Modern artificial intelligence models have an environmental impact, especially when it comes to image generation.

But as demand for these models grows, UCLA engineering professor Aydogan Ozcan and his team have reimagined a way to generate AI images with less energy.

“Computer engineering, in specific, provides a lot of interesting opportunities to exercise math and physics for building systems and devices,” said Ozcan, the Volegenau Chair of Engineering Innovation in the Samueli School of Engineering. “That’s why I’m very interested in using light to build imaging systems, microscopy systems and computing systems.”

His work on optical generative AI models has been a natural offshoot of research on materials, computing and optics – or light – Ozcan said.

Generative AI models are often slow and power-intensive because the models are iterative – meaning the model repeatedly tailors potential answers until it comes up with the response for the user, which Ozcan said inspired his optical AI work.

Ozcan’s optical AI model takes randomly generated patterns made by a computer program and displays them like a “barcode,” he said. A laser then hits the “barcode” and the light is reflected onto a decoder, which reflects the light once more onto a display – creating a unique image, he added.

The optical AI model’s process is more energy efficient than typical generative image models, Ozcan said.

The optical AI does not start off being able to make every image, however, meaning the lab has to build “teacher” AI models, Ozcan said. These models show the optical AI how to turn random noise into usable images until it masters image creation, he added.

The iterative processes of many of today’s AI models use graphics processing units, which digitally process many numbers for AI queries.

GPUs are located inside of specialized computers called servers, which are housed in data centers, said Liam de Villa Bourke, a fourth-year environmental science student. Data centers for other types of tech use central processing units, but GPUs – which are comparatively more energy intensive – are necessary for AI, he added.

Jalesa Rosario, a doctoral student in information studies, said she recognizes that AI has more than just environmental implications. Generative AI has the potential to scrape user data and enable cybersecurity threats, she added.

“Does that small, imagined potential outweigh the current disasters it is causing?” she said. “The answer is to be determined.”

Ozcan said he worries about AI misleading people and contributing to unemployment. However, he added that he believes AI is going to make lives better for humanity by helping to advance medicine and technology – and that regulation can solve many of the problems AI has the potential to create.

Ozcan said he has modest ambitions for optical generative models like the one he created, adding that he envisions they could be used in specific cases for visual displays.

“Public opinion must be aware of it (AI) and demand from governments policies about it before it’s late,” he said. “The only way out of catastrophic AI hurting our citizens, our people, our kids and the next generations is to be educated, proactively understand pros and cons.”