AI confounds humans with improved chip designs

Described in Nature Communications, the wireless technology processors feature a combination of standard electronic circuits as well as electromagnetic structures including antennas, resonators, and signal splitters. Deep learning systems were tasked with creating new chips, resulting in strange designs featuring unusual patterns of circuitry, often with improved efficiency over existing chips. According to the researchers, the AI-created chip designs were unintuitive and unlikely to be developed by a human mind.

“We are coming up with structures that are complex and look random shaped and when connected with circuits, they create previously unachievable performance. Humans cannot really understand them, but they can work better,” said lead researcher Kaushik Sengupta, a professor of electrical and computer engineering and co-director of NextG, Princeton’s industry partnership programme for next-generation communications.

These circuits can be engineered towards more energy efficient operation or to make them operable across an enormous frequency range that is not currently possible. Furthermore, the  AI method synthesises inherently complex structures in minutes, while conventional algorithms may take weeks. In some cases, the new methodology can create structures that are impossible to synthesise with current techniques.

According to Sengupta, the circuitry in an advanced chip is so small, and the geometry so detailed, that the number of possible configurations for a chip exceeds the number of atoms in the universe. There is no way for a person to understand that level of complexity, so human designers do not try. They build chips from the bottom up, adding components as needed and adjusting the design as they build.

“Classical designs, carefully, put these circuits and electromagnetic elements together, piece by piece, so that the signal flows in the way we want it to flow in the chip,” said Sengupta.

“By changing those structures, we incorporate new properties. Before, we had a finite way of doing this, but now the options are much larger.”

While AI is clearly a powerful tool, human oversight is still required, in part because that AI can make faulty arrangements as well as efficient ones. It is possible for AI to hallucinate elements that don’t work, at least for now.

“There are pitfalls that still require human designers to correct,” said Sengupta. “The point is not to replace human designers with tools. The point is to enhance productivity with new tools. The human mind is best utilised to create or invent new things, and the more mundane, utilitarian work can be offloaded to these tools.”