Replication Project: Henderson et al. (2025)

This repository reproduces the ROI × Task × Boundary decoding analysis (near-trial ANOVA) reported in Henderson et al. (2025). It also re-creates Figure 2A–C to check whether the qualitative pattern of binary decoding accuracies emerges in an independent implementation. Beyond replication, the repository generates a new set of silhouette stimuli and evaluates their structure, including dimensionality and category separability under multiple decision boundaries.

Overview of this project is best summarized in the html file under the writeup directory.

Recreating the environment

The file henderson2025_env.yml specifies the full conda + pip environment used to run all Python analyses. After cloning the repository (see steps below), build this environment on your machine:

conda env create -f henderson2025_env.yml
conda activate henderson2025_env

All Matlab scripts utilized MATLAB_R2024b.

How to Run the Reproducibility Test

1. Clone the repository into whatever directory of your choosing, and then navigate into it:

git clone https://github.com/psyc-201/henderson2025
cd henderson2025

2. Create a folder called data inside the top layer of the repo directory.

3. Download the data files from the authors’ OSF repo and move them into the appropriate folders:

• Specifically, download the behav_all.zip, unzip it, then place this directory within data
• To download the voxel activation and timing data, download all files with the the author's Samples directory on OSF.
• You may have to download these Samples files one-by-one then move them ALL into a folder called Samples that you make yourself (the zip file seemed to have issues on OSF)
• Both Samples and DataBehavior should be nested within data

4. Open the file decoding_reproducibility/reproducing_decoding_anova_result.py in an IDE. Make sure you have built and activated the Python environment.

5. Run the script.
   It will:

• Load in the data from the .mat files and convert them into a Python format,
• Perform decoding within each ROI,
• Run a 3-way repeated-measures ANOVA (ROI × Task × Boundary),
• Save figures and results to the bold_decoding_anova_results/ folder. All outputted PNGs and CSVs are also available on on OSF.

How to Run the Simulation/Extension w/ New Stimuli

This extension pipeline recreates the representational analyses from Henderson et al. (2025) using a new set of silhouette stimuli. The Python script is organized into clear #%% step blocks, but you must begin by generating the new stimulus images in MATLAB.

1. Generate the new silhouettes in MATLAB

• Open the script stimuli/making_new_blobs.m in MATLAB.
• Open the script stimuli/making_new_blobs.m in MATLAB.
• Confirm that it writes 36 PNG files into the folder: stimuli/new_blob_stim/ These images will be used by all later steps.

2. Build the HDF5 image brick (Python)

• Open image_similarity/image_similarity.py in your IDE.
• Run the cell labeled: #%% step 1: make the image brick for matlab
• This step loads the PNGs and writes a single HDF5 file here: image_similarity/features/images_all.h5py

3. Run MATLAB GIST and prepare SimCLR weights

• Run the MATLAB GIST extraction script image_similarity/gist_matlab/get_gist.m (uses the the HDF5 image brick we created). This will produce the file: image_similarity/features/gist/images_gistdescriptors_4ori_4blocks.h5py
• Download the SimCLR checkpoint file checkpoint_100.tar from the SimCLR GitHub releases page.
• Place the checkpoint inside: image_similarity/features/simclr/ The Python feature extraction will not run until this file is present.

4. Extract SimCLR features (Python)

• In image_similarity.py, run the cell labeled: #%% step 3: extract simclr features with resnet
• This step computes PCA-reduced features for blocks 2, 6, 12, and 15 and saves files such as: image_similarity/features/simclr/images_simclr_block2_pca.npy

5. Compute category separability

• Run the cell labeled: #%% step 4: compute category separability and bar plot
• This step loads the GIST and SimCLR features, computes category separability for each decision boundary, and produces the figure: image_similarity/image_category_sep.png

6. PCA visualization

• Run the cell labeled: #%% step 5: pca on gist features and scatter plots
• This produces a 2-panel PCA visualization saved as: image_similarity/image_gist_pca.png

Overview of Directories

• writeup - contains Quarto markdown file and rendered html report. Contains a subdirectory, figs, that houses the figures used in the markdown report. These are available on OSF under the top-level figs directory
• image_similarity - contains the main Python script as well as the gist_matlab and features subdirectories. The produced PNGs from this pipeline are available here under the image_similarity directory. All intermediate files: e.g., h5py, mat, and CSV files, are also available on this OSF link under the subdirectories nested within image_similarity. Basically contains all Gist and SimCLR implementation, along with image_similarity.py
• stimuli - contains 'making_new_blobs.m` and output directory for newly generated shape stimuli
• post_hoc_power_analysis - contains Python script to run simple post-hoc power analysis on the Task x Interaction effect of interest
• original_paper - contains PDF copy of the Henderson et al. (2025) paper
• in_class_plotting_exercise - contains Quarto markdown file to plot the reproduced Figure 2A-2C (decoding accuracy plots) in R
• decoding_reproducibility - contains reproducing_decoding_anova_result.py