Google faces internal battle over AI research to speed up chip design

The Google app logo is seen on a smartphone in this illustration taken September 15, 2017. REUTERS/Dado Ruvic/Illustration

Join now for FREE unlimited access to Reuters.com

Register

OAKLAND, Calif., May 2 (Reuters) – Alphabet Inc (GOOGL.O) Google said on Monday it had recently fired a senior director of engineering after colleagues, whose landmark research it tried to discredit, artificial intelligence software, accused him of harassing behavior.

The dispute, which stems from efforts to automate chip design, threatens to undermine Google’s search reputation in the academic community. It could also disrupt the flow of millions of dollars in government grants for AI and chip research.

Google’s search unit has come under intense scrutiny since late 2020 after workers filed open criticism of its handling of staff complaints and publishing practices. Read more

Join now for FREE unlimited access to Reuters.com

Register

The new episode emerged after the scientific journal Nature published “A Graph Placement Methodology for Rapid Chip Design” in June, led by Google scientists Azalia Mirhoseini and Anna Goldie. They found that AI could take a key step in the chip design process, known as floorplanning, faster and better than an unspecified human expert, a subjective benchmark.

But other Google colleagues in a paper published anonymously online in March — “Stronger Baselines for Evaluating Deep Reinforcement Learning in Chip Placement” — found that two alternative commodity software-based approaches outperformed AI. One beat it on a well-known test and the other on a proprietary Google rubric.

Google declined to comment on the leaked draft, but two employees confirmed its authenticity.

The company said it refused to publish Stronger Baselines because it failed to meet its standards, and soon after fired Satrajit Chatterjee, one of the main movers behind the work. He declined to say why he fired him.

“It’s unfortunate that Google took this turn,” said Chatterjee attorney Laurie Burgess. “It’s always been his goal to have transparency about science, and he’s been urging Google to fix it for two years.”

Google researcher Goldie told The New York Times, which first reported the firing on Monday, that Chatterjee had harassed her and Mirhoseini for years by spreading false information about them.

Burgess denied the allegations and added that Chatterjee did not disclose Stronger Baselines.

Patrick Madden, an associate professor specializing in chip design at Binghamton University who read both papers, said he had never seen a paper before Nature that lacked a good point of comparison. .

“It’s like a benchmark problem: everyone gets the same puzzle pieces and you can compare how close you get to whatever it takes,” he said. “If they produced results on a standard benchmark and they were stellar, I would sing their praises.”

Google said the comparison with a human was more relevant, and software licensing issues prevented it from mentioning the tests.

Studies by large institutions such as Google in well-known journals can have an outsized influence on the funding of similar projects in the industry. A Google researcher said the leaked document unfairly opened the door to questions about the credibility of any work published by the company.

After “Stronger Baselines” emerged online, Zoubin Ghahramani, vice president of Google Research, wrote on Twitter last month that “Google supports this work published in Nature on ML for Chip Design, which has been independently replicated, open source and used in production at Google.”

Nature, citing a UK public holiday, had no immediate comment. Madden said he hopes Nature will come back to the publication, noting that the peer reviewer notes show at least a demand for results on the benchmarks.

“Somehow it never happened,” he said.

Join now for FREE unlimited access to Reuters.com

Register

Reporting by Paresh Dave. Editing by Gerry Doyle

Our standards: The Thomson Reuters Trust Principles.

Comments are closed.