Addressing different types of bias in AI

AI has evolved into a major technical field with important commercial applications since it originally captivated the imagination of computer scientists and mathematicians in the 1950s. The technology required for large-scale surveillance are fast maturing, with image classification, face recognition, video analysis, and speech recognition systems all making substantial development by 2021. Can we trust the judgment of AI systems? However, as exciting as these new machine learning capabilities are, there are important factors to consider when designing, executing, and deploying machine learning models. Bias is one of the most significant difficulties in in machine learning. In this paper, we will review the basic information of bias, types of bias and how to address the bias in AI.

The term “bias” has a wide range of meanings. In general, Data ethicists referring to the concept of bias in social science. AI bias is an anomaly in the output of machine learning algorithms. These could be due to the prejudiced assumptions made during the algorithm development process or prejudices in the training data. One might think why we should care too much about bias. Well, The commercial application with biased AI can have consequences of moral burden and in some cases it can lead to pay very directly for their actions. The small amounts of bias can rapidly increase exponentially because of feedback loops in machine learning. In addition, Human bias can lead to larger amounts of machine learning bias. Also, Human decision makers and algorithmic decision makers are not used in a plug and-play interchangeable way in practice. As a result, we will review the different types of bias and how to address them.

Harini Suresh and John Guttag of MIT describe six categories of bias in machine learning in their study “A Framework for Understanding Unintended Consequences of Machine Learning.” These six categories of bias include Historical Bias, Measurement Bias, Aggregation Bias, Representation bias, Evaluation Bias and Deployment bias.

Historical bias: People are prejudiced, processes are biased, and society is biased, all of which contribute to historical bias. Suresh and Guttag mentioned that “Historical bias is a basic, structural issue with the initial step of the data generation process,”, “and can occur even when flawless sampling and feature selection are used.” For instance, here are a few examples of historical race bias in the US, 1) Doctors were substantially less likely to propose cardiac catheterization (a useful procedure) to Black patients when they were presented identical files. 2)When negotiating for a secondhand car, Black persons were offered $700 higher start prices and considerably less compromises.3) While Responding to Craigslist apartment rental advertising using a Black name received less responses than responding with a white name.

Aggregation bias: Aggregation bias occurs when models fail to aggregate data in a fashion that includes all of the relevant components, or when a model lacks the requisite interaction terms, nonlinearities, and so on. This is especially common in medical contexts. For example, the treatment of diabetes is frequently based on univariate statistics and research involving small groups of heterogeneous people.

Representation bias: Maria et el observed that models for forecasting occupation not only reflected, but magnified, the gender imbalance in the underlying population! This form of representation bias occurs frequently, especially in simple models. When there is a clear, obvious underlying link, a basic model will frequently assume that it holds all of the time.

So now we understand the different kinds of Bias in AI , how we can address them? Various types of Bias require different mitigating strategies. Author have described the three approach to address bias. One of the solution is to have inclusive and diverse dataset. While a more diverse inclusive dataset can help with representation bias, it can’t help with historical or measurement bias. Bias exists in all datasets. Many experts in the field have agreed on a series of recommendations to enable better documentation of the decisions, context, and specifics about how and why a particular dataset was created, what scenarios it is appropriate to use in, and what the limitations are. This way, those using a particular dataset will not be caught off guard by its biases and limitations.

Andrew et al suggested to have robust auditing process. Many AI teams are putting their systems through auditing processes so that we can constantly monitor what forms of bias, if any, these AI systems are exhibiting, so that we can at the very least recognize the problem if it exists and take actions to correct it. Many facial recognition teams, for example, are systematically testing their system’s accuracy on different subsets of the population to see if it is more or less accurate on dark-skinned versus light-skinned people. So, having transparent systems and regular auditing methods increases the chances of detecting an issue promptly, if one exists, so that we can address it.

A diverse workforce will also help in the reduction of bias. Individuals in your workforce who are more diverse are more likely to notice different problems, and they may even help make your data more diverse and inclusive in the first place if you have a diverse workforce. I believe that by having more different points of view while developing AI systems, we may all expect to produce less biased applications.

References

1. Andrew Ng (2021) AI for Everyone, Available at: https://www.deeplearning.ai/program/ai-for-everyone/ (Accessed: 27 April 2021).
2. Daniel Zhang et al (2021)“The AI Index 2021 Annual Report,” AI Index Steering Committee, Human-Centered AI Institute, Stanford University, Stanford, CA, March 2021.
3. Howard, J. and Gugger, S., (2020). Deep Learning for Coders with Fastai and PyTorch: AI Applications Without a PhD. 1st ed. Canada: O’Reilly Media, Inc., 1005 Gravenstein Highway North, Sebastopol, CA 95472.
4. Maria et al. (2019) “Bias in Bios.” Proceedings of the Conference on Fairness, Accountability, and Transparency (2019): n. pag.

About Author: Pinkesh Patel

Pinkesh have Over 16 years of experience in R&D, portfolio management, and business development in life science & retail Industry. He is mentor and investor at gold and diamond Jewelry firm ‘Proyasha Diamonds’. Pinkesh has Received B.A. Honors in Pharmacology from London Metropolitan university and MBA from Anglia Ruskin University.