Vision Transformers provably learn spatial structure

Authors: Samy Jelassi, Michael Sander, Yuanzhi Li

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Lastly, we empirically verify that a Vi T with positional attention performs similarly to the original one on CIFAR-10/100, SVHN and Image Net.
Researcher Affiliation Academia Samy Jelassi Princeton University sjelassi@princeton.edu Michael E. Sander ENS, CNRS michael.sander@ens.fr Yuanzhi Li Carnegie Mellon University yuanzhil@andrew.cmu.edu
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] See supplementary material.
Open Datasets Yes On the experimental side, we validate in Section 6 that Vi Ts learn spatial structure in images from the CIFAR-100 dataset... competitive with the vanilla Vi T on the Image Net, CIFAR-10/100 and SVHNs datasets.
Dataset Splits Yes Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] See Appendix.
Hardware Specification Yes Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] See Appendix.
Software Dependencies No The paper mentions 'Adam W' as an optimizer but does not provide specific version numbers for software dependencies or libraries.
Experiment Setup Yes For the small datasets, we use a Vi T with 7 layers, 12 heads and hidden/MLP dimension 384. For Image Net, we train a 'Vi T-tiny-patch16-224' [24]. Both models are trained with standard augmentations techniques [18] and using Adam W with a cosine learning rate scheduler. We run all the experiments for 300 epochs, with batch size 1024 for Imagenet and 128 otherwise and average our results over 5 seeds. We refer to Appendix A for the training details.