Are Convolutional Neural Networks or Transformers more like human vision?

Modern machine learning models for computer vision exceed humans in accuracy specific visual recognition tasks, notably on datasets like ImageNet. , high accuracy can be achieved in many ways. The particular decision found by a machine learning system is determined not only by the data which the system is exposed, but also the inductive biases of the model, are typically harder to characterize. In this work, we follow a recent of in-depth behavioral analyses of neural network models that go beyond as an evaluation metric by looking at patterns of errors. Our focus is comparing a suite of standard Convolutional Neural Networks (CNNs) and a-proposed attention-based network, the Vision Transformer (ViT), which the translation-invariance constraint of CNNs and therefore represents model with a weaker set of inductive biases. Attention-based networks have been shown to achieve higher accuracy than CNNs on vision tasks, and demonstrate, using new metrics for examining error consistency with more, that their errors are also more consistent with those of humans. results have implications both for building more human-like vision, as well as for understanding visual object recognition in humans.