The experimental tool seemed innocuous enough at first glance: Plug in a pixelated image of a face and it would generate a higher-resolution mock-up using machine learning.But it wasn’t long after an independent programmer posted it to Twitter last week that other researchers started to notice a glaring flaw. When prompted with Barack Obama’s blurry likeness, it returned a white man’s face with little resemblance to the former president. Ditto for other prominent people of color such as U.S. Rep. Alexandria Ocasio-Cortez and actress Lucy Liu.The unfortunate tendency has fueled a heated debate in the artificial intelligence field over racial biases in the industry and how they manifest in the software it produces. Some of the industry’s biggest names have weighed in.It comes as organizations and industries everywhere grapple with issues of discrimination and diversity amid worldwide Black Lives Matter protests against racial injustice.But even before this particular moment of reckoning, a growing movement of activists, academics and technologists have been highlighting the biases and lack of accountability pervasive in the black-box algorithms that have come to govern an increasing share of our daily lives.Their position is largely that incidents like the whitewashing depixelator are a symptom of a larger problem rooted in a lack of diversity in the industry, a failure to consult social science fields for a nuanced understanding of racist societal systems and development processes that are unaccountable to people affected by the finished products.Not everyone sees the issue this way. Earlier this week, Facebook Chief AI Scientist Yann LeCun argued in a Twitter thread that the problems inherent to the model can be traced back to biases in the dataset—in this case, an algorithm called Pulse that is built on a more widely used model from Nvidia called StyleGAN. ML systems are biased when data is biased.
This face upsampling system makes everyone look white because the network was pretrained on FlickFaceHQ, which mainly contains white people pics.
Train the *exact* same system on a dataset from Senegal, and everyone will look African. https://t.co/jKbPyWYu4N— Yann LeCun (@ylecun) June 21, 2020 While LeCun conceded there are other points in the algorithm development process where bias might be introduced, he has pushed the view that datasets are the main culprits in the past, and it is a common rebuttal to people in the AI industry who say that more comprehensive changes are needed.Many other prominent researchers pushed back against LeCun’s argument on Twitter. David Ha, a research scientist at Google Brain, pointed out that even if the dataset were vetted for this particular instance, the benchmarks that engineers use to gauge progress in the field are rooted in their own biased datasets. I respectfully disagree w/ Yann hereAs long as progress is benchmarked on biased data, such biases will also be reflected in the inductive biases of ML systemsAdvancing ML with biased benchmarks and asking engineers to simply “retrain models with unbiased data” is not helpful https://t.co/ecYxivQ7Y9— hardmaru (@hardmaru) June 21, 2020 Timnit Gebru, co-lead of the Ethical AI Team at Google, pointed Lecun to a talk from Google Brain research scientist Emily Denton that breaks down the various ways in which bias can infect specifically facial reconstruction tools, including an examination of what researchers consider to normative and “objective” in how data is framed. Yann, I suggest you watch me and Emily’s tutorial or a number of scholars who are experts in this are. You can’t just reduce harms to dataset bias. For once listen to us people from marginalized communities and what we tell you. If not now during worldwide protests not sure when.— Timnit Gebru (@timnitGebru) June 21, 2020 But a chorus of agreement with LeCun from engineers, developers and anonymous Twitter handles shows these ideas aren’t universally accepted within the AI industry.The reckoning within the field of AI has taken other forms in the past few weeks as well. Amazon, IBM and Microsoft all recently committed to ending or pausing their sales of facial recognition tech after widespread reports over the past few years of such algorithms being racially biased and in some cases leading to wrongful accusations.Tools like the facial depixelator can serve as a reminder of how commonplace bias can be in AI models and why measures like these are only the start of fixing the problem.