Spinning electronic neurons compute like a brain

electronics

Editor of the website for technological innovation – 04.10.2022

deposit calculation

A new analog computing platform enabled the construction of a neuromorphic processor made only with components that are already commercially available.

A full prototype “tank computer” with an analog hardware architecture that mimics the brain, demonstrating a low-power, high-speed alternative to today’s digital logic-based computers.

The brain works on a completely different calculation principle than processing binary data. It acts on many different inputs at the same time, with each neuron connected to many others in a time-varying cascade network. If viewed as a biological computer, the brain does not work with ones and zeroes, but in a non-linear analog domain. Mathematically, this logic can be modeled as a “reservoir computer”.

“Reservoir computing is a form of neuromorphic computing that was first proposed in the 2000s. It is most easily understood using the analogy of a lake. If you throw stones into a lake, the waves of each stone interact to produce a complex pattern of ripples on the surface of the water, like a faint memory that contains information about your stone throwing activity.

“By analyzing the waves, you can understand how many stones you threw, the time intervals between them, and even the size of each stone. The lake is the ‘reservoir,’ and the ripple pattern is the state matrix of the reservoir. Similar processes, including mapping and high-throughput, have recently been identified in the brain mouse, suggesting that our brain can also function as a complex reservoir computing system,” explained Professor He Qian from Tsinghua University in China.

The implementation is electronic, but the operation can be described as a “rotating neuron” system.
[Imagem: Xiangpeng Liang et al. – 10.1038/s41467-022-29260-1]

rotating neurons

Several platforms for reservoir calculation have already been implemented, but they always require complex systems for entering data and reading results.

Professor Qian’s team simplified everything by building an electronics-based reservoir computer system with built-in input and readout capabilities.

Using a bunch of electronic circuits, including a readout module consisting of an array of memoristors—the electronic memory components most commonly used in neuromorphic computing—the team built a circuit that models a container of spinning neurons that respond to inputs in a sequential and interconnected manner.

“Our main challenge was to find an equivalent pairing of neural network algorithm and hardware that could be implemented,” Qian said. “In this case, the electronic rotating hardware has a similar function to a lake tank, mapping the low-dimensional inputs into a high-dimensional space, which can be linearly separated using a simple linear classifier.”

Not only did it work, it worked using a thousand times less energy than any other reservoir computing system demonstrated so far – in this first demonstration, the system was used to accurately predict the future sequence of a chaotic time series.

“Electronics is more than just a binary transistor,” Qian says. “There is still much to explore in the rich dynamics offered by electronics for neuromorphic reservoir computing, which is particularly interesting for brain-inspired computing and artificial intelligence due to its low complexity and low training cost.”

Other news about: