Tension Inside Google Over a Fired AI Researcher’s Conduct
This is a story about how Google’s AI research team tried to improve its own technology. The story includes a lot of technical details, but the main point is that the researchers wanted to make improvements to the chip design. Their ideas didn’t work out as expected, though.
Reinforcement learning is used by Google’s chipmakers to make chips faster. This method uses artificial intelligence to create new designs. The researchers claim this method works better than Google’s own methods.
The researchers found a new type of artificial intelligence chip that could make computers perform tasks much faster than before. This chip was made using the brain cells of mice. The chip was used to create an AI system called Mirahosseini. The chip is also known as a neural network processor (NNP).
Samsung and Nvidia both use reinforcement learning to optimize their chips. However, the researchers at Google did not do this. Instead, they used an AI technique called deep reinforcement learning. Deep reinforcement learning uses neural networks to learn how to play games. This technique was first developed by two Stanford University professors in 2012.
Satrajit Chatterjee is a more senior researcher at google. He uses the cover of scientific debate as an excuse to attack the women personally. He was warned by the personnel department, but continues to do this.
The conflict came to a head when he tried to publish a public rebuttel of Mirhoseini & Goldie’s Nature study, but the committee of senior executives said otherwise. He was fired.
Sat Chatterjee has waged war against Azalia and me for over two years now. He started a campaign to discredit my work and baselessly accused me of fabricating and falsifying results. Most Googlers joined the thread expressing support for us and our work.
Some current and former Google researchers were fired after they revealed misconduct by Larry Page and Sergey Brin. Their lawyer says they have evidence that Google improperly suppressed their work. She does not want to share this information.
When asked about Chatterjee, Google spokesperson Jason Freidenfels confirmed that he was terminated with cause. He also provided a statement from Vice President of Google Research, Zoubin Ghahramani, stating that we firmly uphold our standard for respectful discourse among our researchers. Mr. Ghahramani’s statement did not include Chatterjee by name.
Chatterjee was hired by Google in 2018. He was previously a Senior Vice President at hedge fund Two Sigma. He had also worked at Intel before joining Google. Chatterjee joined the Machine Learning Research Group within the Google Research division. This group is responsible for developing new algorithms and techniques for machine learning applications. The episode adds to a string of recent internal conflicts at the company that suggest the freewheeing, engineer-centric culture that it celebrated as a startup may be leaving the company unprepared for some of the challenges of being a multinational corporation with over 100,000 employees.
The two women did not work closely with Chatterjee, who managed the Morpheus project. He tried to get them fired, but failed. Their work was not wrong, but he thought it was fake.
The senior employee is Chatterjee. He is skeptical about the results of Miroseini and goldie’s work. Employees say that he questioned their results. This made them feel stressed and pressured. Their work turned out to be stressful and difficult.
Chip design teams at Google and others are cautious by nature. Nanoscale fabrication is expensive, and any errors in a fabricated chip cannot be corrected once it has been carved out of silicon. Google has said TPUs have enabled groundbreaking advances in its AI research and products. But Chatterjee’s criticisms continue even after Google’s hardware team decides it trusts TPUs enough to let them help design the next generation.
In May 2021, a Google engineer asked about using machine learning to improve chip designs. He was told by others that this was impossible because older methods were better. Then he was told by another engineer that commercial chip design software did a better job than machine learning. Finally, he was told by yet another engineer that machine learning was useless.
Jeff Dean says that Morpheus is already being used to design next-generation TPU chips. He shows slides of the test results comparing Morpheus’ performance to other chip design tools.
Dean is a good friend who helps people. He is a member of the team that published a paper in Nature about how well the Morpheus chips perform compared to other designs. This paper is a public document. It is not secret information. It is not proprietary. It is not private. It is not trade secret. It is not intellectual property. It is not copyright. It is not patentable. It is not copyrighted. It is not patented. It is not a trade secret. It is a public document. So why is this paper being treated as if it is something else? Why is it being treated as if it should be kept secret? Why is it being kept secret? Why is this paper being treated like a trade
An anonymous draft of Chattergee’s paper that leaked online claims that the comparison made in the Nature paper was wrong. He presents results from different experiments using different measurements, claiming that the older software for arranging circuits performed better than the Morpheus algorithm. After three months of reviewing, Google’s committee said he could not release his critique beyond the company. The group also said the experiments and data he presented do not refute the Nature work. However, the committee gave him an opportunity to revise the paper and make changes.
In March this year, the committee ruled that a revised version of the paper was only slightly improved, but still unpublishable because it did not meet the standards set by the committee. Later in March, the committee fired Chatterjee.
Goldie was right when she said that the committee decided to fire her. But Google’s VP of Research Zoubin Ghahrani didn’t say anything about this. Instead he praised Goldie’s work and tried to discredit her critics. He even wrote a letter to the NYT.
Internal conflict at Google over what managers allowed researchers to publish externally has broken into the open. The company said last years that it would tighten its pre-publication review process after outcry over how the co-leaders of Google Brain’s ethical artificial intelligence team, Timnit Geebru and Margaret Mitchell, was forced out of the company by executives who demanded they withdraw or remove the names of themselves from a research paper co-authored with academics that was critically of AI technology used in Google Search and other Google products.
Some AI researchers inside and outside google criticized a memo from dean, the company’s head AI, that claimed G&M paper didn’t meet the company’s bar for publication. Thousands of googlers and outside AI expert signed a public letter criticizing Google, and the disputed paper were later accepted at a leading Peer Reviewed Conference, without Google affiliation.
The paper was never made public. However, the algorithm was used by Google to create new chips.