Exploring out of distribution: Deep neural networks and the human brain

Deep neural networks have achieved state-of-the-art performance across a wide range of tasks. Convolutional neural networks, with their ability to learn complex spatial features, have surpassed human-level accuracy on many image classification problems. However, these architectures are still often unable to make accurate predictions when the test data distribution differs from that of the training data. In contrast, humans naturally excel at such out-of-distribution generalizations. Novel solutions have been developed to improve a deep neural net's ability to handle out-of-distribution data. The advent of methods such as Push-Pull and AugMix have improved model robustness and generalization. We are interested in assessing whether or not such models achieve the most human-like generalization across a wide variety of image classification tasks. We identify AugMix as the most human-like deep neural network under our set of benchmarks. Identifying such models sheds light on human cognition and the analogy between neural nets and the human brain. We also show that, contrary to our intuition, transfer learning worsens the performance of Push-Pull.