Getting started with rcicr
In rcicr: Reverse-Correlation Image-Classification Toolbox

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.width = 4,
  fig.height = 4
)

rcicr implements reverse correlation image classification, a psychophysics technique for visualizing mental representations, for example of faces. It works in two stages:

Stimulus generation. A base image, such as a face photo, is combined with random visual noise. Each trial shows a pair: the "original" (base plus noise) and its "inverted" counterpart (base minus the same noise). On each trial the participant picks whichever of the two looks more like a target category, such as "trustworthy" or "happy". This is a two-image forced-choice (2IFC) task.
Classification image (CI). After data collection, the noise of every chosen original is added up and the noise of every chosen inverted image is subtracted. The average is the classification image: it shows which visual features drove the participant's choices.

This vignette runs both stages on a tiny synthetic example. For the full method, with several participants, scaling choices, z-maps and informational value, see vignette("reverse-correlation-walkthrough", package = "rcicr"). Example datasets and analysis scripts are in rcicr_examples.

library(rcicr)

1. Generate stimuli

generateStimuli2IFC() needs a square base image. To keep this vignette self-contained we make a synthetic greyscale one. In a real study you pass the path to your base face photo(s).

set.seed(42)
base_face_path <- tempfile(fileext = ".png")
png::writePNG(matrix(runif(64 * 64), 64, 64), base_face_path)

Now generate stimuli for a small task: 20 trials and one base image, at a small size so the vignette builds quickly. A real study typically uses img_size = 512 and several hundred trials (Dotsch & Todorov, 2012).

stimulus_path <- tempfile("stimuli")

generateStimuli2IFC(
  base_face_files = list(face = base_face_path),
  n_trials        = 20,
  img_size        = 64,
  stimulus_path   = stimulus_path,
  seed            = 1,
  ncores          = 1,
  save_as_png     = FALSE # set to TRUE to also write stimulus PNGs to stimulus_path
)

rdata_file <- list.files(stimulus_path, pattern = "\\.Rdata$", full.names = TRUE)[1]

This writes an .Rdata file to stimulus_path holding the noise parameters of every trial. That file is the only link between stimulus generation and CI computation. Keep it: every analysis function below needs it as its rdata argument.

2. Collect (or, here, simulate) responses

In a real experiment you now run the 2IFC task and record, per trial, which image each participant chose: 1 for the original, -1 for the inverted one. This vignette has no participant, so it simulates random responses. Random responses carry no signal, so the resulting classification image shows nothing; never do this in a real analysis.

responses <- sample(c(1, -1), 20, replace = TRUE)

3. Compute the classification image

generateCI() looks up the noise parameters of the stimuli that were shown, weights them by the responses, and averages them into one classification image.

ci <- generateCI(
  stimuli     = 1:20,
  responses   = responses,
  baseimage   = "face",
  rdata       = rdata_file,
  save_as_png = FALSE
)

names(ci)

The result holds four pixel matrices:

ci$ci is the raw noise.
ci$scaled is that noise rescaled for display. The default method, 'independent', picks the lowest scaling constant that avoids clipping this particular image; ?generateCI describes the others.
ci$base is the base image.
ci$combined overlays the scaled noise on the base image.

image(ci$combined, col = gray.colors(256), axes = FALSE, asp = 1)

Because the responses were random, this classification image is just noise. With real data, patterns tied to the participants' choices emerge here.

Next steps

batchGenerateCI() and batchGenerateCI2IFC() compute one CI per participant or condition from a data frame. By default they put the whole batch on one scale with autoscale(), so the images can be compared by eye.
computeInfoVal2IFC() computes the informational value: a z-score-like measure of how much signal a CI holds, compared with a simulated null distribution.
plotZmap() shows which regions of a CI carry reliable signal.

Each function's help page, such as ?generateCI or ?batchGenerateCI, lists its options and has runnable examples.