In the following video, I'll take a step back from GANs and discuss bias. An issue that's seen around the world and seeps into many aspects of life to which machine learning and GANs are no exception. The purpose of these videos is to bring awareness to bias in machine learning and more specifically in GANs, which is the first step towards eliminating bias in your models. You get a brief introduction to the machine bias article in the journal ProPublica, and a discussion of how racial disparity was found in proprietary software that's used across the country in criminal sentencing and the impact that this has. First off, I'll go over the main points from the ProPublica news article called Machine Bias. In the United States criminal justice system, courts are increasingly relying on risk assessments or the likelihood you're going to commit a crime in the future. Many risk score calculations are now computerized and they're starting to rely on machine learning more and more. ProPublica used a public records request to assess one of the two leading commercial models for it's machine bias series, and this algorithm is called COMPAS. The purpose of COMPAS was to make pretrial sentencing less biased, so its purpose was to reduce bias, but ProPublica's findings suggest that the algorithm itself is actually significantly biased. One important thing to note about COMPAS is that the calculations used for the scores are considered proprietary and the company doesn't share exactly what they are. The calculations are broadly based on a questionnaire filled out by defendants, as well as on their criminal records. There's no question about race there, but it asks questions like, was one of your parents ever sent to jail, and is it wrong for a hungry person to steal? So getting a little bit philosophical and maybe moral. Then judges can then increase or decrease that sentence length based on their evaluation of the risk score, and so yes, these questions do not question race at all, but we'll see how some of these proxy questions might be trying to get at that or might be biased due to the type of question asked. Unfortunately, the COMPAS algorithm only predicted recurrence of violent crime correctly about 20 percent of the time. ProPublica found that models predicted black defendants were 77 percent more likely and more at risk for committing a future violent crime and 45 percent more likely for any kind of future crime. You can see in the graph that while the count of white defendants in high risk scores decreased together, this graph for black defendants shows that the likelihood remains the same. This starts to suggest that the scores for white defendants were skewed towards these lower risk categories while the scores for black defendants were not. As one example, Gregory Lugo , who is a white man, versus Mallory Williams, who is a black woman, were both arrested for DUIs, but had different profiles on their prior offenses. Lugo was actually on his fourth serious event with three DUIs prior to this as well as a battery, and Williams was actually on her first serious offense with only two misdemeanors beforehand, which is less than a DUI. Even though Williams's history, when it comes to criminal behavior, was much less to fear then Lugo's, her risk assessment score was actually much higher, a six as opposed to a one here. What kind of effects can these risk assessment scores have on individuals? Paul Zilly, who's a recovering addict that relapsed and stole a lawnmower, had his plea deal overturned, which was agreed upon from both parties, the prosecutor and the defendant. His sentence increased from a year in county jail to two years in state prison after the judge saw his high risk score. But after an appeal, his sentence was then reduced from two years to 18 months, and the appeals judge had stated that without COMPAS, Zilly would have been given one year or six months instead, meaning that this risk score actually made it so that he would go to jail for much longer. On the other hand, Sade Jones, who was 18 at the time of her arrest with no previous record, had been given a medium risk score after taking an unlocked bicycle and her bond was raised from the recommended $0 to a $1000. Still now, she struggles to find a job due to the misdemeanor on her record. These are some of the consequences of what a higher risk score does, even though the person may not have necessarily done bad things afterwards, and it's largely just based on what the model has predicted. If the model has predicted in a skewed way, in a biased way, then that might not be very fair. The researchers also found that the model correctly predicted repeat offenders only 61 percent of the time across all areas of risk of recidivism, risk of violent crime, and risk of failure to appear. The accuracy was only 20 percent for violent crime alone, however. The issue starts to arise because the errors were different for those who are white versus those who are black. For the false positives, almost 45 percent of African-Americans, who were labeled to be at high risk of committing a future offense, didn't actually commit one, and that's compared to 23.5 percent of white Americans. Then for the false negatives, i.e, those who were labeled low risk but actually did re-offend, so the ones we should've caught, 28 percent of African-Americans and 48 percent of white Americans. So it's noticeable something is broken with the fairness of this model in the evaluation method once these details were uncovered. In summary, models that were created to reduce bias in the criminal justice system by predicting who is and who isn't likely to re-offend, actually seem to only reflect our own biases back at us and reinforced the difficulty of re-entry for populations that are historically under-served and discriminated against. Proprietary risk assessment software created by for-profit companies only make understanding and improving these systems more difficult because their calculations aren't disclosed, and often any investigations into their accuracy are done by the same people who built the software actually. Something else to consider is that these models don't view people holistically, like people's sincere efforts to rehabilitate themselves or the effect a longer sentence could have on their ability to take care of their families. So a really important question to ask yourself is, who are these models really serving today?
