Startup's 'oscillator-based' AI technology could be 1,000 times more energy efficient than conventional computing

Researchers have unveiled a new "super-efficient" artificial intelligence (AI) model that can generate images by using a network of physical oscillators rather than traditional calculation-based computing infrastructure.

The new model, known as "Un-0," was created by Unconventional AI, a recently launched technology company founded by a group of prominent AI researchers.

These include Michael Carbin, an associate professor who leads the Programming Systems Group at MIT; Sara Achour, an assistant professor of computer science and electrical engineering at Stanford University; MeeLan Lee, a former Google engineer; and Naveen Rao, former head of AI for analytics company Databricks. The scientists outlined details of this new model in a technical blog post published June 25 on the company's website. The model is also publicly available through GitHub.

Un-0 represents the first proof of concept for the company’s underlying technology, which combines Achour’s work in nonlinear physical substrates — a physical material or hardware device that performs mathematical computations by letting its own natural, continuous laws of physics run — with Carbin’s research into machine learning and physical dynamics. The model itself is a "physical dynamical system," which uses physical motion over time to perform computations.

Oscillator-based AI computing

Conventional computers work by using a system of transistors — tiny electrical switches that can be toggled on to let current flow through them or toggled off to block it. These signals can be read by a computer chip as either a "1" or a "0" — and layering millions, or even billions, of these transistors together allows them to perform complex mathematical equations.

Neural networks, like the kind that power established "stable diffusion" AI image generation tools such as Midjourney or Dall-E, work by layering millions of these calculations on top of each other. Essentially, the system starts with an image made of pure static. The network then examines the static and tries to mathematically predict what visual information (or noise) it needs to subtract from the image to get closer to the target picture. This process is repeated between 20 and 50 times (or occasionally up to 100, although the returns beyond 50 are marginal), with each pass getting closer to a recognizable image.

But whereas those systems use raw mathematics to drive computational processes, Unconventional AI's concept is based on physics and physical movement. At the core of the theory are oscillators — physical devices that produce a continuous waveform, like a metronome.

According to the scientific principles at work, two oscillators that share a physical connection — even if they’re moving at completely different rates — will eventually settle into the same rhythm by mutually influencing each other's movement. By scaling up this principle to thousands of physically linked oscillators — known as a "Kuramoto model" — the startup AI posited that the concept could be used to perform computational tasks such as image generation.

In practice, different patterns of oscillator angles, or "phases," are used to represent different classes of images, such as shoes or trains. The model takes a large collection of oscillators, set at random angles, and then introduces a smaller subgroup of oscillators already set to the specific configuration of angles. This smaller subgroup acts as a prompt for the desired image category.

These oscillators are then physically connected to the wider group, using a preset configuration of different connection strengths. When the oscillators are set into motion, this "control group" naturally pulls the rest of the oscillators toward the desired pattern over time.

After a while, the system takes a snapshot of all of the oscillators' phases, which becomes a grid of numbers. This grid is then fed into a "decoder" system, which translates the numbers into color pixel information to form an image.

Challenging AI's energy consumption

One of Unconventional AI's most eye-catching claims has been its stated goal of having its model use 1,000 times less power than current systems do. In traditional AI image generation models, the calculations needed to perform operations involve flipping billions of tiny transistor switches on and off trillions of times per second, to force the electrical current to move in specific patterns through the circuit.

Although each transistor isn't particularly power-intensive, the cumulative energy usage of a single server running an AI image generation tool can be enormous. For example, it reportedly took 1,287 MWh of energy — enough to power the average U.K. home for more than 475 years — to train OpenAI's GPT-3 model, Alex de Vries, a doctoral candidate at the VU Amsterdam School of Business and Economics, reported in a 2023 article published in the journal Joule.

With the Un-0 model, however, the idea is that rather than forcing transistors to rapidly flip between open and closed, the system consists of a series of closed loops, where the natural path of the current forms the individual oscillators. Because the current is allowed to flow unobstructed, the researchers said in the study, the system is theoretically much more energy efficient than traditional computing architecture.

The company's initial proof-of-concept model uses a simulation of these oscillators running on traditional computing hardware, but the scientists' goal is to one day build their own oscillator-based computing chips on which to run these calculations.

Testing the model

To test the model's performance, Unconventional AI put it through two common AI industry image generation benchmarks: CIFAR-10, a dataset of low-resolution color images split across 10 categories, and ImageNet 64×64, a much larger collection of over 1.2 million pictures at a higher resolution.

These tests allow researchers to measure how closely the generated images match reference material. This metric is known as the model's Fréchet inception distance (FID), where a smaller number represents a higher degree of accuracy. In the study, researchers found that adding more oscillators significantly improved the model's results.

In the CIFAR-10 test, scores ranged from an FID of 11.01 with 1,024 oscillators to 8.76 with 4,096 oscillators. In the more demanding ImageNet 64x64 test, a pool of 6,656 oscillators achieved a score of 8.41 FID, while 16,384 oscillators clocked in at 6.74.

These results are comparable to those of early image generation models, including Google's pioneering BigGAN and OpenAI's iDDPM, which paved the way for its more modern DALL-E tool. However, the study authors stressed that the results "should be read as reference points rather than strictly identical measurements."

"We view Un-0 as a promising first approach with quality that overlaps with that of several established image generation families when they were first introduced to the community," company representatives said in the technical blog post. "Un-0's quality matches where today’s leading generative methods began. Conventional generators are still stronger on absolute quality and parameter efficiency — closing that gap with new algorithms and model architectures is the work ahead."

The scientists released the model weights — the internal mathematical parameters that the model alters as it learns — as well as training and ablation scripts — specialised code files used to build and test the system — allowing other researchers to test the models and run their own simulations. They hope to close the gap with new algorithms and models.

"Taken together, Un-0's system of coupled Kuramoto oscillators offers the promise of learning with physical dynamics at a scale that's beyond what has been done before," they said in the technical blog post. "Un-0 points in the direction of the opportunity for a new computer that exploits physics to achieve our top-line goal of energy efficiency."