This article presents a flexible and general optimisation scheme for converting floating-point networks using Variable Bit Depth (VBD) compression as a step towards efficient, low-power inference.
Szabolcs Cséfalvay has been working with Imagination for almost 10 years. His research to date has covered face detection, path tracing, 3D reconstruction and neural networks. His main current research interests are the algorithmic information content and performance optimisation of neural networks, and path tracing related algorithms. He received his Mgr. degree in Computer Graphics from Charles University in Czechia.