Deep neural network training acceleration with geometric data structures

23^rd May 2023, 14:00 Ashton Lecture Theatre
Konstantinos Tsakalidis
University of Liverpool

Abstract

The efficiency of deep learning applications deteriorates significantly as the sizes of the training data and of the neural networks grow larger. In this talk we will identify beyond-state-of-the-art open problems in the intersection of deep learning with computational geometry. Motivated by the recent application of dynamic data structures for geometric halfspace range searching in the acceleration of deep neural networks' training and preprocessing complexity, we revisit efficient algorithms for constructing geometric multi-dimensional data structures and maintaining them dynamically.