Researchers at the National Institute of Standards and Technology (NIST) have built a superconducting switch that “learns” like a biological system and could connect processors and store memories in future computers operating like the human brain.
Above – Illustrations showing the basic operation of NIST’s artificial synapse, which could connect processors and store memories in future neuromorphic computers operating like the human brain. A synapse is a connection or switch between two brain cells. NIST’s artificial synapse is a tiny metal cylinder that processes incoming electrical spikes to customize spiking output signals based on a tunable internal design. Researchers apply current pulses to control the number of nanoclusters pointing in the same direction, as depicted in the “disordered” versus “ordered” illustrations. This design, in which different inputs alter the alignment and resulting output signals, is inspired by how the brain operates. Credit: NIST
The NIST switch, described in Science Advances (link is external), is called a synapse, like its biological counterpart, and it supplies a missing piece for so-called neuromorphic computers. Envisioned as a new type of artificial intelligence, such computers could boost perception and decision-making for applications such as self-driving cars and cancer diagnosis.
Superconducting computing chips modeled after neurons can process information faster and more efficiently than the human brain. This is a key benchmark in the development of advanced computing devices designed to mimic biological systems. And it could open the door to more natural machine-learning software, although many hurdles remain before it could be used commercially.
Schneider’s team created neuron-like electrodes out of niobium superconductors, which conduct electricity without resistance. They filled the gaps between the superconductors with thousands of nanoclusters of magnetic manganese.
By varying the amount of magnetic field in the synapse, the nanoclusters can be aligned to point in different directions. This allows the system to encode information in both the level of electricity and in the direction of magnetism, granting it far greater computing power than other neuromorphic systems without taking up additional physical space.
The synapses can fire up to one billion times per second — several orders of magnitude faster than human neurons — and use one ten-thousandth of the amount of energy used by a biological synapse.
A synapse is a connection or switch between two brain cells. NIST’s artificial synapse—a squat metallic cylinder 10 micrometers in diameter—is like the real thing because it can process incoming electrical spikes to customize spiking output signals. This processing is based on a flexible internal design that can be tuned by experience or its environment. The more firing between cells or processors, the stronger the connection. Both the real and artificial synapses can thus maintain old circuits and create new ones.
Even better than the real thing, the NIST synapse can fire much faster than the human brain—1 billion times per second, compared to a brain cell’s 50 times per second—using just a whiff of energy, about one ten-thousandth as much as a human synapse. In technical terms, the spiking energy is less than 1 attojoule, lower than the background energy at room temperature and on a par with the chemical energy bonding two atoms in a molecule.
“The NIST synapse has lower energy needs than the human synapse, and we don’t know of any other artificial synapse that uses less energy,” NIST physicist Mike Schneider said.
The new synapse would be used in neuromorphic computers made of superconducting components, which can transmit electricity without resistance, and therefore, would be more efficient than other designs based on semiconductors or software. Data would be transmitted, processed and stored in units of magnetic flux. Superconducting devices mimicking brain cells and transmission lines have been developed, but until now, efficient synapses—a crucial piece—have been missing.
The brain is especially powerful for tasks like context recognition because it processes data both in sequence and simultaneously and it stores memories in synapses all over the system. A conventional computer processes data only in sequence and stores memory in a separate unit.
The NIST synapse is a Josephson junction, long used in NIST voltage standards. These junctions are a sandwich of superconducting materials with an insulator as a filling. When an electrical current through the junction exceeds a level called the critical current, voltage spikes are produced. The synapse uses standard niobium electrodes but has a unique filling made of nanoscale clusters of manganese in a silicon matrix.
The nanoclusters—about 20,000 per square micrometer—act like tiny bar magnets with “spins” that can be oriented either randomly or in a coordinated manner.
“These are customized Josephson junctions,” Schneider said. “We can control the number of nanoclusters pointing in the same direction, which affects the superconducting properties of the junction.”
The synapse rests in a superconducting state, except when it’s activated by incoming current and starts producing voltage spikes. Researchers apply current pulses in a magnetic field to boost the magnetic ordering, that is, the number of nanoclusters pointing in the same direction. This magnetic effect progressively reduces the critical current level, making it easier to create a normal conductor and produce voltage spikes.
The critical current is the lowest when all the nanoclusters are aligned. The process is also reversible: Pulses are applied without a magnetic field to reduce the magnetic ordering and raise the critical current. This design, in which different inputs alter the spin alignment and resulting output signals, is similar to how the brain operates.
Synapse behavior can also be tuned by changing how the device is made and its operating temperature. By making the nanoclusters smaller, researchers can reduce the pulse energy needed to raise or lower the magnetic order of the device. Raising the operating temperature slightly from minus 271.15 degrees C (minus 456.07 degrees F) to minus 269.15 degrees C (minus 452.47 degrees F), for example, results in more and higher voltage spikes.
Crucially, the synapses can be stacked in three dimensions (3-D) to make large systems that could be used for computing. NIST researchers created a circuit model to simulate how such a system would operate.
The NIST synapse’s combination of small size, superfast spiking signals, low energy needs and 3-D stacking capability could provide the means for a far more complex neuromorphic system than has been demonstrated with other technologies, according to the paper.
Neuromorphic computing promises to markedly improve the efficiency of certain computational tasks, such as perception and decision-making. Although software and specialized hardware implementations of neural networks have made tremendous accomplishments, both implementations are still many orders of magnitude less energy efficient than the human brain. We demonstrate a new form of artificial synapse based on dynamically reconfigurable superconducting Josephson junctions with magnetic nanoclusters in the barrier. The spiking energy per pulse varies with the magnetic configuration, but in our demonstration devices, the spiking energy is always less than 1 aJ. This compares very favorably with the roughly 10 fJ per synaptic event in the human brain. Each artificial synapse is composed of a Si barrier containing Mn nanoclusters with superconducting Nb electrodes. The critical current of each synapse junction, which is analogous to the synaptic weight, can be tuned using input voltage spikes that change the spin alignment of Mn nanoclusters. We demonstrate synaptic weight training with electrical pulses as small as 3 aJ. Further, the Josephson plasma frequencies of the devices, which determine the dynamical time scales, all exceed 100 GHz. These new artificial synapses provide a significant step toward a neuromorphic platform that is faster, more energy-efficient, and thus can attain far greater complexity than has been demonstrated with other technologies.