Thyer-et-al-2021

Overview and setup

Analysis code for Thyer et al 2021 (in prep). Working title: "Multivariate decoding of visual working memory load provides evidence for item-based pointers". This is the code used to generate analysis and figures. Data will be made available on OSF. Put data in folder titled data in top directory.

This decoding uses the Mord package which is a ordinal logistic regression package. Download the package and put it in a folder called mord in the top directory. Alternatively, just use the sklearn.linear_model.LogisticRegression for slightly worse decoding accuracies.

Files

decode_eeg.py

• Contains vast majority of functionality for loading and wrangling data, cross-validation, classification, plotting, and stats testing. All notebooks import this package.

• Experiment

  • Handles loading data. Directories, data info, subjects, etc.

• Experiment_Syncer

  • Handles synchronizing subjects across multiple experiments using unique IDs. Also uses Experiment functions.

• Wrangler

  • Handles data augmentation before classification. That includes trial binning, selecting classes, cross-validation, and rolling over time.

• Classification

  • Handles actual classification. Basically standardizing data, training, and testing models.

• Interpreter

  • Takes in results from classification. Handles all of the plotting and statistical testing of results.

• ERP

  • Less connected than the other classes. Handles plotting of ERPs.

decode_eeg.ipynb

• Basic notebook for load classification within an experiment. Not actually used, since decode_eeg_loop.ipynb handles all within-experiment decoding. Still a useful template.

decode_eeg_loop.ipynb

• Loops through experiments and trial bin size parameter and performs classifications.

decode_load_single_feature.ipynb

• Trains classifiers on experiment 1 (color) and tests on experiment 2 (orientation) and vice versa.

decode_load_single_feature_set_size.ipyng

• Train classifiers on mixture of experiment 1 (color) and tests on experiment 2 (orientation) with set size 1 vs 2, 2 vs 3, & 3 vs 4.

decode_load_single_feature_to_conjunction.ipynb

• Trains classifiers on experiment 1 and 2 (single feature, color or orientation) and tests on experiment 3 (conjunction, color & orientation).

decode_preds_dont_double.ipynb

• Trains classifiers on experiment 1 and 2 (single feature, color or orientation) 1 vs 2 and 2 vs 4. Also trains classifiers on experiment 3 (conjunction, color & orientation) 1 vs 2 and 2 vs 4. Then compares predictions from those classifiers.

decode_load_acc_and_k.ipynb

• Classifies load on all unique subjects across all three experiments. Also calculates K for each subject. Then correlates classification accuracy and K.

decode_load_colinearity_hyperplane.ipynb

• Using data from Diaz et al. 2021, train classifier on set size 2 ungrouped and set size 4 ungrouped. Then test on set size 4 grouped (4 items, but arranged to look like 2). Measure distance from hyperplane for each condition.

behavior_analysis.ipynb

• Plots behavior. Specifically accuracy across set sizes.

plot_erp.ipynb

• Plots frontal, central, and parietal/occipital ERPs for each experiment.

decode_load_eyetracking.ipynb

• Classifies load in Experiment 1 using EOG data. Plots accuracy, setsize pair accuracy, and confusion matrices.