Why Should De-Biasing Be Made A Priority in Machine Learning

558 0

The era of AI & Machine Learning has already begun. From self driving cars to online recommendation in social media and e-commerce, we see instances of its widespread presence everyday. The increasing popularity of machine learning is also reflected in the numerous machine learning courses that are available today. 

Though Machine Learning models are a definite plus to our society, there are still instances when they do not work as intended, thus defeating the whole purpose of convenience and ease that it stands for. Most of the time when ML models fail to work as expected, the reason can be found in the training method and training data. Biases can enter ML models at various intervals such as during data modelling, information assortment, data preparation or evaluation. Some of the usual biases in machine learning models are Sampling, Confirmation, Anchoring and Performance bias. 

Sampling bias relates to when models are trained using data that is not fully representative of future cases. In Confirmation bias, data will be taken, prioritized and recollected in ways that prove the previously held ideas. Anchoring bias on the other hand, will cause over dependence on primary data that is used. And Performance bias will affect how predictive power, performance uniformity, and generalizability of data is seen. 

How to Eliminate Bias In Machine Learning Models?

Following are some of the steps that can be taken to ensure a fair machine learning design:

  • Be careful during annotation: Humans are responsible for interpreting unstructured data such as pictures and texts that are given structured class labels for training ML models. For instance, annotators can provide positive or negative class labels to pictures of people or a specific. This will then be imbibed by the particular ML model which will start applying this learning to other similar texts or people.  Though annotations made by humans are standard for several different tasks, they are still vulnerable to culturally planted biases.

  • Make de-biasing a priority:  De-biasing should be considered an integral step throughout the training and development of the model. There are many ways to ensure that. One such way is to remove all demographic details, both explicit and implicit, from the training data. Yet another way to ensure de-biasing is to include certain fairness measures into the machine learning model’s design.

  • Incorporate fairness measures into traditional ML metrics: The performance of machine learning classification models are determined using the traditional metrics that specializes in overall performance, all-round model generalizability, and performance at class level. There is always a space to improve it by including some measures promoting fairness and eliminating biases in machine learning models. These kinds of performance metrics are essential for achieving situational awareness.  
  • Include certain mass constraints with representativeness of data during sampling: The traditional way of sampling ensures that the training data is statistically representative of any future cases that the particular machine learning model can come across in the future. One concern with this method is that it does not consider the case of minorities, individuals who are statistically less common. As you know, ML models are trained to find out patterns from the data and then generalize it to a large group. So when the training data does not contain information about individuals who are statistically minorities in the data set, the ML model will ignore all the learning concerning such people. 

Conclusion

There were many cases in recent years which showcased the disastrous consequences of machine learning models trained using biased data. These models exhibited the biases with which they were trained. For instance, there was one case where a bot trained using twitter data, started spewing racial slurs and had to be shut down. So hopefully in the future, the above mentioned steps can be taken while developing the model.

 

About Srilatha Kante

Srilatha Kante

Srilatha Kante, SpringPeople technology evangelist and master trainer, has over a decade of experience in software development and testing. Some of her areas of expertise include resource management, automation framework development, SDLC, QA process, test planning, mule ESB, API,RAML, and Anypoint Platform technologies.


Posts by Srilatha Kante

Leave a Reply

Your email address will not be published. Required fields are marked *

CAPTCHA

*