Algorithm Bias in AI

In the present era with the advancement and evolution of Artificial Intelligence, Algorithmic Bias has become a matter of great concern and an important research area that has grown a great awareness among the general audience over the past few years. For all those who are very new in this field and are not aware of the “Algorithmic Bias” term, Let me give you an example. Suppose you want to buy a shoe and you are made to think of the best shoe available in the market.

As your favorite color is red so you thought of a shoe with red color patterns on it. Now if I ask someone else to think of his/her desired shoe. Then he/she will think of a shoe having a pattern or design of his/her preferred color (Let it be blue). This is what is called Bias which generally happens during the training of a machine learning model. We humans are very much biased towards our personal choice.

An ML model works by recognizing the patterns and predicting the result from a dataset usually provided by a user. As the dataset is collected by a human so we can’t ignore the fact that data provided will be too uneven and will not be completely neutral as each individual is biased to some or other thing. Even with the utmost care, we can’t completely ignore and separate the human bias while accumulating data due to which this bias also becomes an integral part of the ML technology we create.

You must have mostly seen people debating on the point that computer systems are racist and they discriminate against people. But that’s not the actual case, just like humans behave and make assumptions based on their upbringing, training and their personal experience, same goes with the computers too. It is not created by God or it doesn’t exist in a void. It tends to think in the same way it has been trained and taught from the given training data. We need to take a step back and have to take a closer look at the training data that we are providing and filter it in such a way that the predicted result is fully neutral without any bias.

Even though there are many methods but I believe there are two main methods in which the training data is causing algorithm bias within the Artificial Intelligence:-

1. Bias caused by individuals while collecting the data.
2. Unintentional natural bias caused during the data collection process

Due to all these biases caused during the accumulation of data, the data gets more centric towards a particular category and becomes a representer of only some particular group of people ultimately leading to degrading the model’s performance.

For example, suppose a university in the United States developed a facial recognition system that could classify males and females. The dataset used for training the system was mainly collected from the CCTV footage of the busiest market in the united states.

Now the issue here was that US has a great ratio of Whites as compared to the Blacks So the model became too centric towards the whites and learned their’s each and every one feature for classification. But when this system was made globally available and was used by a university in Africa for testing the accuracy then they got very less precision.

Do You Know Why ????

The degradation of the model’s performance was due to the poor recognition of dark-skinned people in Africa. As the training data used for model training was not at all cleaned during pre-processing due to which it was not much diverse and was much centered towards white-skinned people so the system could not correctly read the features of the input images provided for testing and randomly predicted the result due to which the precision and recall score was reduced ultimately affecting the accuracy of the model.

Now coming to the point where this bias really affected a large scale of consumers,

1. Amazon once started a project in 2014 whose main motive was to review the applicant’s resumes and rate those applicants by using various algorithms. They built somewhat around 500 models and defined 50000 key terms which would be picked up. But a year later in 2015, they found out that the system was not rating the female candidates and was completely biased towards the male population. This was due to the reason that Amazon used the last 10 years' historical data for the training purpose and these datas were very much male dominant and the male to female ratio was too high. The model got trained in a similar way and penalized to all those resumes which contained key terms like Woman, Female, Her, etc ultimately affecting a large no of female candidates who had applied for the job.

2. An US organization developed a health care risk predictor for determining which patients had a high risk of disease and required an intense care management program. The thought process behind developing this kind of program was to treat the sickest patient with additional attention and provide them with all kinds of resources and medical support. But while developing the algorithm, the developers used previous patients’ medical reports as training data, and there the income and race had a high co-relation. Due to which instead of being trained to predict the sickest patient, the algorithm ended up finding the sickest who can spend the most money for treatment which ultimately created a race bias among the whites and the blacks and affected millions of people.

3. An audit conducted by a few researchers in California found out that the Advertisement system of some social media platforms like Facebook had great discrimination against women. These researchers discovered some ads for some job lists and all these list datas required similar kinds of requirements but for all different agencies and companies. But Facebook was targeting males for all male-related jobs like Software Engineer at Nvidia but displayed all kinds of teaching, Nursing, Salesman kind of jobs to the females like Sales associate at Lalchand Jewellers. This resulted in a great gender bias against women and many cases were registered against these social media platforms. Finally, these platforms settled the lawsuit by dropping the ad targeting option for all the jobs postings.

Can we remove the biases from AI completely ???

Frankly speaking, According to me technically we can remove the biases from the AI system but it’s very difficult to do so. The effectiveness and accuracy of a machine learning model will be as good as the quality of the dataset provided to it during training. The removal of algorithmic biases completely depends on us. If we can clean our training dataset and can eradicate the uneven datas having unconscious assumptions and stereotypes on race, age, completely then we reduce the bias to a great extent. But there exists a large no of human bias and this number is constantly increasing day by day so just like it’s nearly impossible to have a bias-free human mind so is the case with the AI system.

The only effort we can apply is by performing some kind of optimization to reduce the AI bias.

Considering a very layman attempt, we can try to reduce the prejudices of humans from the training data. But it’s not an easy task as it sounds. Another approach includes removing all kinds of classes and labels like sex, race, color, etc which contributes to algorithm bias but the issue in it is it will remove a large set of samples from the dataset and for a large scale model training, we can’t provide incomplete dataset as it will affect the model’s score and reduce its performance.

There are no shortcuts for removing bias but there are great recommendations provided by various consultancy agencies like Mckinsey which focuses largely on AI bias reduction. Mckinsey and company highlighted 6 potential methods to developers and AI practitioners for making AI reach its max capability by reducing bias and increasing people’s trust in this system.

1. We should be completely aware and understand the algorithm completely in which the Machine Learning model can help in correcting the bias and those where there is a high risk of unfairness.
2. We should develop 3 most important strategies -:

• Technical strategy that helps us to find out the source of bias and what features of it is affecting the accuracy
• Operational strategy helps us in improving the quality of data collected.
• Organisation strategy involves creating a workplace that will help in calculating the metrics score and keep it transparent for all.

3. Once we identify the biases we have in the system then we should consider all the human interfered processes and try to modify and improve them.

4. We should fully explore and learn all the possible use cases where automatic decisions will be preferred and where human interference will be required.

5. We need to research thoroughly and have to invest more in data collection and bias research and have to follow a multidisciplinary principle that will include social scientists and experts who can understand the fault and can develop a technique to solve it.

6. We should invest heavily in diversifying the field as it will help us in neutralizing our dataset and the model will not target a specific group on the basis of race, age, etc. So we can remove unwanted algorithm bias by maintaining a diverse team.

There are various online tools and libraries which are available which we can integrate into our system and can detect bias.

★ Recently IBM released an open-source library named AI FAIRNESS 360 for detecting and mitigating biases in unsupervised learning algorithms which have nearly 40 contributors on Github. It helps the AI practitioners to detect and test the partialities in the model and training data with a set of metrics scores. It also helps in reducing bias with the help of 12 algorithms like Learning Fairness representation, Rejecting option classification.

★ IBM’s cloud data and AI team announced a premium AI product named IBM Watson OpenScale which won the AI excellence award. It performs bias checking and helps in removing it in real-time when AI is making its decisions

★ Google has also developed its own tool named What-if that can test the model’s score in hypothetical situations and can visualize its performance across multiple subsets of training data and for different fairness metrics of machine learning.