Link to the paper: https://www.nature.com/articles/s41467-018-07052-w
Neuromorphic engineering is a flourishing research area using computational models of human brains for information processing. Most neuromorphic hardware is built with very-large-scale integration (VLSI) of silicon transistors to mimic neurons and synapses – the basic processing elements in a neural network. Despite Moore’s law scaling, neuromorphic processors still suffer from the trade-off between scalability and biological fidelity. Sophisticated circuits with up to thousands of transistors are needed to emulate the rich dynamics of biological neurons and synapses at the cost of power consumption and chip size.
Memristors are tiny crossbar-shaped electronic devices made of a nanoscale switching material sandwiched between two metal layers. They provide an alternative for advancing neuromorphic engineering. Memristors are typically passive circuit elements that have non-volatile memory of their histories, offering an electronic analogue to biological synapses. Those are the tiny unions between neurons through which learning is achieved by changing connection strength. Another type called active memristors have only volatile (transient) memory. They respond to electrical stimuli in a highly nonlinear fashion and can produce a signal gain in some operating regimes. The implication is that such tiny elements can in principle replace the role of transistors in microprocessors. They are ideal for building scalable and energy-efficient artificial neurons.
Resistive switching in active memristors is driven by entirely different mechanisms than silicon transistors. Vanadium dioxide spontaneously changes from an insulator to a metal at a moderate critical temperature even lower than the boiling point of water. Such a transition can occur within a picosecond and consumes only femto-Joule levels of energy if the device geometry is only tens of nanometers. Conveniently, there is no need for an external heater because passing a minuscule current through the device can trigger the transition by Joule heating.
In fact, active memristors can directly mimic the voltage-gated protein ion channels located on nerve cell membranes that allow passage of ions in and out of the cell. Neurophysiology has taught us that the coordinated opening and closing of a pair of oppositely energized sodium and potassium ion channels can generate action potentials – electrical impulses that carry information processed in the brain. From a circuitry point of view, a spiking memristor neuron can be considered a pair of resistively coupled electrical oscillators, each made of an energized memristor mimicking the protein ion channel, a capacitor mimicking the cell membrane, and a resistor to represent the resistive ion flows in the electrolyte.
Using vanadium dioxide active memristors scaled to within cubes of 100 nanometers on a side, we have developed electronic neuron emulators and demonstrated all three basic classes of biological neuron excitability. We have reproduced 23 types of known biological neuron spiking patterns, and expect more to be demonstrated soon. Active memristor neurons have size and power scaling superior to conventional electronic neurons and show promise as building blocks for an all-memristor neuromorphic computer architecture in the future.
Developing vanadium dioxide suitable for neuron applications was challenging. Vanadium is a transition metal and has multiple oxidation states, leading to a variety of vanadium oxide phases with diverse vanadium-to-oxygen ratios. Only one phase is suitable for our application. We surmounted this challenge with a tightly controlled fabrication process, ensuring a nearly pure vanadium dioxide phase instead of other crystal phases. Another challenge was realizing biological neuron spiking patterns in a compact memristor neuron emulator circuit. We were inspired by neuroscience literature and made innovations in the circuit topology and test protocols to reach our goal. For example, to realize phasic neurons that only respond to the change of an input stimulus rather than its absolute amplitude, we simply replaced the resistor in the input stage with a capacitor or added a capacitor before it, since capacitors allow passing of alternatively changing current but block steady direct current.
We are excited that inherent nonlinear dynamics of such tiny memristive switching devices can reproduce the complex dynamics of real neurons in an extremely compact circuit. The long-term impact may not be fully appreciated yet, since most of today’s artificial neural networks use simplified neuron emulators that only have a small subset of neurocomputational properties compared with real neurons. Since each neuronal spiking behavior derived from evolution has its purpose in the sense of neurocomputation, eliminating the trade-off between scalability and biological fidelity is an important step toward future applications. In the near term, a promising application is real-time energy-efficient edge computing devices such as onboard signal processing for autonomous cars, wearable electronics, and the internet of things.