The Convoluted Nature of Cognition: Revealing the Priors of Deep Learning Models Through Iterated Learning

Abstract

With the continuous improvement of deep learning models and their resulting integration into real world systems comes an increasing need to critically analyze the inductive biases of such models. For this project, we probe the inductive biases of Convolutional Neural Networks (CNNs) by observing their end behavior in an iterated retraining algorithm. The Markov chains which define iterated learning, a knowledge transmission process in which each agent learns a distribution of hypotheses based on data produced by the previous agent, have been shown to converge to the priors of the learning agents. By using CNNs as the learning agents in an iterated retraining algorithm, the resulting softmax distributions and divisions of sampled labels which they produce are able to reveal if the CNNs contain any inherent inductive biases and how they incorporate these biases into their classifications. When we pretrained the models to establish and control their priors, we found that the posterior distribution conditioned on pretraining the CNNs on the MNIST digits dataset is narrower than that on the CIFAR-10 natural images dataset. Additionally, we observed that the iterated retraining process clearly reflected the priors of the learning agents; models with stronger priors were able to make predictions with increased confidence and classify the images to a higher accuracy. Thus, iterated learning proves to be a useful algorithm in analyzing the inductive biases of deep learning models.