Building a 2D-compatible multilayer neural network

Neural network hardware implementations have to reconcile simple hardware topologies with often complex neural architectures. Field programmable neural arrays (FPNA) are defined for that. Their computation scheme creates numerous virtual neural links by means of a limited set of communication links, whatever the device, the arithmetic, and the neural structure. Their concrete use has proved that they allow to have the computation power of standard neural models with a reduced set of neural resources easy to map directly to digital hardware. A simple pattern classification problem is chosen in this paper so as to show how FPNA allow to replace complex standard neural architectures by hardware-friendly neural structures. FPNA have been applied to numerous other problems with similar benefits. They are now applied to high-dimensional real-world applications, such as multiband speech recognition