The Futility of Bias-Free Learning and Search

Along with Professor George Montañez, students Jonathan Hayase ’20, Julius Lauw ’20, Dominique Macias ’19, Akshay Trikha ’21 and Julia Vendemiatti ’21 published a paper titled “The Futility of Bias-Free Learning and Search.” Learning algorithms are machines that turn data resources into predictions. Their paper shows that unless algorithms do this conversion in a biased way, predisposing their predictions toward predetermined outcomes, they cannot predict any more accurately than random guessing. The paper proves that finding a good bias for a given problem is difficult, when searching among any set of data resources that on average isn’t itself positively biased. These results apply to machine learning algorithms, AI systems, genetic learning algorithms and many other forms of search and optimization.