Modeling machine learning: A cognitive economic approach

Abstract

We investigate whether the predictions of modern machine learning algorithms are consistent with economic models of human cognition. To test these models we run an experiment in which we vary the loss function used in training a leading deep learning convolutional neural network to predict pneumonia from chest X-rays. The first cognitive economic model we test, capacity-constrained learning, corresponds with an intuitive notion of machine learning: that an algorithm chooses among a feasible set of learning strategies in order to minimize the loss function used in training. Our experiment shows systematic deviations from the testable implications of this model. Instead, we find that changes in the loss function impact learning just as they might if the algorithm was a human being who found learning costly.