AI bias will explode. But only the unbiased AI will survive.
Within five years, the number of biased AI systems and algorithms will increase. But we will deal with them accordingly – coming up with new solutions to control bias in AI and champion AI systems free of it.
Many AI systems are trained using biased data
AI systems are only as good as the data we put into them. Bad data can contain implicit racial, gender, or ideological biases. Many AI systems will continue to be trained using bad data, making this an ongoing problem. But we believe that bias can be tamed and that the AI systems that will tackle bias will be the most successful.
A crucial principle, for both humans and machines, is to avoid bias and therefore prevent discrimination. Bias in AI system mainly occurs in the data or in the algorithmic model. As we work to develop AI systems we can trust, it’s critical to develop and train these systems with data that is unbiased and to develop algorithms that can be easily explained.
IBM researchers developed a methodology to reduce the bias that may be present in a training dataset, such that any AI algorithm that later learns from that dataset will perpetuate as little inequity as possible.
As far as I can tell, what IBM means here is that they replace some of the politically incorrect real data with some politically corrected faked data. IBM links to this post:
Reducing discrimination in AI with new methodology
December 6, 2017 | Written by: Kush R. Varshney
Principal Research Staff Member and Manager, IBM Research
I finally had a chance to watch Hidden Figures on my long journey to Sydney, where I co-organized the second annual ICML Workshop on Human Interpretability (WHI). The film poignantly illustrates how discriminating by race and gender to limit access to employment and education is suboptimal for a society that wants to achieve greatness. Some of my work published earlier this year (co-authored with L. R. Varshney)
No discrimination involved in K.R. Varshney and L.R. Varshney working together.
explains such discrimination by human decision makers as a consequence of bounded rationality and segregated environments; today, however, the bias, discrimination, and unfairness present in algorithmic decision making in the field of AI is arguably of even greater concern than discrimination by people.
AI algorithms are increasingly used to make consequential decisions in applications such as medicine, employment, criminal justice, and loan approval. The algorithms recapitulate biases contained in the data on which they are trained. Training datasets may contain historical traces of intentional systemic discrimination, biased decisions due to unjust differences in human capital among groups and unintentional discrimination
White people making more money than black people is an example of unjust differences in human capital. In contrast, the Varshney family working together is due to their high human capital, which is a just difference.
, or they may be sampled from populations that do not represent everyone.
My group at IBM Research has developed a methodology to reduce the discrimination already present in a training dataset so that any AI algorithm that later learns from it will perpetuate as little inequity as possible. This work by two Science for Social Good postdocs, Flavio Calmon (now on the faculty at Harvard University) and Bhanu Vinzamuri, two research staff members, Dennis Wei and Karthikeyan Natesan Ramamurthy, and me will be presented at NIPS 2017 in the paper “Optimized Pre-Processing for Discrimination Prevention.”
The starting point for our approach is a dataset about people in which one or more of the attributes, such as race or gender, have been identified as protected. We transform the probability distribution of the input dataset into an output probability distribution subject to three objectives and constraints:
Group discrimination control,
Individual distortion control, and
By group discrimination control, we mean that, on average, a person will have a similar chance at receiving a favorable decision irrespective of membership in the protected or unprotected group. By individual distortion control, we mean that every combination of features undergoes only a small change during the transformation to prevent, for example, people with similar attributes from being compared, causing their anticipated outcome to change. Finally, by utility preservation, we mean that the input probability distribution and output probability distribution are statistically similar so that the AI algorithm can still learn what it is supposed to learn.
Given our collective expertise in information theory, statistical signal processing, and statistical learning, we take a very general and flexible optimization approach for achieving these objectives and constraints. All three are mathematically encoded with the user’s choice of distances or divergences between the appropriate probability distributions or samples. Our method is more general than previous work on pre-processing approaches for controlling discrimination, includes individual distortion control, and can deal with multiple protected attributes.
We applied our method to two datasets: the ProPublica COMPAS prison recidivism dataset (an example containing a large amount of racial discrimination whose response variable is criminal re-offense) and the UCI Adult dataset based on the United States Census (a common dataset used by machine learning practitioners for testing purposes whose response variable is income). With both datasets, we are able to largely reduce the group discrimination without major reduction in the accuracy of classifiers such as logistic regression and random forests trained on the transformed data.
On the ProPublica dataset with race and gender as protected attributes, the transformation tends to reduce the recidivism rate for young African-American males more than any other group.
In other words, in the real data, young African-American offenders tend to re-offend at a higher average rate than other groups. In IBM’s fake data, however, not so much.
On the Adult dataset, the transformation tends to increase the number of classifications as high income for two groups: well-educated older women and younger women with eight years of education.
Do younger women with eight years of education tend to earn a lot of undeclared dollar bills?
… These situations call for an end-to-end auditable system that automatically ensures fairness policies as we lay out in this vision (co-authored with S. Shaikh, H. Vishwakarma, S. Mehta, D. Wei, and K. N. Ramamurthy); the optimized pre-processing I’ve described here is only one component of the larger system.
Of course, the names S. Shaikh, H. Vishwakarma, S. Mehta, D. Wei, and K. N. Ramamurthy are just in the real data set of names of artificial intelligence experts. In IBM’s fake data set, though, the names of the artificial intelligence experts are D. Washington, P. Garcia, P. Badeaux, L. Jones, and T. Glasper.
By the way, the headline is a reference to this classic 1970s joke about IBM salesmen:
Three women were talking about their husbands and their love making, and the first one says, “My husband is an athlete and when he makes love to me, he is so powerful that I am swept up in his body, and it’s wonderful”
The second woman says, “My husband is a violinist, and when we make love, he knows how to play me like I’m a fine musical instrument, and it’s overwhelming and beautiful”
The third woman says, “My husband is a salesman for IBM, and he doesn’t actually make love to me, he just sits on the edge of the bed and tells me how good it’s going to be when I finally get it.”