In Search or a Robust Digital CNN System
Wenhai Fang, Cheng Wang, Lambert Spaanenburg · 2006
Microelectronics promises a high component density and low power dissipation to embedded systems. Unfortunately such a component will always suffer from various error types that make the chip respond differently from its functional simulation. This is especially true for cellular neural networks (CNN), which makes the determination of robust, low-precision parameters to guarantee small footprint and reliable operation an important design consideration. This paper describes the digital word width effects in a CNN implementation that must be considered to achieve a small size for a reliable system. It discusses the automated design space exploration using a field-programmable gate-array (FPGA) implementation to perform an optimal CNN parameters selection