Thursday, 10 December 2020

Google CEO Promises to Investigate Exit of Top AI Researcher

Google CEO Sundar Pichai.

Google CEO Sundar Pichai has apologized to other Google employees for how the company handled the firing/dismissal of AI expert Timnit Gebru. The memo did not state if Pichai or any Google employee had reached out to Dr. Gebru, or whether a direct apology to her would be forthcoming. While the CEO may have intended for his missive to calm the situation, its tone and framing could have the opposite effect.

Pichai’s memo to employees is the latest escalation in what has become a bafflingly strange series of events. Last week, Dr. Gebru, a leading AI ethicist prominently recruited by Google announced she had been fired, though the head of Google AI, Jeff Dean, maintains that she resigned. Dr. Gebru’s immediate manager, Samy Bengio, has issued a statement of support for her.

What’s publicly known is this: Dr. Gebru was fired (or resigned, in Google’s telling) for refusing to withdraw or modify a paper she and other Google employees had written, and for making a series of demands the company found unacceptable. Dr. Gebru and her allies within Google and outside of it have disputed this telling of events. The paper is a survey of previous research on the limitations and weaknesses of AI models currently used for language analysis. (This much, at least, both sides of the dispute agree on).

If that sounds boring to you, you aren’t alone. In fact, that’s one of the strangest things about this entire affair. According to Wired, which has read the document, the most remarkable thing about the paper “is how uncontroversial it is.” The paper does not attack Google or Google technology. It simply cites previous studies showing that AI models can eat huge amounts of electricity and it discusses the problem of building an AI model on biased source material. One of the studies on bias that Dr. Gebru’s paper cites was itself published by Google earlier this year.

The high power consumption of AI is no secret — it’s a major reason why a lot of companies are trying to design new, more-efficient AI accelerators. Google, Facebook, and other leaders in the field have already indicated at various points that they were emphasizing improving the efficiency of their current AI deployments. Similarly, there’s no controversy in the field over the idea that building an AI model on biased data will produce a biased model. The concept of GIGO — Garbage In, Garbage Out — is scarcely new. It’s only been in the past few years that researchers have begun paying attention to the problem, but its existence is non-controversial.

Google has invested a great deal of time and money into the types of large-scale language processing models that Dr. Gebru’s paper critiqued, and deployed its own language model, BERT, to aid in processing long search results. But the paper in question wasn’t specifically critical of BERT.

The company has claimed that it accepted Dr. Gebru’s resignation because she demanded to know who, specifically, had deemed that her paper did not meet Google’s required standards for publication. What has not been explained to anyone’s satisfaction is why her paper was deemed unsatisfactory in the first place. Google said that her paper ignored more recent research in the field that shows newer models as more power-efficient than those in the past, as well as work discussing recent work on bias. As MIT Technology Review notes, however, there are 128 citations listed in the paper, which was a collaboration between Dr. Gebru and six other authors, including four Google researchers.

Another explanation for the rejection is that her submission violated Google’s two-week requirement for how and when papers must be submitted prior to being approved for publication. This explanation has been publicly disputed by previous Google employees.

The criticism of how this situation played out isn’t just coming from external parties. Several thousand Google employees have signed an open letter demanding a fully transparent investigation into the circumstances surrounding what they frame as Dr. Gebru’s firing.

Google’s public argument, broadly stated, is that it accepted Dr. Gebru’s resignation after she made inappropriate demands for transparency. It also alleges that her paper was submitted improperly and in a way that was a meaningful violation of Google’s rules. A large number of people, including individuals that worked directly with Dr. Gebru, have directly challenged these assertions.

Pichai’s memo is obviously intended to respond to this problem. It does not do so very well.

How to Memo, But Badly

Axios has the full document, but I’ll hit a few high points. Anything italicized is from Pichai’s memo. Here’s the opening paragraph:

One of the things I’ve been most proud of this year is how Googlers from across the company came together to address our racial equity commitments. It’s hard, important work, and while we’re steadfast in our commitment to do better, we have a lot to learn and improve. An important piece of this is learning from our experiences like the departure of Dr. Timnit Gebru.

A phrase like “learning from our experiences like the departure of Dr. Timnit Gebru” is passive-voice PR-speak that pins all of the action on Dr. Gebru and implies she left the company rather than Google terminating her. The question of whether Dr. Gebru resigned or was terminated by Google is one of the central points of dispute. One of the two open letters published by Google employees in support of Dr. Gebru opens by stating: “Dr. Gebru did not resign” (italics original). Pichai’s decision to frame his opening paragraph as though this is a settled point both dodges responsibility and ignores the contested nature of the claim.

The first thing Google needs to do, Pichai writes, is:

assess the circumstances that led to Dr. Gebru’s departure, examining where we could have improved and led a more respectful process The second thing it needs to do is accept responsibility for the fact that a prominent Black female leader with immense talent left Google unhappily.

The first is another example of awkward, bland PR-speak intended to minimize the idea that Google took action. The second sounds good — Google is taking responsibility for something — except, again, it treats the entire issue of Dr. Gebru’s departure as settled.

The three demands of the open letter signed by several thousand Google employees were for Google to explain why the paper was rejected, explain why it wanted Dr. Gebru and her colleagues to withdraw their language model research, and for Google to “make an unequivocal commitment to research integrity and academic freedom.” Pichai’s memo makes numerous mentions of how the furor surrounding Dr. Gebru’s dismissal has raised unwelcome fears for other persons of color at the company, but it does nothing to address the substantive questions Google employees have asked surrounding the circumstances of her departure.

Pichai claims that Google will feel her loss because Dr. Gebru is an expert in areas that Google “must” make progress on, and that this progress depends on our ability to ask ourselves challenging questions. He writes this, apparently, without a trace of irony.

I’m not going to claim to have insider knowledge of what happened between Dr. Gebru and Google, but I can’t remember the last time a bunch of engineers stood up and publicly protested the dismissal/firing of a single individual. It’s clear that this situation looks fishy to an awful lot of people, and Pichai’s declaration that Google will take the right lessons from these events is hard to believe given that this memo ignores every substantive question that’s been raised about the situation. Google may believe it acted defensibly and in complete accordance with its internal policies. But a substantial group of people, some of whom were involved with the paper in question, do not agree.

In his memo, Pichai pledges that it is incredibly important to him that our Black, women, and underrepresented Googlers know that we value you and you do belong at Google. One of the best ways to demonstrate that this is true, under the circumstances, might be to directly engage with the requests for transparency regarding Dr. Gebru’s termination.

Just a thought.

Now Read:



Posted by at

No comments:

Post a Comment

Subscribe to: Post Comments (Atom)