tech news 247 default logo image

In late 2018, Google AI researchers Anna Goldie and Azalia Mirhoseini got the go-ahead to test an elegant idea. Google had invented powerful computer chips called tensor processing units, or TPUs, to run machine learning algorithms inside its data centers—but, the pair wondered, what if AI software could help improve that same AI hardware?

The project, later codenamed Morpheus, won support from Google’s AI boss Jeff Dean and attracted interest from the company’s chipmaking team. It focused on a step in chip design when engineers must decide how to physically arrange blocks of circuits on a chunk of silicon, a complex, months-long puzzle that helps determine a chip’s performance. In June 2021, Goldie and Mirhoseini were lead authors on a paper in the journal Nature that claimed a technique called reinforcement learning could perform that step better than Google’s own engineers, and do it in just a few hours.

The results won media coverage and notice in the world of semiconductors. In a commentary on the Nature paper, Andrew Kahng, a professor at UC San Diego, predicted the technique would be quickly adopted by chipmakers. “To long-time practitioners,” he wrote, “Mirhoseini and colleagues’ results can indeed seem magical.” Google’s data centers now contain TPU chips created with help from Morpheus. Samsung and Nvidia have independently said they also use reinforcement learning to optimize chip designs.

Yet in parallel to their success, according to five current and former Google employees, and documents seen by WIRED, Mirhoseini and Goldie spent years fending off a series of unproven claims that their results were wrong or even falsified.

Satrajit Chatterjee, a more senior researcher at Google, used the cover of scientific debate to undermine the women personally, the employees claim. They spoke anonymously because they were not authorized to discuss company matters. Multiple complaints about Chatterjee’s behavior toward the women were made to Google’s personnel department, and he received a written warning, some employees said, but he continued to criticize the women’s results.

The conflict came to a head in March of this year, after Chatterjee sought permission from research managers to publish a public rebuttal of Mirhoseini and Goldie’s Nature study. A committee of senior executives formed to review that paper denied his request, saying its results did not refute the earlier work. The same month, Chatterjee was fired.

On May 2, Goldie posted a document on an internal Google discussion list describing the committee’s rejection of Chatterjee’s paper and accusing him of a series of unproven attacks on the Morpheus coleads and their work. “Sat Chatterjee has waged a campaign of misinformation against me and Azalia for over two years now,” Goldie wrote. “He started a campaign to discredit our work [and] baselessly alleged that Azalia and I fabricated and falsified results.”

The document was posted in a thread where Googlers were reacting to a New York Times article that first reported Chatterjee’s firing, alongside complaints from his attorney that Google researchers were attacking him to shut down a scientific discussion. Most Googlers who joined the thread expressed support for the two women and their work; some current and former Google researchers did so publicly on social media.

Laurie M. Burgess, Chatterjee’s attorney, declined to make her client available for interview and denied he had acted inappropriately, saying he had evidence Google improperly suppressed his work. Burgess said she did not want to share that evidence and did not respond to an email asking detailed questions about Chatterjee’s behavior toward Goldie and Mirhoseini and their project.

When asked about Chatterjee, Google spokesperson Jason Freidenfelds provided a company statement confirming he was “terminated with cause.” Freidenfelds also provided a statement from Zoubin Ghahramani, vice president at Google Research, saying that “we firmly uphold our standard for respectful discourse among our researchers.” Gharamani’s statement did not mention Chatterjee by name.

The episode adds to a series of recent internal conflicts at Google that suggest the freewheeling, engineer-centric culture it celebrated as a startup has left the company unprepared for some challenges of being a multinational with more than 100,000 staff.

Google hired Satrajit Chatterjee in 2018 as a senior machine learning researcher. He was previously a senior vice president at hedge fund Two Sigma and had also worked at Intel. When Chatterjee joined, Mirhoseini and Goldie already worked in the company’s most prominent machine learning lab, Google Brain. Chatterjee joined a separate, smaller research group inside Google’s research division.

The two women did not work directly with Chatterjee, but in 2019, Goldie’s internal document claims, he asked to manage the Morpheus project. After being politely declined, employees say, Chatterjee began raising doubts about the pair’s work with senior researchers they needed to collaborate with or win support from, suggesting their work was wrong or even fabricated.

As a more senior employee, Chatterjee’s questions could be influential. As a result, employees say, other senior staff at times became skeptical of Goldie and Miroseini’s work, questioning their results.

The effect was to turn Miroseini and Goldie’s work at Google into a stressful, split reality, insiders claim. At the same time as running a successful project with support from Google’s chip designers, they say the pair had to do extra work to respond to allegations that their results were wrong or even false.

Chip design teams at Google and elsewhere are generally cautious by nature, because nanoscale fabrication is expensive and any errors in a chip cannot be fixed once it has been carved into silicon. Google has said that TPUs have enabled breakthroughs in its AI research and services, and rents out the chips through its cloud unit. Yet Chatterjee’s criticisms of Morpheus continued even after Google’s hardware leaders decided they trusted it enough to let it help design the next generation of the company’s TPUs.

In May 2021, a Google employee posted to an internal email list asking if anyone had applied machine learning to the design of circuit boards. Mirhoseini replied to say that Morpheus could help. But Chatterjee chimed in to claim that older techniques outperformed machine learning tools and that commercially available chip design tools provided the best results.

Jeff Dean, Google’s head of AI, joined the discussion to say that Morpheus was already being used to design the next generation of TPU chips. The technology had won out in extensive tests against human chip experts and commercial chip design tools, Dean said, while also attaching a slide deck of the results.

Dean also linked to the team’s recently published, peer-reviewed Nature study. It reported that the Morpheus team’s code laid out blocks of TPU circuits better than Google engineers using commercial chip design tools. The authors did not disclose the details of those chip segments, saying they were confidential to Google, but also included results for an open source processor design freely available to anyone. The paper’s results were later replicated by another research team inside Google, and code for the experiments was open sourced.

The Nature paper became a target for Chatterjee, sources say, and he started work on a paper claiming to rebut its findings with coauthors from inside and outside of Google. Late in 2021 he sought permission to publish it outside the company. Google assembled a special committee of five senior chip and AI experts outside the management chains of those involved in the dispute to review the claimed rebuttal paper and determine its fate. The group also attempted to replicate results from the Morpheus team.

An anonymous draft of Chatterjee’s paper that leaked online questions the comparisons made in the Nature paper. It presents results from different experiments, using different measurements, claiming to show that older software for arranging circuits on a chip could outperform Morpheus algorithms.

After three months of review, Google’s committee said Chatterjee could not release his critique outside the company. The group said the experiments and data presented did not refute the Nature work, in part because they didn’t fully re-create its experiments. But the committee did give Chatterjee and his coauthors the opportunity to revise the paper.

In March this year, the group ruled that a revised version was only slightly improved, and still unpublishable. “Scientific publications coming out of Google research must be held to very rigorous standards, which is why the committee put significant effort into reviewing this work before coming to a decision,” an email seen by WIRED from the committee said. Later in March, Chatterjee was fired.

Goldie mentioned the committee’s decision in the document posted to an internal discussion list early in May. She also said that Google managers had investigated and replicated her team’s results and determined that Chatterjee’s claims “were completely unfounded.”

Google’s vice president for research Zoubin Ghahramani joined the discussion and thanked Goldie. He also posted a statement sent to The New York Times, and later WIRED, saying that Google stands behind the Nature paper and that the authors had helped the company design more useful and efficient AI hardware. The project’s researchers have published additional peer-reviewed papers on the technology and filed related patent applications.

Another person familiar with the dispute offered a different view, saying that Google leaders’ support for the original paper makes it difficult to discuss valid questions about chip design in a rebuttal. The person said they had never seen Chatterjee act inappropriately.

Internal conflicts at Google over what managers allow researchers to publish externally have broken into the open before. The company said last year that it would tighten its prepublication review process after outcry at how the coleads of Google Brain’s ethical AI team, Timnit Gebru and Margaret Mitchell, were forced out of the company. Executives had demanded they withdraw or remove their names from a research paper coauthored with academics that was critical of AI technology used in search and other Google products.

Some AI researchers inside and outside Google derided a memo from Dean, the company’s head of AI, that claimed Gebru and Mitchell’s paper “didn’t meet our bar for publication.” Thousands of Googlers and outside AI experts signed a public letter criticizing the company, and the disputed paper was later accepted at a leading peer-reviewed conference, without Google affiliations.

It is unclear whether a version of Chatterjee’s paper will be formally published or peer reviewed outside Google. Inside the company, algorithms from Goldie and Mirhoseini’s project are still being used to help design future generations of chips.

Source: Wired – AI