Can AI be Racial? If humans can, why can't be robots!!!

If there is one buzz word that keeps reverberating every now and then, it has to be AI (Artificial Intelligence). With the imminence of AI there has been several discussions going on whether AI would replace humans and all the jobs would be taken over by robots or not. Though this question is yet to be answered there is one human trait AI is adopting pretty fast i.e. the attribute of "being bias". So whether AI replaces human or not, but the human traits are here to stay with the Robots.

In 2016 diversity.ai did organize an online beauty contest where the jury were robots. This contest was driven by artificial intelligence and this idea was supposed to remove the bias, preference of individuals and comparison on non level playing field.

This was driven by deep learning which allows for the rapid and automatic assessment of diversity and removes the possible discrimination by race, sex, age and other parameters. Whether it's supervised or unsupervised learning the AI was supposed to get the best out of all the sample set provided. As it happens in a jury comprising of humans, there were several experts among the robots as well evaluating specific attributes while assessing beauty. But as it happens with humans, there were bias and prejudice that was predominant in determining the beauty as a function of fairness, symmetry and past experience.

AI is primarily driven by the data set it refers to and hence if the data set is not diverse the result is going to be weird. Hence in case of Beauty.AI when certain population groups were under-represented in the training sets, these populations are left out or were subjected to higher error rates. Hence the AI was able to declare the white men and women more beautiful compared to their black counterparts. This is true in face recognition as well, where the AI was not able to detect the face of the black men or women i.e. the success rate is around 60-70% where as it is approx. around 96% for white men and women. It is always difficult to distinguish between legitimate correlations (i.e. man is to king and as woman is to queen) and biased correlations (i.e. man is to doctor as woman is to nurse). So it is really difficult to get the data sets which represents such diverse perspective.

So how does AI work and how can we remove the bias from the system? There has been numerous debate going on this now and increasing diversity in teams is getting traction because of this bias problem as well.

Using a dataset with photos of black dogs and white cats when used extensively, results in the software incorrectly labeling a white dog as a cat. In this case even though the AI was wrong, it was very sure it was right with a confidence of 96%, making it harder to detect the error. Researchers are looking for places where the software had high confidence in its decision, finding mistakes and noting the features that characterize the error.

AI systems are black boxes and hence the input data goes in and the answer comes out without an explanation for the decision. The neural network logic works at the back end but does not disclose how the answer was finalized. It is difficult to figure out how bias creeps in the system. Hence there is no way for an individual to challenge the outcome with logic and hence invariably this just has to be accepted.

So how to get rid of this Bias? It's a daunting task, but some steps can be taken for the time being to address this and some of the great minds in the field of AI are working on this:

• Establish a discussion forum for thought leaders and innovators in AI to discuss on racial, gender and other bias that cripples the effectiveness of AI.
• Develop a range of guidelines and validation mechanisms to test the deep learning systems and other cognitive computing solutions for racial, gender, age and ethnic bias;
• Develop an open access data sets to allow developers through out the world to train the algorithms on minority data sets.

So whether AI would completely replace humans and be a different species or replace a generation of humans to be a new generation with the same bias as the existing human race would be interesting to watch. Whether it would happen in a decade or a century time will only tell.