MIT researchers claim that they successfully created analog synapses one million times faster than those found in the human brain.
Just as digital processors need transistors, and our biological brains need neurons and synapses to make connections between different brain regions, analog requires programmable resistors. According to a press release, these resistors can form a network of analog synapses and neurons when properly configured.
These analog synapses are not only incredibly fast, but also highly efficient. And that is important because as digital neural networks become more powerful and advanced, their energy requirements increase, thus increasing their carbon levels significantly.
Gain incredible speed
Researchers have achieved nanosecond speeds – faster than biological synapses in the human brain – by eliminating conventional organic carriers and choosing high-tech inorganic phosphosilicate glass (PSG).
“The speed is really amazing. Normally we would not apply such strong fields on devices to prevent them from turning to ashes. But on the contrary, Proton ended up shutting down at huge speeds across the device stack, especially a million times faster than what we had before. And this motion is not damaged by the small size and low proton mass. “It’s almost like a phone call.” Author and MIT postdoc Murat Onen speaks.
Ju Li, senior author of research papers and professor of nuclear science, explains:
“The activity potential in biological cells increases and decreases with the time scale of milliseconds, since the voltage difference of about 0.1 V is inhibited by the stability of water. Here we apply up to ten volts on the film. “Special hard glass with a nanoscale thickness that guides the protons without permanently damaging them. And the stronger the field, the faster the ionization device.”
Because PSG can withstand high voltages without breaking the proton can travel at ridiculous speeds, while it is also very energy efficient. In addition, the material is widely available and simple to manufacture; So it is not only the fastest option but also the most practical option.
These programmable resistors greatly accelerate neural network training while reducing the cost and energy required to complete the training. This can accelerate the development of in-depth study models that can be applied to the detection of fraud, medical imaging analysis, or self-driving vehicles.
“Once you have an analog operating system, you will no longer be a training network for others running. Instead, you will become a training network with unprecedented complexity that no one else could afford, and therefore run beyond them all. “On the other hand, this is not a faster car, this is a spacecraft.”
MIT scientists hope their findings will help boost the field of analog-deep learning, an area of growth and innovation of artificial intelligence.