[ML_Fairness]Fairness in Machine Learning

Nowadays, everyone sees that Deep Learning shows state-of-the-art performance and making human life will be very much improved through the technologies. As the importance of good algorithms, the quantity of data is important to companies using it.

© 이 우철

The quantity of data really cares because it enhances the performance. The performance connects to the profit of company so that the gap seems bigger bigger today.

But, how about the other side of data? Is data really help predict unknowns? Isn’t it biased?

Also, how about the algorithm? If the prediction of model decides who to hire and who many of imprisonment years to the murderer, is it totally free from false negative cases? False negative cases will be ‘murder acquitted in the court’ or ‘company fails to offer the job to applicant who is really competent’.

© 이 우철

In machine learning, also including deep learning, input data is trained. Trained data which finished learning followed by optimization, validation is set solid as model when it is inferred. When new input data is procured, output(whether it is by ‘classification’ or ‘regression’) comes out. Therefore, simply speaking, model is ‘inductive reasoning algorithm’ which makes broad generalizations from specific observations, from specific to general.

Therefore, to make a good model, to utilize it well, observations, which is input data in the deep learning framework, should not be biased. If input data is biased, model will be biased and the consequence of it will be also unfair and polluted.

© 이 우철

Also, the weakest point of deep learning is, it is literally ‘deep’ so that it is hard to tell what is going on deep in the model. If model shows bias, will human society admit it without any doubt?

The problem of ‘black box of ML Models’ will make human reluctant to use it and believe it unless it makes perfect/ 100% fair output.

Here, FAT in ML, Fairness, Accountability and Transparency in Machine Learning, is becoming more and more important.

Fairness of AI Systems

According to Microsoft Research, fairness of AI systems is defined as follows:

• AI systems can behave unfairly for a variety of reasons.
  • societal biases reflected in the training data and in the decisions made during the development and deployment of these systems.
  • characteristics of the data (e.g., too few data points about some group of people)
  • characteristics of the systems themselves
• “Unfairness of AI” can be hard to distinguish between these reasons, especially since they are not mutually exclusive and often exacerbate one another.
• Therefore, we define whether an AI system is behaving unfairly in terms of its impact on people — i.e., in terms of harms — and not in terms of specific causes, such as societal biases, or in terms of intent, such as prejudice.

Type of Harms(Unfairness of AI)

• There are many types of harms (see, e.g., the keynote by K. Crawford at NeurIPS 2017).

  • Allocation harms : It can occur when AI systems extend or withhold opportunities, resources, or information. Some of the key applications are in hiring, school admissions, and lending.

  • Quality-of-service harms : It can occur when a system does not work as well for one person as it does for another, even if no opportunities, resources, or information are extended or withheld. Examples include varying accuracy in face recognition, document search, or product recommendation.

Potential Area

• AI in Finance
  • Banking
  • Mortgage Rates
  • Credit Ratings
• AI in the Court
  • Judge
  • Attorney
• AI in Job Market
  • Interviewer
  • CV Screener

Conclusion

With exponential development of maching learning algorithm and the magnititude of data that can feed into learning, ML is approaching to human life with faster pace than ever before

To aid in the decision-making process, data used in learning, visibility, transparency learning algorithms and fairness/interpretability of the outcome is crucial.

Deeper and more various sighted research in Fairness ML will help people to cognitively feel comfortable when accepting AI into their life decisions.

Z. Reference

• https://en.wikipedia.org/wiki/ML_Fairness
• https://fairlearn.github.io/user_guide/fairness_in_machine_learning.html#fairness-of-ai-systems
• https://www.youtube.com/watch?v=fMym_BKWQzk&ab_channel=TheArtificialIntelligenceChannel
  
  The End.

• 2020 21