Convolutional neural networks—the machine-learning systems routinely used for image recognition—are inherently limited, according to a new study by @lamaral and colleagues. #ConvNets fail on data above a certain threshold of complexity. They lean too heavily on shortcuts—spurious correlations that don’t generalize—and on localized features such as texture, to the neglect of the overall scene. h/t @manlius https://doi.org/10.1063/5.0213905 #AI