Why is the usage machine learning a successful measure against the "confirmation bias"?

The "confirmation bias" is a thought or decision error, which is based on the fact that people tend to be prefer approval of their own thoughts and decisions and therefore tend to remain in existing solutions and thinking patterns rather than to treat new solutions and hypotheses equally weighted and objectively.

An increased tendency towards confirmation bias usually exists when someone has already invested a lot of energy and/or money in the development of a solution or a model and thus no longer wants to move away from it, even if new solutions promise obvious advantages.

In the case of technological problems, such a bias is naturally often existent in the case of complex problems by means of analytical, rule-based control models and algorithms. For example, if you invest an entire PhD thesis to create analytical, equation-based formulas to solve a problem, you naturally don't like to hear that others solve the same problem in a fraction of the effort using machine learning or soft computing techniques.

The above graphic schematically shows the development effort depending on the complexity of the problem or the number of functional requirements. In contrast to data- and example-based approaches, for example with machine learning methods, this increases exponentially in the case of analytical, equation-based approaches or with bottom-up approaches. In line with the significantly lower, necessary efforts when using machine learning procedures, it is usually much easier mentally to leave a path that may have been wrongly taken and to revise previous (mis)decisions and to pursue new or adapted approaches and hypotheses. This solution-free, broad and flexible exploration of many and diverse solutions and hypothesis is an inherent part of machine learning and is thus better immunized against the confirmation bias.

Last update on 2020-12-28 by Andreas Kuhn.

Go back