In a recent study published in Science, a team of researchers at the Massachusetts Institute of Technology (MIT) have produced what are known as protonic programmable resistors, also known as the building blocks of analog deep learning, which are capable of processing data 1 million times faster than human brain synapses. This study is intriguing because it continues to expand the field of artificial intelligence since analog deep learning allows for faster computation with a fraction of energy being used in an age where time, energy, and money required to build such complex neural network models continues to shoot up.
Programmable resistors in deep analog learning are comparable to the transistors being the core elements for digital processors. By layering these programmable resistors in repeating arrays, scientists can produce a network of analog artificial “neurons” or “synapses” that will carry out computations like a digital neural network.
To carry out the study, this multidisciplinary team of MIT researchers used a practical inorganic material in the manufacturing process that allows their devices to operate 1 million times faster than previous versions, also equivalent to 1 million times faster than human brain synapses. Along with the increased processing speed, this inorganic material also makes the resistor very energy-efficient while being reconcilable with silicon manufacturing techniques. This notable change allows for manufacturing devices at the nanometer scale, opening doors for combination with deep-learning applications involving commercial computing hardware.
"The working mechanism of the device is electrochemical insertion of the smallest ion, the proton, into an insulating oxide to modulate its electronic conductivity. Because we are working with very thin devices, we could accelerate the motion of this ion by using a strong electric field, and push these ionic devices to the nanosecond operation regime," Bilge Yildiz, who is a Breene M. Kerr Professor in the departments of Nuclear Science and Engineering and Materials Science and Engineering at MIT, and a co-author on the study.
These programmable resistors greatly increase the training speed of a neural network while also significantly cutting down on cost and energy to carry out that training. This study could allow scientists to produce deep learning models much faster with the hope of applying this technology in uses such as self-driving cars, fraud detection, or medical analysis.
As always, keep doing science & keep looking up!