Browsing by Author "Verburgh, Cornel"
Now showing 1 - 1 of 1
Results Per Page
Sort Options
- ItemParallelising inference on cluster graphs(Stellenbosch : Stellenbosch University, 2020-12) Verburgh, Cornel; Du Preez, Johan; Barnard, Arno; Stellenbosch University. Faculty of Engineering. Dept. of Electrical and Electronic Engineering.ENGLISH ABSTRACT: Cluster graphs (CGs), a technique used in machine learning, are computationally very complex and often are slow to converge to an answer. There are also various approaches that lead to convergence. Some approaches converge faster than others. However, finding the optimal approach is non-trivial. We investigated the use of a new parallel inference algorithm, comparing it to the current state of the art inference algorithms, such as parallel splash belief propagation and residual belief update. These were tested on several CGs, ranging from a simple sudoku solver to a PGM that does satellite image denoising. The results from these tests were as follows: for satellite image denoising a 5.3 times speedup was achieved, plateauing at 10 threads; for the sudoku solver the speedup ranged from 2.0 times to 4.0 times speed-up, normally plateauing around six threads. From the results it is clear that the algorithms perform better the more clusters are present. It is also important to note that hyperthreading affects the speed-up, as is shown by the reduction in CPU instructions per cycle the more threads are being used.