Finding Biological Plausibility for Adversarially Robust Features via Metameric Tasks
Authors: Anne Harrington, Arturo Deza
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | we performed a psychophysics experiment using a metamer task similar to Freeman & Simoncelli (2011); Wallis et al. (2019); Deza et al. (2019b) where we evaluated how well human observers could distinguish between images synthesized to match adversarially robust representations compared to nonrobust representations and a texture synthesis model of peripheral vision (Texforms (Long et al., 2018)). |
| Researcher Affiliation | Academia | Anne Harrington & Arturo Deza Center for Brains, Minds and Machines Massachusetts Institute of Technology {annekh,deza}@mit.edu |
| Pseudocode | No | The paper does not contain any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Link to Code/Data: https://github.com/anneharrington/Adversarially-Robust-Periphery. |
| Open Datasets | Yes | Both models were trained on a subset of Image Net (Russakovsky et al., 2015), termed Restricted Image Net (Table 1). |
| Dataset Splits | No | The paper mentions models were trained on 'Restricted Image Net' and images were chosen from its 'test set', but it does not specify the training, validation, or test split percentages or exact counts for their own experimental setup. |
| Hardware Specification | No | The paper describes the monitor and experimental setup hardware: 'Experiments were ran on an Ubuntu-Linux Machine version 14.04.5 LTS with MATLAB 2015a s Psychtoolbox version 3.0.14... We used a 34 inch diagonal 75 Hz LCD monitor, that measured 80cm width and 34 cm height with a visual display resolution of 3440 pixels width by 1440 pixels height.' However, it does not specify the CPU, GPU, or memory used for model training or synthesis. |
| Software Dependencies | Yes | Experiments were ran on an Ubuntu-Linux Machine version 14.04.5 LTS with MATLAB 2015a s Psychtoolbox version 3.0.14. |
| Experiment Setup | Yes | All stimuli were synthesized at a size of 256 256 pixels, this was equivalent to 6.67 6.67 degrees of visual angle (d.v.a.) when performing the psychophysical experiments... The oddity task consisted of finding the oddball stimuli out of a series of 3 stimuli shown peripherally one after the other (100ms) masked by empty intervals (500ms) while holding center fixation... We had 12 participants complete both the oddity and matching 2AFC experiments. |