Data and analysis code from "Beyond accuracy: quantifying trial-by-trial behaviour of CNNs and humans by measuring error consistency"

Error consistency is a quantitative analysis for measuring whether two decision making systems systematically make errors on the same inputs. The paper is available on arXiv.

dependencies

The R analysis scripts have the following dependencies which can be installed via install.packages("package-name"). Data analysis was performed using R version 3.5.1.

library(lattice)
library(jpeg)
library(R.matlab)
library(graphics)
library(pROC)
library(psych)
library(grid)
library(gridExtra)
library(stats)
library(png)
library(pBrackets)
library(PET)
library(TeachingDemos)
library(binom)
library(RColorBrewer)
library(ggplot2)
library(scales)
library(xtable)
library(viridis)
library(binom)

The Brain-Score parsing script has the following dependencies:

pip3 install jupyterlab
pip3 install urllib
pip3 install bs4
pip3 install numpy

data-analysis

Scripts to analyse the data from raw-data/ and plot figures to figures/.

The main analysis script is data-analysis.R. Confidence intervals are simulated via simulate-confidence-intervals.R. Brain-Score metrics are parsed from the Brain-Score website with get_data_from_Brain_Score.ipynb.

documentation

Project-related documentation.

figures

Figures generated by scripts from data-analysis/ using data from raw-data/.

raw-data

Experiments with the prefix noise-generalisation are from this paper; the raw data is copied from the corresponding github repository. Experiments with the prefix texture-shape are from this paper; the raw data is copied from the corresponding github repository.