Today, we’re going to discuss erroneous algorithms in machine learning (this term is used to describe computer-based predictive analytics). To set up the context, I’m going to share a little something I learned back in my political science days as an undergrad. OK, something I learned besides where on campus was the best place to take a nap. (For any Temple University students — SAC, upper level, near the conference rooms. You’re welcome.) There is a central question nearly everyone has asked within the poli sci realm: Is there a perfect form of government? Short answer: No. I will not bore everyone to tears by detailing why not but there is a very simplistic way to understand this dilemma: Government was created by people, people are flawed and, therefore, government will naturally be flawed. Side note: Please do not turn the comments section into a political thunderdome over this proclamation.
This brings me back to the topic of algorithms. Recently, The New York Times published an article that pulled from various studies completed about bias in online marketing ads based on algorithms. An algorithm is a formula. It’s what Google uses when a person types in “best running shoes for beginners” to produce search results. A person creates algorithms, using the principles of predictive analytics, usually some calculus and a dash of black magic. Machines, however, learn from human behavior and adjust algorithms over time. This is known as a learned algorithm.
The Times article gives a great example. When you type into Google or Bing “best running shoes” it auto-completes the thought. But the crux of the article was how search results are being corrupted by the negative, and deeply stereotypical, side of society. For instance, ads targeting applicants for high-paying executive jobs appeared in the search results for men over double the rate as they did for women. A separate study revealed ads for arrest records appearing in searches for African-American-centric names.
People are leaning on machine learning data and calculations because we see this way as the ultimate truth. Machines have no prejudice and will just report the facts. But if they are implanted with bad search algorithms, not necessary created with malice but lack of social understanding, this is like building a house on a cracked foundation.
This sets up the discussion “Oh my God, this is how Skynet started” (this is a reference to the storyline for the “Terminator” series). The machines are learning without us! Artificial intelligence! Before you start building that underground bunker, keep in mind a few things. For starters, data scientists are still trying to understand this phenomenon. It has been suggested if an algorithm shows signs of this behavior to rewrite it. These signs would be present during testing. Ah, yes! That magical thing I suggested a few blogs ago: Always create a test plan.
Test your algorithms. Then test them again. Also, I’m going to drop some additional poli sci knowledge. Niccolo Machiavelli, who wrote The Prince, did not fake his own death. He simply wrote about it. So to everyone who thinks Tupac Shakur faked his own death because he named one of his posthumous albums Machiavelli, this is wrong. But “California Love” is still an awesome song.
This cat is re-creating my most common activity during college.