Generalisation in humans and deep neural networks

Authors: Robert Geirhos, Carlos R. M. Temme, Jonas Rauber, Heiko H. Schütt, Matthias Bethge, Felix A. Wichmann

NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We compare the robustness of humans and current convolutional deep neural networks (DNNs) on object recognition under twelve different types of image degradations.
Researcher Affiliation Academia 1Neural Information Processing Group, University of Tübingen 2Centre for Integrative Neuroscience, University of Tübingen 3International Max Planck Research School for Intelligent Systems 4Graduate School of Neural and Behavioural Sciences, University of Tübingen 5Department of Psychology, University of Potsdam 6Bernstein Center for Computational Neuroscience Tübingen 7Max Planck Institute for Biological Cybernetics 8Max Planck Institute for Intelligent Systems
Pseudocode No The paper does not contain any pseudocode or clearly labeled algorithm blocks.
Open Source Code Yes Our carefully measured behavioural datasets twelve experiments encompassing a total number of 82,880 psychophysical trials as well as materials and code are available online at https://github.com/rgeirhos/generalisation-humans-DNNs.
Open Datasets Yes We thus developed a mapping from 16 entry-level categories such as dog, car or chair to their corresponding Image Net categories using the Word Net hierarchy [51]. We term this dataset 16-class-Image Net since it groups a subset of Image Net classes into 16 entrylevel categories (airplane, bicycle, boat, car, chair, dog, keyboard, oven, bear, bird, bottle, cat, clock, elephant, knife, truck). Our new dataset consisting of 83K carefully measured human psychophysical trials provide a useful reference for lifelong robustness against image degradations set by the human visual system.
Dataset Splits No The paper mentions training and testing but does not explicitly provide specific percentages or sample counts for training, validation, and test splits.
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models, memory specifications, or types of computing environments used for the experiments.
Software Dependencies Yes Training was done using Tensor Flow 1.6.0 [58].
Experiment Setup Yes The remaining aspects of the training followed standard training procedures for training a Res Net on Image Net: we used SGD with a momentum of 0.997, a batch size of 64, and an initial learning rate of 0.025. The learning rate was multiplied with 0.1 after 30, 60, 80 and 90 epochs (when training for 100 epochs) or 60, 120, 160 and 180 epochs (when training for 200 epochs).