Deep Alignment Network - from MIMD to SIMD Platform

Authors: Kivanc Yuksel, Prof. Władysław Skarbek
Institute: Warsaw University of Technology, Dept. of Electronics and Information Technology

The paper considers the following software engineering problem for digital media: given a software tool for processing tensor signals, like Deep Neural network (DNN) defined for MIMD architecture (Multi Instruction, Multi Data), redefine this algorithm to SIMD architecture (Single Instruction, Multiple Data). While for mapping multiple instructions, the standard signal processing approach is applied, for mapping tensors of any dimensionality, 2D RGBA textures (Red, Green, Blue, and Alpha channels) are used as the target data structure. To illustrate the tensor mapping concept, Deep Alignment Network (DAN), contemporary important application for Human Computer Interfacing, is selected and its efficiency analyzed. The testbed for comparisons of DAN's MIMD and SIMD architectures, was based on Javascript (MIMD) and WebGL (SIMD) software platforms. It appears that expected speedup of SIMD versus MIMD architecture is on the reasonable level: 350 image frames per minute versus seven image frames per minute.

Keywords: Deep neural networks, facial salient points, deep alignment network, JavaScript, WebGL