Proposal of Adaptive-Learning Neuron Circuits with Ferroelectric Analog-Memory Weights

Novel adaptive-learning neuron circuits are proposed, in which a pulse frequency modulation (PFM) system is used and the interval of output pulses is continuously changed through the learning process. Key components of the circuits are adaptive-learning metal-insulator-semiconductor field-effect transistors (MISFETs) in which the gate insulator films are composed of a ferroelectric material and their polarity is gradually changed by applying input pulses to the gates. In the multiple-input neuron circuits, it is proposed that the adaptive-learning MISFETs be connected in parallel and the function corresponding to the excitatory and inhibitory synapses in a human brain be realized by applying positive and negative pulses to the gates of MISFETs. Finally, the layout of the synapse array between two neuron layers is discussed and the effectiveness of using a silicon-on-insulator (SOI) structure is demonstrated.