Unconventional AI Debuts Oscillator-Based Un-0 Model Series

Unconventional AI Inc. has developed an artificial intelligence architecture that improves the power efficiency of image generation models.

This technology is the basis of the company’s new neural network series, Un-1. released on Thursday.

Unconventional AI is led by Chief Executive Officer Naveen Rao (pictured, second from left), former corporate vice president of Intel AI Platform Group. In December, the company raised $475 million from a consortium that includes Amazon.com founder Jeff Bezos. We are developing chips that can run AI models using significantly less power than current graphics cards.

Not all processors are based on standard silicon transistors. Several startups are developing so-called in-memory computing devices that use a combination of transistors, capacitors, and small energy storage devices. Quantum processors, on the other hand, often use materials such as sapphire instead of silicon.

The Un-0 model series is part of Unconventional AI’s efforts to develop more efficient AI chip architectures. The company says the Un-0 is optimized to work with oscillators rather than standard transistor-based circuits. An oscillator is a device that emits a signal, such as an electrical pulse, at regular time intervals.

Unconventional AI says it may be possible to assemble large numbers of tiny oscillators to build machine learning accelerators. Such components are already produced in large quantities in the semiconductor industry, as they are used in chips such as central processing units. In particular, the CPU relies on an oscillator to set the pace at which other circuits perform calculations.

Un-0 does not run on a physical oscillator chip. Instead, it uses thousands of simulated oscillators to generate images. Oscillators are linked together. This means that signals generated by one virtual device affect the output of other virtual devices, and vice versa.

There are six Un-0 models that vary in size and output quality. The smallest one consists of 1,024 virtual oscillators, and the largest one has 16,384 virtual oscillators. Unconventional AI trained its model using two open source datasets containing thousands of images optimized for machine learning projects: CIFAR-10 and ImageNet-64.

The training process unfolded differently than a standard AI project. Typically, developers approach the task by optimizing AI model components such as weights. In contrast, Unconventional calibrated the way Un-0’s simulated oscillators influenced each other and the frequencies at which they generated signals.

The workflow for standard AI models to generate media files starts with an image containing random noise. Un-0 starts the process in the same way, but the subsequent steps are different.

First, a small group of oscillators generate instructions that tell the model what type of image to create. This instruction prompts Un-0’s other oscillators to interact. According to Unconventional AI, interactions generate a series of numbers that can be combined to create an image.

The company ran a series of benchmark tests to evaluate the Un-0’s output quality. They determined that the model was comparable to “the quality of the leading traditional image generation methods at the time they were first published.” As a result, Unconventional AI believes that future advances have the potential to significantly improve the power efficiency of AI applications.

photograph: Lightspeed Venture Partners