Learning Bias As we begin to apply the massive potential of machine learning to human language, researchers have noticed an unfortunate tendency towards sexist or racist outcomes. Researchers at Princeton University have demonstrated that even as ubiquitous a tool as Google Translate translates gender neutral pronouns to male or female, depending on the content of the sentence. So, the gender neutral statement in Turkish “one is a doctor” is translated as “he is a doctor.” While the gender neutral statement “one is a nurse” is translated as “she is a nurse.” The algorithm learns the biases of human society from our own use of language. The Princeton researchers went a step further and decided to apply a modified version of the Implicit Association Test to the Common Crawl, a vast body of text from the internet that is frequently used to train AI’s. Unsurprisingly, they found the Common Crawl riddled with the kind of everyday bias that is a common but largely unacknowledged part of human communication. Why does this matter? Most of the alarmist rhetoric around machine learning centers on the possibility of a superintelligence taking over the planet as the dominant life form. This research indicates that a far more likely outcome is that machine learning systems built by humans will amplify the logical errors and biases of humans. Biases in Google Translate may be unfortunate, but biases in a machine learning algorithms dedicated to identifying criminal suspects or evaluating mortgage applications are disastrous for the people involved. Because machine learning is a black box and back propagation is not an exact science, rooting out biases once the system has run is not practical. A vast data set like the Common Crawl that is stripped of gender or racial or socio-economic identifiers might be a good place to start. In a nutshell: Sapir-Whorf may or may not apply to humans, but it clearly applies to machine learning. Read More Upside/Downside This week Adidas sent out an email to people who ran the Boston Marathon, the title of which is “Congratulations on surviving the Boston Marathon.” Let me give you a second to think about that. You see the problem? Good. Naturally, the brand quickly apologized for the unintentional awkwardness of their email. And next week, some other brand will be enjoying the type of PR meltdown created by hasty and ill-considered marketing communications. Perhaps this is just the world we live in now. Marketing organizations are pushing out communications via social and digital 24/7. Dedicated war rooms are staffed at all hours to help brands stay on top of every trending story. The Adidas war room (I assume it exists) made a boo-boo which will presumably be “fixed” by adding more layers of approval. Problem solved. Why does this matter? There are many embittered members of Generation X in marketing who are celebrating their slide into middle-age by engaging in ex post facto condemnation of digital and social media. I’m not one of those. But neither am I one of those credulous twits who try to prove their ongoing relevance by collecting Instagram followers. Instead, I assume any marketing communication should be evaluated by its potential upside and downside. The upside of a Pepsi commercial is selling more branded soda and the downside is offending an entire generation with your tone deaf attempt at social relevance. As dumb as it may sound, the upside and downside are at least balanced in that case. The likelihood of the upside and the low incidence of the downside make the decision to run a commercial a reasonable decision. (Despite recent events.) But I cannot understand the upside of the Adidas email. How many shoes and how much athletic gear did they expect that email to sell? I doubt they even paused to consider the question. In all likelihood the email was just part of the “marketing package” they received as part of their sponsorship. The upside, then, is reminding a small group of marathon runners that Adidas exists. The downside is a PR disaster. The same could be said of many of the “timely” communications pumped out by digital and social media war-rooms. In a nutshell: E(x | P(x)) - E(y | P(y)) ≥ 0 Where x is the potential benefit and y is the potential disaster. Read More React Redone Facebook announced this week that they have completely rewritten React, their popular javascript library for building user interfaces. React was something of an unexpected hit in the development community. It became extremely popular both because it offered portability to native-type performance on Android and iOS, but also as an alternative to Google’s Angular for building web apps. Clearly, Facebook sees React as a potential platform for a wide range of their ambitions and wanted to build it again from the ground up to accommodate their future plans. Whenever a popular tool like this is rewritten, there is a danger that applications built under the old version will break. But Facebook assures React developers that the new React will be backwards compatible. Based on their strong API contract, they anticipate "only small breaking changes." Meaning, that because most aspects of React operated independently, the possible compatibility problems will be limited to a few, small things. Why does this matter? I have been critical of Mark Zuckerberg and Facebook in the past – perhaps because Mr. Zuckerberg is extremely wealthy and my own wealth is somewhat less than extreme. But my other objection to Facebook is that Facebook does things to suit Facebook, even at the expense of the company’s fans and users. React was popular and open-source. This was an interesting departure for a company dedicated to monetizing almost everything. React, like most popular open source technologies was adopted and broadly applied, perhaps more broadly than Facebook had intended. Facebook's assumption that the new React will only break a few things is based merely on their own experience. Because they cannot know how other developers have used React, they can’t be sure how many compatibility issues will occur. That’s why trusting large corporations to control open source technologies is a poor idea. In a nutshell: If Facebook can unilaterally rewrite React, it isn’t open source. It’s just free. Read More