Artificial neurons improve computer chips by replicating biological functions

Researchers from the USC Viterbi School of Engineering and the School of Advanced Computing have developed an artificial neuron that mimics the complex electrochemical behavior of biological brain cells.

This innovation, documented in Nature Electronics, is a breakthrough in neuromorphic computing technology. This innovation could enable orders of magnitude reduction in chip size, reduce energy consumption by orders of magnitude, and advance artificial general intelligence.

Unlike existing neuromorphic chips based on traditional digital processors or silicon technology, which simply simulate neural activity, these artificial neurons physically embody or emulate the analog dynamics of their biological counterparts. Just as neurochemicals initiate brain activity, they can be used to initiate computations in neuromorphic (brain-inspired) hardware devices. They differ from previous iterations of artificial neurons, which were simply mathematical equations, in that they are physical replicas of biological processes.

The study, led by USC computer and electrical engineering professor Joshua Yang, who led the seminal paper on artificial synapses more than a decade ago, introduces a new type of artificial neuron based on so-called diffusion memristors. The paper in Nature Electronics explores how such artificial neurons could enable a new class of chips that power nearly all electronics today, complementing and enhancing today’s silicon-based technologies that rely on the movement of electrons for computations.

Instead, the diffusion device introduced by Yang and colleagues to build neurons will rely on the movement of atoms. Such neurons operate more like how our brains work, are more energy efficient, and could enable new chips that usher in what is known as artificial general intelligence (AGI).

How the device works

In biological processes, the brain uses both electrical and chemical signals to drive activity within the body. Neurons, or nerve cells, start out as electrical signals that, when they reach the spaces or gaps at the ends of the neurons (synapses), are converted into chemical signals to convey and process the information. When the information is passed on to the next neuron, some of those signals are converted back into electrical signals across the neurons.

This is a physical process that Yang and colleagues were able to emulate with high fidelity in several key aspects. A big advantage: Diffused memristor-based artificial neurons require the space of only a single transistor, rather than the tens or hundreds of transistors used in traditional designs.

In particular, in biological models, ions or charged particles help generate electrical signals that cause movement within neurons. In the human brain, such processes rely on chemicals (such as ions) such as potassium, sodium, and calcium to force this action.

In the paper, Yang, director of USC’s Center of Excellence in Neuromorphic Computing, emulates the process of using silver ions in oxides to generate electrical pulses to perform computing for activities such as movement, learning, and planning.

“Although the ions in artificial synapses and neurons are not exactly the same, the physics governing their movement and dynamics are very similar,” he says. “Silver is easily diffusible, giving us the dynamics we need to emulate biological systems so that we can achieve the function of neurons with very simple structures.”

The new device, which can create a brain-like chip, is called a diffusion memristor, after the movement and dynamic diffusion of ions created by the use of silver.

He added that the team chose to use ion dynamics to build an artificial intelligence system because “that’s what happens in the human brain, and for good reason: the human brain is an ‘evolutionary winner’, the most efficient intelligent engine.”

“It’s more efficient,” he said, explaining, “It’s not that our chips and computers aren’t powerful enough to do anything. It’s just that they’re not efficient enough. They use too much energy.”

This is especially important given the level of energy required to run large software models with vast amounts of data, such as machine learning in artificial intelligence.

Yang went on to explain that, unlike the brain, “our existing computing systems are not designed to process large amounts of data or learn on their own from just a few examples. One way to improve both energy and learning efficiency is to build artificial systems that operate according to principles observed in the brain.”

If you’re looking for pure speed, Electron is the best choice for fast processing as it runs the latest in computing. But, he explains, “Ions are a better medium than electrons for embodying brain principles. Because electrons are lightweight and volatile, computing with electrons allows for software-based rather than hardware-based learning, which is fundamentally different from the way the brain works.”

In contrast, “the brain learns by moving ions across membranes, enabling energy-efficient adaptive learning directly in hardware, or more precisely in what people call ‘wetware’,” he says.

For example, young children can learn to recognize handwritten digits by just seeing a few examples of each digit, whereas computers typically require thousands of digits to accomplish the same task. But the human brain accomplishes this amazing learning while consuming only about 20 watts of power, compared to the megawatts of power required by today’s supercomputers.

This new method brings us one step closer to mimicking natural intelligence.

Professor Yang pointed out that the silver used in the experiment is not easily compatible with conventional semiconductor manufacturing, and alternative ionic species need to be investigated to achieve similar functionality.

The efficiency of these diffused memristors includes not only energy but also size. A smartphone typically has about 10 chips, but there are billions of transistors and switches that control the on/off (0s and 1s) that power calculations.

“Instead [with this innovation]we only use one transistor footprint for each neuron. We are designing building blocks that will ultimately reduce chip size by orders of magnitude and reduce energy consumption by orders of magnitude. So in the future, AI will be able to run sustainably with similar levels of intelligence without consuming unsustainable energy. ” says Yang.

“Now that we have demonstrated artificial synapses and neurons as functional, compact building blocks, the next step is to integrate them in large numbers and test how closely they can mimic the efficiency and function of the brain.”

“Even more exciting is the prospect that such brain-faithful systems could help uncover new insights into how the brain itself functions,” Yang concludes.