Over the weekend reports of racial/gender bias in Twitter's AI-based image cropping have started blowing up. I wanted to add some context from my perspective as an ex-employee and as a contributor to the research the product is based on. First: the community is right to worry and talk about these issues, and to hold Twitter accountable for ensuring its products work for everyone. If indeed its product fails in a racially or gender-biased way, Twitter should learn from that and fix it, and I'm confident they will.

I enjoyed this podcast with @jack, touching on "THE ALGORITHMS", transparency, political bias and incentives.

https://twitter.com/mikiebarb/status/1291733232015126530

This is not to say I agree with all of Jack's answers and views. Some are good, others underwhelming. The good bits are admission of mistakes (necessary but not sufficient for trust), promise of transparency to address political or algorithmic bias.

After 4 years at Twitter, and more with the Magic Pony team, time has come for me to move on and make it official by tweeting a picture of my laptop and badge, COVID-style, in a box. A ton of stuff happened in these 4 years. We joined weeks before the Brexit referendum, and months before the 2016 US election. Twitter has changed beyond recognition, growing up, embracing our role in serving public discourse.

In my first ever subtweet, let me say this:
If you're known and tweet something controversial that may send the wrong message or cause upset, you shouldn't try to clarify your stance. Amit you were wrong and apologise. Get someone to read your tweets before sending them. Ok, scrap the subtweet, I'm talking of course about Yann LeCun's tweets. As scientists, we are used to arguing and defending viewpoints. Even if there's a technical/philosophical debate to be had about dataset vs model bias, it's is not about that anymore.

https://twitter.com/ylecun/status/1275176322617290752?s=19

Early career (postdoc, PI) researchers working on applying ML in a variety of scientific disciplines: there are some great 5-year fellowships in Cambridge you should take a good look at:
jobs.cam.ac.uk/job/25959/ Hosted at the CS Department (where some great ML momentum is building) and working in collaboration with outstanding scientists across many departments in Cambridge, the scope for impact in these positions is great.

It’s telling of my privilege that I'm only outraged when seeing videos on Twitter of behaviors many people experience first hand - and that my primary concern is what to tweet in reaction so it does not appear empty virtue signaling. So as I should have done earlier, I'm tweeting What this past week exposed about attitudes and behaviours in the US is unbelievable and eye-opening. What we see from police, counter-protestors and politicians is simply barbaric and backwards, so at odds with US’s self-image as a country that leads the world on modern values.

So @autonomousdotai sent me one of their SmartDesk 2.0 to try (cause I'm such an influencer) and they asked me to share my honest opinion about them. Fair deal. (They either haven't read any of my tweets or must be particularly into dry sarcasm and the use of satire in marketing) First impression: it is a heavy beast and took some time to assemble. It has 2 linear motors to do fine adjustment of height. I was very skeptical of their utility but I find myself use this a lot. I share the desk with my wife who's shorter so the controls are actually used.

OK, day 1 of homeschooling.

To see how it's going, you can follow my @vivino profile:
vivino.com/users/ferenc.h… Homeschool maths is actually quite enjoyable (for both of us so far). I figured we can use Lego:

It was 6°C this morning and I saw a class of ~8-10 year olds playing outside in their indoors uniform (i.e. no coats or anything). For girls this is a blouse, thin cardigan, short skirt, optional leggings.

So here's a thread on respiratory infections and exposure to cold👇 Although being in the cold is not what MAKES YOU catch a cold, a bit of literature search reveals evidence that it's likely a contributing factor to increased susceptibility, which may partially explain the seasonality of flu, rhinovirus and coronavirus epidemics.

Convince me to like PR-AuC as a metric for evaluating classifiers. Also, some questions below. Does PR-AuC have interesting interpretations?
(area under the ROC curve is the probability of the classifier assigning a higher score to a random negative example than to a random positive one)

Last day playing barista at ⁦@Meadows_Camb⁩ tomorrow. There’s still time to come by if you missed out. It’s been enormous fun. cc @lawrennd @sebnowozin @liyzhen2 @DaniCMBelg @ryotat
I will give out fee drinks for citations.

We’re excited to reveal a new partnership between Twitter and UC Berkeley:
a new lab, lead by @mrtz and @beenwrekt, dedicated to understanding and improving how ML systems work inside social systems.
blog.twitter.com/en_us/topics/c… We know that there is still la lot for us to learn about the consequences of deploying ML systems to hundreds of millions. This partnership will allow us to create a mix of fundamental and applied research to address these questions and to turn our insights into real-world impact

Here we go again: filter bubble+confirmation bias. All of us who already think it's stupid are flooded with great information and explanation on what WTO terms mean.

We falsely think that all people should surely be getting smarter about this, and therefore opinion would shift.

https://twitter.com/uk_domain_names/status/1073221524545363973

In reality, nobody who does not already think it's stupid will see this thread, or care read it. This applies to my tweet you're reading (you probably agree with my main conclusions already). Same with newspapers or the TV-show below:

https://twitter.com/DavidLammy/status/1088738148438499328

Book Recommendation

Who Can You Trust? by @rachelbotsman
★★★★☆

A very timely book about trust and trustworthiness and especially how this applies to technology, and how technology can enable trust at scale. I gave it four stars for reasons below.👇 The book introduces several concepts and frameworks one can use to think about trust. One of my favourites - and one of my pet peeves - is the distinction between trust (subjective, perception) and trustworthiness (objective, real, often unmeasurable).

Few days ago I tweeted things I should not have. It was bad, I regret and apologize.
This sort of stuff undermines the effort of colleagues, and my own, to articulate the important role various disciplines play in taking ML forward and to create a welcoming and healthy community. I meant to call out a person for a pattern of behaviour. I should not have done that, I did not give enough context, I worded things poorly, the rest is history.

Most painfully, I turned a personal thing into an us-vs-them message which I do not otherwise endorse.