Blog

Unlike conventional digital processors or existing neuromorphic chips based on silicon technology that merely simulate neural activity, these artificial neurons physically embody or emulate the analog dynamics of their biological counterparts. Just as neurochemicals initiate brain activity, chemicals can be used to initiate computation in neuromorphic, or-brain-inspired hardware devices. By being a physical replication of the biological process, they differ from prior iterations of artificial neurons that were solely mathematical equations.

The work, led by USC Computer and Electrical Engineering Professor, Joshua Yang, who also led the work in a seminal paper on artificial synapses over a decade ago, introduces a new type of artificial neuron based on the so-called “diffusive memristor”. The Nature Electronics paper explores how such artificial neurons can enable a new class of chips that complement and augment today’s silicon-based technologies, which power nearly all modern electronics and rely on the movement of electrons for computation. Instead, the diffusive device introduced by Yang and colleagues to build the neurons would rely on movement of atoms. Such neurons can enable newer chips that would operate more similarly to how our brain works, would be more energy efficient and could lend themselves to usher in what’s known as artificial general intelligence (AGI).

How it works:
In the biological process, the brain uses both electrical and chemical signals to drive action in the body. Neurons or nerve cells start out with electrical signals that when they reach the space or gap at the end of the neuron called the synapse, the electrical signals are converted into chemical signals into order to pass on and process the information. Once the information crosses to the next neuron, some of those signals are once again converted to electrical signals through the body of the neuron. This is the physical process that Yang and colleagues have succeeded in emulating with high fidelity in several critical aspects. The big advantage: their diffusive memristor-based artificial neuron requires only the space of a single transistor, rather than the tens to hundreds used in conventional designs.

In particular, in the biological model, ions or charged particles help generate the electrical signals to cause action within the neuron. In the human brain, such processes rely on chemicals (e.g., ions) like potassium, sodium, or calcium to force this action.

In the current paper, Yang, who is Director of the Center of Excellence on Neuromoprhic Computing at USC, uses silver ions in oxide to generate the electrical pulse and emulate the processes to perform computing for activities such as movement, learning, and planning.
“Even though it’s not exactly the same ions in our artificial synapses and neurons, the physics governing the ion motion and the dynamics are very similar,” says Yang.

Yang explains, “Silver is easy to diffuse and gives us the dynamics we need to emulate the biosystem so that we can achieve the function of the neurons, with a very simple structure.”  The new device that can enable a brain-like chip is called the “diffusive memristor” because of the ion motion and the dynamic diffusion that occurs with the use of silver.

He adds, the team chose to utilize ion dynamics for building artificial intelligent systems  “because that is what happens in the human brain, for a good reason and since the human brain, is the ‘winner in evolution–the most efficient intelligent engine.”

“It’s more efficient,” says Yang.

This is critical, explains Yang, “It’s not that our chips or computers are not powerful enough for whatever they are doing. It’s that they aren’t efficient enough. They use too much energy.” This is particularly relevant given the level of energy needed to run large software models with a huge amount of data like machine learning for artificial intelligence.

Yang goes on to explain that unlike the brain, “Our existing computing systems were never intended to process massive amounts of data or to learn from just a few examples on their own.
One way to boost both energy and learning efficiency is to build artificial systems that operate according to principles observed in the brain.”

If you are looking for pure speed, electrons that run modern computing would be the best for fast operations. But, he explains, “Ions are a better medium than electrons for embodying principles of the brain. Because electrons are lightweight and volatile, computing with them enables software-based learning rather than hardware-based learning, which is fundamentally different from how the brain operates.”

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

For example, a young child can learn to recognize handwritten digits after seeing only a few examples of each, whereas a computer typically needs thousands to achieve the same task. Yet, the human brain accomplishes this remarkable learning while consuming only about 20 watts of power, compared to the megawatts required by today’s supercomputers.

Potential Impact:
This new method is one step closer to mimicking natural intelligence.

Yang noted that silver used in the experiment is not readily compatible with conventional semiconductor manufacturing, and that alternative ionic species will need to be investigated for similar functionalities.

The efficiency of these diffusive memristors include not only the energy, but size. Normally one smart phone has about 10 chips but billions of transistors or switches that control the on/off or 0’s and 1’s that underpin computation.

“Instead [with this innovation], we just use a footprint of one transistor for each neuron. We are designing the building blocks that eventually led us to reduce the chip size by orders of magnitude, reduce the energy consumption by orders of magnitude, so it can be sustainable to perform AI in the future, with similar level of intelligence without burning energy that we cannot sustain,” says Yang.

Now that we have demonstrated capable and compact building blocks, artificial synapses and neurons, the next step is to integrate large numbers of them and test how closely we can replicate the brain’s efficiency and capabilities. “Even more exciting,” says Yang, “is the prospect that such brain-faithful systems could help us uncover new insights into how the brain itself works.”