Active Memristor Neurons for Neuromorphic Computing
Abstract
To rival mammals intelligence, future artificial intelligence (AI) needs to be both strong capable of self-learning with raw inputs, and broad capable of handling multiple contexts rather than narrowly restricted problems. It is unlikely that AI implemented on conventional computing platforms will eventually achieve both goals. This is because even if the brains connectivity is fully understood and reproduced, the neuron and synapse computing primitives, if built with non-biomimetic CMOS circuits, are not capable to emulate the rich dynamics of biological counterparts without sacrificing the energy consumption and size. Memristors may provide an alternative approach to realize strong and broad AI. The nonvolatile, stochastic and adaptive switching dynamics make the passive memristor a close electronic analogue to biological synapses. Owing to their superb scalability, memristor crossbars can potentially reach the synapse density of the brain. A complementary device, active memristor, can be used to construct electronic equivalent of the biological neurons. Active memristors show volatile resistive switching and are locally active within a hysteretic negative differential resistance (NDR) regime incurrent-voltage characteristics. The NDR regime endows active memristors a.c. signal gain, a must have for neural signal processing. Previous demonstrations of neuron circuits based on active memristors only showed basic spiking behaviors exhibited by leaky integrate-and-fire (LIF) models, and thus offer limited computational complexity. Using scalable VO2 active memristors, we show that memristor neurons actually possess most of the known biological neuronal dynamics and all three classes of neuron excitability. Furthermore, our simulations show that the size and power scaling of memristor neurons project toward biologically competitive neuron density and energy efficiency. Our work indicates that scalable and biomimetic active memristor neurons, in combination with
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 25, 2019
- Accession Number
- AD1075276
Entities
People
- Wei Yi