Time coding output neurons in digital artificial neural networks

Previous research has proposed several platforms for the direct hardware implementation of neural networks. Most of these platforms employ a small number of special-purpose processors that emulate a given number of neurons. By contrast, this paper proposes a light-weight hardware model, called time coded output (TiCO) neuron, which codes its internal activation as the duration of its output pulses. Since the implementation of such a neuron requires only about 200 logical elements, standard state-of-the-art field-programmable gate arrays may host several hundreds of neurons on a single chip. Due to its output coding and its low footprint, TiCO neurons are particularly suited for the usage in embedded systems.