This week NBC published an article exploring the source of training data sets for facial recognition. It makes the claim that we ourselves are providing, without consent, the data that may well be used to put us under surveillance.

In January IBM made a database available for research into facial recognition algorithms. The database contains some 1 million face descriptions that can be used as a training set. Called “Diversity in Faces” the stated aim is to reduce bias in current facial recognition abilities. Such bias is rampant often due to too small and too heterogenous (compared to the global population) data sets used in training. That stated goal is ethically sound it seems, but the means used to get there raises a few questions with me. Specifically if the means live up to the same ethical standards that IBM says it seeks to attain with the result of their work. This and the next post explore the origins of the DiF data, my presence in it, and the questions it raises to me.

What did IBM collect in “Diversity in Faces”?
Let’s look at what the data is first. Flickr is a photo sharing site, launched in 2004, that started supporting publishing photos with a Creative Commons license from early on. In 2014 a team led by Bart Thomee at Yahoo, which then owned Flickr, created a database of 100 million photos and videos with any type of Creative Commons license published in previous years on Flickr. This database is available for research purposes and known as the ‘YFCC-100M’ dataset. It does not contain the actual photos or videos per se, but the static metadata for those photos and videos (urls to the image, user id’s, geo locations, descriptions, tags etc.) and the Creative Commons license it was released under. See the video below published at the time:

YFCC100M: The New Data in Multimedia Research from CACM on Vimeo.

IBM used this YFCC-100M data set as a basis, and selected 1 million of the photos in it to build a large collection of human faces. It does not contain the actual photos, but the metadata of that photo, and a large range of some 200 additional attributes describing the faces in those photos, including measurements and skin tones. Where YFC-100M was meant to train more or less any image recognition algorithm, IBM’s derivative subset focuses on faces. IBM describes the dataset in their Terms of Service as:

a list of links (URLs) of Flickr images that are publicly available under certain Creative Commons Licenses (CCLs) and that are listed on the YFCC100M dataset (List of URLs together with coding schemes aimed to provide objective measures of human faces, such as cranio-facial features, as well as subjective annotations, such as human-labeled annotation predictions of age and gender(“Coding Schemes Annotations”). The Coding Schemes Annotations are attached to each URL entry.

My photos are in IBM’s DiF
NBC, in their above mentioned reporting on IBM’s DiF database, provide a little tool to determine if photos you published on Flickr are in the database. I am an intensive user of Flickr since early 2005, and published over 25.000 photos there. A large number of those carry a Creative Commons license, BY-NC-SA, meaning that as long as you attribute me, don’t use an image commercially and share your result under the same license you’re allowed to use my photos. As the YFCC-100M covers the years 2004-2014 and I published images for most of those years, it was likely my photos are in it, and by extension likely my photos are in IBM’s DiF. Using NBC’s tool, based on my user name, it turns out 68 of my photos are in IBM’s DiF data set.

One set of photos that apparently is in IBM’s DiF cover the BlogTalk Reloaded conference in Vienna in 2006. There I made various photos of participants and speakers. The NBC tool I mentioned provides one photo from that set as an example:

Thomas Burg

My face is likely in IBM’s DiF
Although IBM doesn’t allow a public check who is in their database, it is very likely that my face is in it. There is a half-way functional way to explore the YFCC-100M database, and DiF is derived from the YFCC-100M. It is reasonable to assume that faces that can be found in YFCC-100M are to be found in IBM’s DiF. The German university of Kaiserslautern at the time created a browser for the YFCC-100M database. Judging by some tests it is far from complete in the results it shows (for instance if I search for my Flickr user name it shows results that don’t contain the example image above and the total number of results is lower than the number of my photos in IBM’s DiF) Using that same browser to search for my name, and for Flickr user names that are likely to have taken pictures of me during the mentioned BlogTalk conference and other conferences, show that there is indeed a number of pictures of my face in YFCC-100M. Although the limited search in IBM’s DiF possible with NBC’s tool doesn’t return any telling results for those Flickr user names. it is very likely my face is in IBM’s DiF therefore. I do find a number of pictures of friends and peers in IBM’s DiF that way, taken at the same time as pictures of myself.

Photos of me in YFCC-100M

But IBM won’t tell you
IBM is disingenuous when it comes to being transparent about what is in their DiF data. Their TOS allows anyone whose Flickr images have been incorporated to request to be excluded from now on, but only if you can provide the exact URLs of the images you want excluded. That is only possible if you can verify what is in their data, but there is no public way to do so, and only university affiliated researchers can request access to the data by stating their research interest. Requests can be denied. Their TOS says:

3.2.4. Upon request from IBM or from any person who has rights to or is the subject of certain images, Licensee shall delete and cease use of images specified in such request.

Time to explore the questions this raises
Now that the context of this data set is clear, in a next posting we can take a closer look at the practical, legal and ethical questions this raises.

Some things I thought worth reading in the past days

  • A good read on how currently machine learning (ML) merely obfuscates human bias, by moving it to the training data and coding, to arrive at peace of mind from pretend objectivity. Because of claiming that it’s ‘the algorithm deciding’ you make ML a kind of digital alchemy. Introduced some fun terms to me, like fauxtomation, and Potemkin AI: Plausible Disavowal – Why pretend that machines can be creative?
  • These new Google patents show how problematic the current smart home efforts are, including the precursor that are the Alexa and Echo microphones in your house. They are stripping you of agency, not providing it. These particular ones also nudge you to treat your children much the way surveillance capitalism treats you: as a suspect to be watched, relationships denuded of the subtle human capability to trust. Agency only comes from being in full control of your tools. Adding someone else’s tools (here not just Google but your health insurer, your landlord etc) to your home doesn’t make it smart but a self-censorship promoting escape room. A fractal of the panopticon. We need to start designing more technology that is based on distributed use, not on a centralised controller: Google’s New Patents Aim to Make Your Home a Data Mine
  • An excellent article by the NYT about Facebook’s slide to the dark side. When the student dorm room excuse “we didn’t realise, we messed up, but we’ll fix it for the future” defence fails, and you weaponise your own data driven machine against its critics. Thus proving your critics right. Weaponising your own platform isn’t surprising but very sobering and telling. Will it be a tipping point in how the public views FB? Delay, Deny and Deflect: How Facebook’s Leaders Fought Through Crisis
  • Some of these takeaways from the article just mentioned we should keep top of mind when interacting with or talking about Facebook: FB knew very early on about being used to influence the US 2016 election and chose not to act. FB feared backlash from specific user groups and opted to unevenly enforce their terms or service/community guidelines. Cambridge Analytica is not an isolated abuse, but a concrete example of the wider issue. FB weaponised their own platform to oppose criticism: How Facebook Wrestled With Scandal: 6 Key Takeaways From The Times’s Investigation
  • There really is no plausible deniability for FB’s execs on their “in-house fake news shop” : Facebook’s Top Brass Say They Knew Nothing About Definers. Don’t Believe Them. So when you need to admit it, you fall back on the ‘we messed up, we’ll do better going forward’ tactic.
  • As Aral Balkan says, that’s the real issue at hand because “Cambridge Analytica and Facebook have the same business model. If Cambridge Analytica can sway elections and referenda with a relatively small subset of Facebook’s data, imagine what Facebook can and does do with the full set.“: We were warned about Cambridge Analytica. Why didn’t we listen?
  • [update] Apparently all the commotion is causing Zuckerberg to think FB is ‘at war‘, with everyone it seems, which is problematic for a company that has as a mission to open up and connect the world, and which is based on a perception of trust. Also a bunker mentality probably doesn’t bode well for FB’s corporate culture and hence future: Facebook At War.

Peter in his blog pointed to a fascinating posting by Robin Sloan about ‘sentence gradients’. His posting describes how he created a tool that make gradients out of text, much like the color gradients we know. It uses neural networks (neuronal networks we called them when I was at university). Neural networks, in other words machine learning, are used to represent texts as numbers (color gradients can be numbers e.g., on just one dimension. If you keep adding dimensions you can represent things that branch off in multiple directions as numbers too.) Sentences are more complex to represent numerically but if you can then it is possible, just like with colors, to find sentences that are numerically between a starting sentence and an ending sentence. Robin Sloan demonstrates the code for it in his blog (go there and try it!), and it creates fascinating results.

Mostly the results are fascinating I think because our minds are hardwired to determine meaning. So when we see a list of sentences we want, we need, we very much need, to find the intended meaning that turns that list into a text.

I immediately thought of other texts that are sometimes harder to fully grasp, but where you know or assume there must be deeper meaning: poems.

So I took a random poem from one of Elmine’s books, and entered the first and last sentence into the tool to make a sentence gradient.

The result was:

I think it is a marvellous coincidence that the word Ceremony comes up.
The original poem is by Thomas Hardy, and titled ‘Without Ceremony’. (Hardy died in 1928, so the poem is in the public domain and can be shown below)

Without Ceremony

It was your way, my dear,
To vanish without a word
When callers, friends, or kin
Had left, and I hastened in
To rejoin you, as I inferred.

And when you’d a mind to career
Off anywhere – say to town –
Your were all on a sudden gone
Before I had thought thereon
Or noticed your trunks were down.

So, now that you disappear
For ever in that swift style,
Your meaning seems to me
Just as it used to be:
‘Good-bye is not worth wile’