AN ANALOG NEURONLESS RECONFIGURABLE FEED FORWARD NEURAL NETWORK

Mario Costa - Davide Palmisano - Eros Pasero

   Today Feed Forward neural Networks (FFNs) use paradigms tied to mathematical frame-works really hard to be transferred in a hardware realisation. This fact makes analog neural integrated circuits heavy to design. Here we propose an alternative formal and general model that can use the native computational properties of the basic electronic circuits without any previous assumption about their behavior. A practical framework is described to train such analog FFNs off-chip. This is especially useful whenever the weight storage elements cannot be re-programmed on the fly at a high rate. Then we will present how this formal model allowed us the development of innovative and non conventional neural architectures we named "neuron-less" (N-LESS).

Keywords: analog neural networks, analog integrated circuits, analog memories, analog signal processing

[full-paper]