Some of us out there, myself included, believe that one day, we’ll live in a world where computers will have a hand in every aspect of life; from driving to creating food and probably making our beds for us. But for now, we’ll have to deal and work on the problems that technology and our internal biases produce:
Google Photos, y'all fucked up. My friend's not a gorilla. pic.twitter.com/SMkMCsNVX4— jackyalciné (@jackyalcine) June 29, 2015
The speed at which this was screenshotted and tweeted represents the anger I felt at the moment. A company with offices in probably every part of the planet, holding nearly 95% of my e-mail traffic and maintainer of my phone and tablet’s operating system and hardware couldn’t be bothered to ensure that their software couldn’t reach a broad audience of people who tend to be heavy technology consumers?
Frustration aside, if we took away the people aspect involved in this complication and focused solely on software; one can say that this was just a misclassification due to pictures being provided that fell outside of the domain of the facial recognition algorithm’s provided models. That statement in itself leads to the bigger problem; which involves more of the wet-ware that was involved in building said models and tweaking said algorithms. I’m not going to repeat what many news outlets have reported on this situation at this point1; Twitter can help you follow the sequence of events. What will be addressed here are the issues that led to this problem and maybe the providing of solutions for free.
Having a Diverse & Inclusive Quality Assurance Team
Having a strong QA (quality assurance) team allows developers and product managers work with comfort knowing that products are working to the expectations that they both can agree on. There’s no question that Google has the such2. Scott presented this point on Twitter:
When software teams have representative diversity, things like this get picked up in QA. https://t.co/YZZHTXQ2WR— Scott Hanselman (@shanselman) July 2, 2015
My thoughts align with this. Like . With Google being a global company, you’d assume that racial and gender diversity wouldn’t be that difficult of a problem for them to solve. That’s an Utopian way of looking at things though given the state of the world of tech when it began and how it’s behaved. Saying it bluntly, it just doesn’t make sense barring the obvious why the world’s most popular search engine was incapable of recognizing the face of a dark skinned Black person.
In a reply to Scott’s tweet, Erica mentions her experience as a beta user for Google Photos:
Of course, a few questions began to come to mind. For starters, what kind of photos were used in the acceptance cycle (or during Erica’s testing experience) that wouldn’t have allowed what I’ve experienced from happening to them? Were lower quality images (like in the range of two mega pixels) used? Did the tests include personal content or were they purely stock level imagery?
Higher Quality === Better Matching?
For those curious, here’s the details on one of the pictures from that collection:
All of the shots were taken on a Nexus 4 using the front-facing camera. The phone is still in my possession after these few years. Tech pundits are free to use this as an excuse as to why the recognition was poor but it’s easily contested. Again, one question remains: did the team have a large enough collection of darker toned people to help circumvent, if not avoid this problem or have a quantity equivalent in size, to photos of other people they have in their collection when they began to build models of what a face looks like? It just seems like the darker the toner of skin, the less likely one would have gotten a match with their facial recognition algorithm.
But it’s not; the faces aren’t invisible, the features not obscured. I’m really convinced that due to a lack of priming from a company that can literally cover my city with printed sheets of paper with its image search index the images of the two of us were classed as the distinct relatives of Grogg. Mind you, other photos with this person where it matched perfectly fine; with the only difference was that we were facing the sun and not putting our backs to it.
For the continued privacy of my friend, I can’t even place images of matches with her; so don’t ask for said images.
So What? We Taking Group Photos?
One can only assume how much of a deal this might be to Google. The chief architect didn’t think it was a good thing. My tweet following is one a LOT of people somehow over looked; at least the first sentence.
@jackyalcine Holy fuck. G+ CA here. No, this is not how you determine someone's target market. This is 100% Not OK.— (((Yonatan Zunger))) (@yonatanzunger) June 29, 2015
Will they lose money over the fact that darker-toned Black faces aren’t that matchable or detectable by their systems? Not necessarily, as far as I can tell. Things of monetary value to companies for profit tend to be prioritized. The whole sensationalizing of this event was partially necessary just because things like this show the implicit, unchecked and ever-present bias that exists in technology. It can be matching a face or matching the level of respect and value of other companies that get funding; the handling of Black people in technology seems unfathomable for these tech giants who are otherwise “too big to fail”.
Someone suggested that I take a bunch of photos and zip them over Google. Supposedly I’d sit in comfort that my few hundred photos out of the potential millions already primed and favored in an algorithm that probably leans towards more on the neural side than on something similar to a Markov chain. What now? That would be a band aid over the more pressing but “invisible” problem: Anyone but the creator of a tool is a second thought.
One solution would be having a team that’s diverse and from intersections broad in scopes of culture, background and experiences so that if a tool or device were to be created; the intended audiences become apparent; not only because the presenter is palatable. Some companies find this difficult. I find their viewing of that as terrifying. Let’s be real; companies hire people who think they won’t give them any problem at all. The less trips to HR, the better! But when you have a company that’s homogeneous in its White structure, it takes a hell of amount of “work” on their behalf to shift culture (if it genuinely exists) and values to respect a more fluid and dynamic environment for people. It’s like taking a tub of vanilla ice cream and trying to make it into a Neapolitan variety without losing any vanilla. Fortunately, we have the option of expanding here; that tub does not.
Another solution very specific to the Google Photos situation would be consulting with a photography agency that specializes in people of different backgrounds and have them provide paid content that’d help in the priming process of their image recognition models whilst giving credit to said agencies to help them gain more of a foot hold in their respectable field. If Google’s willing to do this3, they know my email address and can hit me up about a few people willing to cooperate.
This went from a rant about why machine learning and dark Black faces can be problematic for Silicon Valley’s White technologies to frustration over the infallible homogeneous nature of technologies that are built nowadays that legitimately give no shits about anyone but the ‘stereotypical White person’.