Can Cognitive Bias in AI/ML Models be eliminated ?
Recently the Artificial Intelligence / Machine Learning (AI/ML) space has been buzzing to say the least. Chat GPT has raised quite a storm of debates and questions about how AI is going to take over the jobs and then consequently enslave all of humanity and the conquer the world! A bit too early I would say to even think like this. Its like that chicken screaming the Sky is falling down.
While AI/ML is here to stay, I see it as more of a help to humans. AI/ML is only as intelligent as the humans who create and train it, and nothing more. We as humans understand so little about our own brains yet, that giving more credit to AI/ML as being smarter is a bit over the edge (in my opinion).
What is AI/ML ?
AI/ML is a small piece of code that takes huge chunks of information (data in computer terms) and then tries to make some sense and draw some insights from it using algorithms. The algorithms learn and adapt based on the input data and essentially the model grows. The way the algorithm is tweaked and modified based on the data is called training the model. Much like we learn in steps and then based on some new information, change our thoughts and beliefs as we go along.
Cognitive Bias
We as humans, have different perspectives about anything and everything in life. No two persons will have the exact same way of thinking, interpreting and processing facts and information. This is largely due to different cultures, upbringings, situations, educations and even our brain wiring.
The AI/ML algorithm is only as diverse, inclusive and adaptive as the humans that develop it. The models will be limited to the awareness and interpretations of the humans making it. Not to say that there’s anything wrong about it. But if you take into account that relying on these models as we go ahead, there could be a large population that could be excluded from these models, just by virtue of the face that the humans who wrote them were not aware of the other side.
To put this into perspective, I recently attended a workshop on AI/ML pipelines. There was a statistic mentioned there which stated that 86% of AI/ML models do not make it to production. Which means that currently 20% of models are doing all the processing. If you apply the Pareto Principle, this sounds reasonable enough. Now applying this same principle to the 20% models, it means that 20% of these models are doing 80% of the predictions. Either way you look at it, 20% of data is going to get missed.
For example, if the fact that 20% of people are left handed is missed out, a lot of applications which work just fine for right handed people are not going to work for 20% of the population. And this even applies to a humble pair of scissors.
This is Cognitive Bias, where one way of thinking result in another way to thinking being excluded.
Can Cognitive bias be removed completely ?
Recently there has been a lot of awareness about inclusivity and diversity at workplaces. There have been a lot of efforts made to include under represented populations, like women, minority ethnic communities and so on, a part of mainstream. This has been done by creating specific guidelines and sensitizing people about this.
AI/ML models need a similar set of guidelines and frameworks. Which may not completely eliminate this bias, but at least make an attempt to reduce it. A conscious effort needs to be made while creating these models to try and include all possibilities. Also having more diversity in the developing teams will help this cause. What do you think ? Can cognitive bias be completely removed ?
Thanks for reading :) !
