Do Invariances in Deep Neural Networks Align with Human Perception?

Authors: Vedant Nanda, Ayan Majumdar, Camila Kolling, John P. Dickerson, Krishna P. Gummadi, Bradley C. Love, Adrian Weller

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We then conduct an in-depth investigation of how different components (e.g. architectures, training losses, data augmentations) of the deep learning pipeline contribute to learning models that have good alignment with humans. Code: github.com/nvedant07/Human-NN-Alignment. We strongly recommend reading the arxiv version of this paper: https://arxiv.org/abs/2111.14726.
Researcher Affiliation Academia Vedant Nanda1,2, Ayan Majumdar2, Camila Kolling2, John P. Dickerson1, Krishna P. Gummadi2, Bradley C. Love3,4, Adrian Weller3,5 1University of Maryland, College Park, USA 2Max Planck Institute for Software Systems (MPI-SWS), Germany 3The Alan Turing Institute, London, England 4University College London, London, England 5University of Cambridge, Cambridge, England
Pseudocode No No explicit pseudocode or algorithm blocks labeled as 'Pseudocode' or 'Algorithm' were found.
Open Source Code Yes Code: github.com/nvedant07/Human-NN-Alignment.
Open Datasets Yes Feather et al. studied representational invariance for different layers of DNNs trained over Image Net data (using the standard cross-entropy loss).
Dataset Splits No No specific details on train/validation/test splits, percentages, or cross-validation setup are explicitly provided in the main text.
Hardware Specification No No specific hardware details (e.g., GPU models, CPU types, or cloud instance specifications) used for running experiments are provided in the paper.
Software Dependencies No The paper mentions software like 'Py Torch Lightning', 'Pytorch', 'Num Py', 'Matplotlib', and 'robustness (Python Library)' but does not provide specific version numbers for these dependencies.
Experiment Setup Yes This is achieved by performing gradient descent on x0 (in our experiments we use SGD with a learning rate of 0.1) to minimize a loss of the following general form: