UCLA

A crossbar array of non-volatile memory devices based on carbon nanotubes (CNTs) is developed to emulate the localized, massively parallel, power-efficient learning capabilities of biological synapses. By incorporating signal processing, memory, and learning into a single element the data transmission overhead present in typical digital systems is eliminated. The electronic synapse is operated as a two-terminal device, in a fixed-amplitude pulse mode, and its resistance depends on the polarities and duration of overlap of the pulses occurring on the two electrodes. The materials and structure of the synapse is optimized to symmetrically increase or decrease its resistance state, or synaptic weight, depending on the correlation between the signals on each terminal and to retain the resistance state in the presence of uncorrelated signals on the terminals. An array of 400 synapses are fabricated and interfaced with an 'integrate-and-fire' circuit to implement a learning algorithm. A speech recognition task is demonstrated to study the overall performance and power efficiencies of the circuit elements. The energy efficiency during unsupervised learning is >10^14 operations per joule, exceeding previous learning demonstrations with resistive memory-based crossbar arrays.