Neuromorphic computing is a promising concept for low-power, energy-efficient spiking networks with the capability of self-learning, adaptation, and recognition of speech, gesture, and objects. Development of the neuromorphic computing technology is currently facing 2 main barriers: First, there is no comprehensive understanding how the brain really works; and second, there is no consensus about what technology might provide synaptic and neural circuits at the best tradeoff between cost, power consumption, and performance. The resistive switching memory (RRAM) is one of the main contender for neuromorphic components, thanks to its low current operation, small area and tunable resistance. Demonstration of brain-inspired learning feature with RRAM synapses may pave the way for future high performance, low cost neuromorphic processor and brain-in-a-chip.
This talk will report on the recent advances on neuromorphic hardware for unsupervised learning of visual patterns. First, I will describe a RRAM synapse capable of spike-timing dependent plasticity (STDP) with one-transistor/one-resistor (1T1R) structure. Second, I will show the learning and recognition capability of a neuromorphic chip with a microcontroller neuron and an array of RRAM synapses. Learning of single/multiple patterns, tracking of patterns, and recognition will be shown in hardware. These results support RRAM as a promising technology for future neuromorphic processors.