Selected Inductive Biases in Neural Networks To Generalize Beyond the Training Domain


Artificial neural networks in computer vision have yet to approach the broad performance of human vision. Unlike humans, artificial networks can be derailed by almost imperceptible perturbations, lack strong generalization capabilities beyond the training data and still mostly require enormous amounts of data to learn novel tasks. Thus, current applications based on neural networks are often limited to a narrow range of controlled environments and do not transfer well across tasks.