Scientists have long drawn inspiration from the brain to design computer systems. Some researchers have recently gone even further by making computer hardware with a brain-like structure. These “neuromorphic chips” have already shown great promise, but they used conventional digital electronics, limiting their complexity and speed. As chips get larger and more complex, the signals between their individual components are backed up like cars on a congested highway and reduce computation to a crawl.
Now, a team from the National Institute of Standards and Technology (NIST) has demonstrated a solution to these communication challenges that could one day allow artificial neural systems to run 100,000 times faster than the human brain.
The human brain is a network of approximately 86 billion cells called neurons, each of which can have thousands of connections (called synapses) with its neighbours. Neurons communicate with each other using short electrical impulses called spikes to create rich, time-varying patterns of activity that form the basis of cognition. In neuromorphic chips, electronic components act like artificial neurons, routing spike signals through a brain-like network.
By getting rid of the conventional electronic communication infrastructure, the researchers designed networks with tiny light sources at each neuron that broadcast optical signals to thousands of connections. This scheme can be particularly energy efficient if superconducting devices are used to detect single particles of light called photons – the smallest possible optical signal that could be used to represent a spike.
in a new Natural electronics article, NIST researchers have realized for the first time a circuit that behaves much like a biological synapse but only uses single photons to transmit and receive signals. Such a feat is possible using superconducting single-photon detectors. Computation in the NIST circuit occurs when a single photon detector encounters a superconducting circuit element called a Josephson junction. A Josephson junction is a sandwich of superconducting materials separated by a thin insulating film. If the current flowing through the sandwich exceeds a certain threshold value, the Josephson junction begins to produce small voltage pulses called fluxons. Upon detection of a photon, the single-photon detector pushes the Josephson junction above this threshold and the fluxons accumulate as current in a superconducting loop. Researchers can tune the amount of current added to the loop per photon by applying a bias (an external current source powering the circuits) to one of the junctions. This is called synaptic weight.
This behavior is similar to that of biological synapses. The stored current serves as a form of short-term memory, as it provides a record of how many times the neuron has produced a spike in the near past. The duration of this memory is defined by the time it takes for the electric current to decay in the superconducting loops, which the NIST team says can range from hundreds of nanoseconds to a few milliseconds, and likely beyond. This means that the hardware could be suitable for problems occurring on many different time scales – from high-speed industrial control systems to quieter conversations with humans. The ability to set different weights by changing the bias of the Josephson junctions allows for longer term memory that can be used to make gratings programmable so that the same grating can solve many different problems.
Synapses are a crucial computing component of the brain, so this demonstration of superconducting single-photon synapses is an important step on the way to realizing the team’s full vision of superconducting optoelectronic networks. However, the lawsuit is far from over. The team’s next step will be to combine these synapses with on-chip light sources to demonstrate fully superconducting optoelectronic neurons.
“We could use what we’ve demonstrated here to solve computational problems, but the scale would be limited,” said NIST project leader Jeff Shainline. “Our next goal is to combine this advance in superconducting electronics with solid-state light sources. This will allow us to establish communication between many more elements and solve important and consequential problems.
The team has already demonstrated light sources that could be used in a complete system, but further work is needed to integrate all components on a single chip. The synapses themselves could be improved by using detector materials that operate at higher temperatures than the current system, and the team is also exploring techniques to implement synaptic weighting in larger-scale neuromorphic chips.
The work was funded in part by the Defense Advanced Research Projects Agency.
Article: S. Khan, BA Primavera, J. Chiles, AN McCaughan, SM Buckley, AN Tait, A. Lita, J. Biesecker, A. Fox, D. Olaya, RP Mirin, SW Nam and JM Shainline. Single-photon superconducting optoelectronic synapses. Natural electronics. Published online October 6, 2022. DOI: 10.1038/s41928-022-00840-9
#NISTs #superconducting #hardware #expand #braininspired #computing