposal is titled “Near-Memory Processing Accelerators for Approximate Nearest Neighbor Search over Large-Scale Datasets.”
From the abstract:
Modern machine learning techniques can learn high-dimensional vector representations of various objects and serve as the foundation for machine learning-powered information processing systems. Important vector applications include approximate-nearest-neighbors (ANN) search objects for a given query, which provides objects similar to the query. Nonetheless, it is difficult to achieve low-latency ANN search with ever-growing datasets and vector dimensions. The execution time of the vector search phase must be sub-milliseconds or less for production systems. However, it takes > 100ms to query a vector on a billion-scale dataset using a CPU, and the query latency on a 100 million-scale dataset is reported to be approximately 5ms even when a GPU accelerator is used. To address this issue, this project aims to investigate the near-memory processing accelerators to reduce ANN query latency, by minimizing the data movement between processors and memory devices.