Twitter has become a medium for many active members of the ML community. Initially, I think the goal was to have a social network to promote your own work, know about other researchers’ works, engage in an informal discussion with others, and exchange ideas. While I like the general idea, I find several drawbacks to the current state of ML-Twitter.
First, within the community, there is this problem of un-even broadcasting power. Many prominent researchers can promote their work in an insanely fast and universal manner. While this might not seem that bad initially, its effects might be. Assume a prominent ML figure with tens of thousands of followers, sharing a preprint before/during a conference review period. Then, many other prominent figures may praise/like/retweet the paper, and consequently, there’s a high chance this paper will be circulated on Twitter and reach the reviewers. Well, obviously, this will create a bias. The reviewer will review the paper with a biased mind, especially if the reviewer is junior and has less confidence. One might argue that it will be the reviewers’ responsibility not to check Twitter, but I think this is impractical at its best. I will probably dedicate another post to this topic.
The second problem is that some people find themselves becoming influencers of the community. They feel popular, and then this will divide them into 3 groups:
- 1. Those who will not care about being very popular. They stay mostly inactive unless something important has happened.
- 2. Those who want to enjoy their fame and write more and more about their personal opinions and engage more.
- 3. Those who feel they have to use this popularity to push some agenda. The agenda could be purely scientific (e.g., going after deep learning) or other important issues (e.g., ethics, bias, diversity & inclusion, etc.). This is not necessarily bad; the thing is that to bring such a big change, one has to raise the stakes, and sometimes this could be dangerous.
Now let’s focus on the third group. As I mentioned, the stakes are higher for this group, and if we look at the previous controversies, we can see that it involves this group oftentimes.
To name some examples, I would mention Judea Pearl, who is pushing the Causal Inference. Yan LeCun, who is pushing Deep Learning, Gary Marcus, who is pushing for Neuro-Symbolic systems while pushing against Deep Learning, Anima Anandkumar, who is pushing for diversity & inclusion, and Timnit Gebru who is pushing for ethical AI and fairness.
I said that the people in this group would raise the stakes. But, within the group, some raise higher than the others. Also, each of them has their own strategy.
The strategy that Anima Anandkumar and Timnit Gebru employ is to escalate the situation to make it viral. I’m not talking about Gebru being unethically and wrongfully fired by Google. I’m talking about the general Twitter discussion.
Even though I stand by Anima Anandkumar’s cause, I often get shocked by the language she uses. I thought a lot about her behavior to understand why she is doing what she’s doing. I found that people are using the same strategy for the Ethical Treatment of Animals (PETA) in another domain. If you are interested, please read this blog post. The gist of the shared principle is that the more controversial something is, the more will be discussed. In other words, only controversial things get spread.
As a result, Anima & Timnit can be found at the center of many heated discussions on Twitter, which means many people will talk about women in stem, sexism, racism, fairness, and so on. But the attention she brings comes with a cost: most of the attention is negative. I know a close friend who happens to be a very active member of the Women in Machine Learning (WiML) community. A few days ago, she expressed her frustration with Anima’s behavior on Twitter to me. Well, the similarity between her comments and this image came very clear to me:
But there’s a tradeoff: you can get everyone to agree in principle that sexism/racism is bad. The price? no one will pay any attention to it. On the other end, you can get everyone to pay attention to sexism/racism, but a lot of people who would otherwise support you will be so mad at the way you represented them. The price? you’ll lose credibility.
Let’s analyze Anima’s strategy on Twitter:
- First, she would use a very harsh and insulting tone and tag the other end with a bad label (e.g., misogynist, patriarch, white supremacist, bigot, masculine bully).
- Then, she brings the allies to the discussion to spread the word and escalate the situation. Also, bring the intuitions and companies int to the discussion and push them to react.
- She would then do the final strike to ban/cancel/boycott the person in the community.
I can’t emphasize enough that this is not necessarily a bad strategy. Anima has just chosen the other end of the spectrum maybe because she’s convinced that the expected gain for her cause will be higher this way.
There are many cases, LeCun v. Gebru, Anandkumar v. Domingos. Let’s discuss the former. First, I will argue why Domingos is so wrong, and then I’ll come back to Anandkumar’s language and the previously mentioned tradeoff.
1- Broader Impact
It all started with Domingos started attacking the “Broader Impact” section of the NeurIPS conference to the best of my knowledge. Well, I think he is wrong. As an author who is publishing a work that can result in this example, I think you should devote a part of your paper to at least discuss cases like this:
Here the model is obviously biased against black people, and if someone goes to use your model in their system, the bias will be propagated.
However, the discussion went further, and the Gebru v. Google and the change of NeurIPS conference came to discussion.
2- Gebru v Google
Well, Domingos said Gebru was fired from Google. His argument? “Timnit said If you don’t this, I’ll resign. Google didn’t do it, and hence by logic, we can conclude it’s a resign”.
I’m afraid that’s not right. Why? Because she has not resigned. She might have intended to resign, but it won’t be counted as resignation. Someone also referred to a similar case that the court decided that the intention of resignation will not be counted as resignation. You can read the emails on the internet.
3- NeurIPS name change
Finally, the name change. Previously, NeurIPS was NIPS, and the argument for changing the conference name was…well..nips.
Domingos arguing that the name is inconsistent and people have a hard time pronouncing it. I think the time has answered the inconsistency issue. In a short time, the community has moved on, and now NeurIPS is the default name. Regarding the pronouncing, I just don’t get it, honestly!
Let’s go to something Pedro Domingos is not aware of or choose to ignore: inclusion. I myself didn’t know that people in one of the most prestigious conferences in the world can even think about harassing others with the name NIPS. I assumed that well-educated people do not have such behavior/beliefs and don’t belong to academia. Well, I wasn’t aware of the numerous instances of harassment in academia. Now I know. I hope others also try.
What about Anandkumar’s language?
At least to me, there was no question about who’s right and who’s wrong. Pedro Domingos was absolutely wrong in all those 3 aspects.
What I didn’t get, was that why Anandkumar is so aggressive? Honestly, her tone resembles right-wing media.
But as I mentioned earlier, I guess this is intentional, and she chose this strategy since she thinks this is effective. She’s bringing the attention in exchange for her credibility. Let’s review her strategy:
(1) Policing and aggressive strike:
(3) Final Strike: Ban/Boycott
She has now asked for removing Pedro Domingos from the UW directory.
Well, you can guess why Anima is so agressive and angry: trading-off attention for her credibility.
Since yesterday, almost everyone in the ML community is aware of this discussion. I guess the majority of the the community has less respect for Pedro Domingos than yesterday.
But how about Anima Anandkumar?