Exploring perceptual straightness in learned visual representations

Authors: Anne Harrington, Vasha DuTell, Ayush Tewari, Mark Hamilton, Simon Stent, Ruth Rosenholtz, William T. Freeman

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

Reproducibility Variable Result LLM Response
Research Type Experimental We measured curvature for a variety of models (Sec A.2.1) to investigate the relationship between model type and output curvature. As shown in Figure 3, we find non-adversarially trained image recognition models have the highest output curvature.
Researcher Affiliation Collaboration Anne Harrington1,2 Vasha Du Tell1,2 Ayush Tewari1 Mark Hamilton1 Simon Stent3 Ruth Rosenholtz1,2 William T. Freeman1 1MIT CSAIL 2MIT Brain and Cognitive Sciences 3 Woven Planet
Pseudocode No The paper does not contain a pseudocode block or algorithm block.
Open Source Code Yes To reproduce our results, links to all the models analyzed can be found in the supplemental material in the Network Comparison Spreadsheet. Sources for the stimuli are in Sec A.2.2 and in the code base linked for the Henaff Bio model in Network Comparison Spreadsheet.
Open Datasets Yes All adversarially trained models have lower curvature than their non-adversarially trained counterparts (Sec A.4), as well as overall, with the majority reducing output curvature below that of the input pixels. Self-supervised DINO (Caron et al., 2021) models have similar output curvature values to their supervised counterparts despite DINO models having been shown to have more semantically meaningful feature correspondences.
Dataset Splits No The paper does not explicitly provide details about training, validation, and test splits (e.g., percentages or counts) for reproducibility beyond mentioning the datasets used.
Hardware Specification No The paper mentions using a single GPU and refers to the MIT Super Cloud and Lincoln Laboratory Supercomputing Center but does not provide specific hardware details like GPU model numbers (e.g., NVIDIA A100, Tesla V100), CPU types, or memory specifications.
Software Dependencies Yes The standard Image Net-trained Res Net (He et al., 2016) model was downloaded from Py Torch s model zoo (Paszke et al., 2019). All Vi T (Dosovitskiy et al., 2020) and standard trained Cross Vi T daggar (Chen et al., 2021) models were downloaded from the timm library (Wightman, 2019).
Experiment Setup No The paper describes the videos used and how curvature was calculated (Sec A.2.2), but it does not provide specific hyperparameters (e.g., learning rate, batch size, number of epochs) or detailed training configurations for the models it analyzes.