Google Fires Senior Researcher In Latest AI Row

Google on Monday acklowedged it recently fired a senior researcher who had questioned a paper the company published on the use of artificial intelligence in designing computer chips.

The researcher, Satrajit Chatterjee, had led work on a paper that cast doubts upon key claims in a high-profile study Google published in the journal Nature last year that found artificial intelligence could design certain parts of computer chips better than humans.

Chatterjee was fired in March shortly after Google reportedly told his team their paper would not be published.

Google confirmed in a statement that Chatterjee’s employment had been “terminated with cause” but declined to elaborate. The termination was reported by the New York Times earlier on Monday.

AI controversy

The company said Chatterjee’s paper did not meet its standards for publication.

The casts light upon the latest divisions within Google’s artificial intelligence research division, less than two years after the company fired two researchers, Timnit Gebru and Margaret Mitchell, who had investigated ethical issues with artificial intelligence.

Those firings led to critiques by Google staff over the company’s handling of personnel complaints and publication practices.

Earlier, in 2018, Google employees protested an AI contract with the US Department of Defence, leading the company to pull out of the project.

The latest matter is significant because Google considers AI one of the most important technologies to its future.

Chip design

The research in question also has a direct bearing on millions of dollars in US government grants into research into AI and computer chips.

In June 2021 Nature published “A graph placement methodology for fast chip design”, research led by Google researchers Azalia Mirhoseini and Anna Goldie.

The paper found AI could complete a key step in the chip design process, called floorplanning, faster and better than an unspecified human expert.

But Chatterjee’s contrasting paper, “Stronger Baselines for Evaluating Deep Reinforcement Learning in Chip Placement”, detailed two alternative approaches using basic software that outperformed the AI.

In the paper, which was leaked anonymously in March, Chatterjee cited a well-known test and a proprietary Google test to make the comparison. Two Google staff confirmed the paper’s authenticity to Reuters.

‘Transparency’

Laurie Burgess, an attorney for Chatterjee, told Reuters and the New York Times that Chatterjee’s objective was “transparency” about the science.

Google’s Goldie told the New York Times that Chatterjee had spread misinformation about her and Mirhoseini for years. She said their paper had been peer-reviewed by Nature and that Google had used their methods to build chips that were in use at Google’s data centres.

Burgess denied Goldie’s allegations and added that Chatterjee did not leak “Stronger Baselines”.

Patrick Madden, an associate professor focused on chip design at Binghamton University who has read both papers, told Reuters that Google’s paper in Nature appeared to lack a good comparison point, since the AI was compared only to an unspecified human expert and lacked benchmarks.

‘Independently replicated’

Google said comparison to a human was more relevant and that software licensing issues prevented it from bringing tests into the paper.

After “Stronger Baselines” was widely circulated online last month, Zoubin Ghahramani, a vice president at Google Research, wrote on Twitter that “Google stands by this work published in Nature on ML for Chip Design, which has been independently replicated, open-sourced, and used in production at Google”.