Insight

This repository accompanies the following paper, currently under review: Mihali, A, Broeker, M, Ragalmuto, F, Horga, G. Introspective inference counteracts perceptual distortion, 2023, bioRxiv. In addition to the co-authors and several colleagues listed in the acknowledgements, also thanks to @astroahad!

This repository contains demo data and analysis and modeling code. We do not provide the experimental code here for now, but mention that to implement the spiral stimuli we adapted the implementation of Peter Scarfe's Contrast Modulated Spiral.

Watch the video of our MAE illusion implementation here. Fixate on the center white dot (here for 15 seconds) and report the direction of motion of the spiral which appears when the central dot changes color from white to yellow.

Software requirements

This package has been tested on macOS Monterey (12.5.1) with the following:

• Matlab 9.6 (2019a, MathWorks, Massachusetts, USA)

• We provide the output results of the Bayesian model fitting procedure. To run our model fitting scripts, please also download the Bayesian adaptive direct search (BADS) optimization algorithm by Luigi Acerbi.

Demo and instructions for use: Data

We have datasets from Exp. 1 and Exp. 2 as alldata_E1.mat and respectively alldata_E2.mat, which are Nsubj * Ncond, with Nsubj being 22 for each. There are 4 conditions, combined such that they are in the following order:

• 1: No-Adapt-See
• 2: No-Adapt-Believe
• 3: Adapt-See
• 4: Adapt-Believe

All the participants performed Adapt-See first, followed by Adapt-Believe (except participant 2 in Experiment 2 who accidentally performed Adapt-Believe before Adapt-See). In Experiment 1, participants then performend No-Adapt-See followed by No-Adapt-Believe; in Experiment 2 the order of these last two conditions was randomized across participants.

All analyses were performed on the alldata structs. For each subject and condition, there are 5 fields, each with variables of length Ntrials (121):

• stims: uniformly distributed on [-0.3, 0.3]
• resp: 0 (left/CCW) or 1 (right/CW)
• conf: 0 for low confidence, 1 for high confidence
• resp_times (sec)
• conf_times (sec)

Scripts

• For individual model fits, full simulated datasets and fits, complement the scripts from here with the data in the corresponding folders: More on data and model fits

• analysis_all.m outputs a large number of data figures from our manuscript: Figures 3, 4, 5 and 7 from main, Figures S2, S3 from Supplementary, and, if we change the model index from mi = 1 to mi = 4, we could also get Figure S6. To overlay the model predictions on top of the data, it calls Predict_bm_alll.m. It loads the entire dataset, psychometric curve fits (psych_curves_fitting_m2_201_E2.mat for exp.2) and, for exp_i = 2, Bayesian model fits (params_all_models6_E2_Nsubj_22.mat).

Psychometric curve fitting

• To fit the psychometric curves with the lapse parameter shared across the 4 conditions, we used the function psych_curve_fitting_offline_m2.m, which uses grid search for optimization and calls loglike.m, which calls function_psi.m.

Bayesian models fitting

• We fit the 6 Bayesian model variants via fitting_exp_alll.m. This script is ready to run but it takes a long time; the results will be saved as params_all_models_E2_Nsubj_22.mat, which we provide.fitting_exp_alll.m calls Fitting_model.m, which finds the parameters that maximize the probability of the data given each model (maximum likelihood estimation) via Loglike_bm_alll.m. Fitting_model.m loads the data, here alldata_E2.mat, as well as the psychometric curve fits, params_psych_curves_exp2_m1_m2_all.mat and outputs the best fitting parameters params_all_models_E2_Nsubj_22.mat.

Bayesian models simulations

sims_mu_like_vs_priorE.m generates Figure S5 and parameter_recovery_Bayesian_model_1_plot.m generates Figure S7.

Bayesian models simulations full

fitting_all_model_sim_and_output_plotting.m generates Figure S8.

Pupillometry analyses

• regressions_pupilAlll.m generates Figure 6 from our paper (once the full dataset will be made available). Note that it loads dv_val_mat.mat - generated by the analysis_all.m script described above, as well as GLMEs_sliding_window_stim_locked_.mat and GLMEs_sliding_window_resp_locked_.mat, generated by the script regressions_pupil_sliding_window_analysis_FIN.m. Note that due to their large size, the majority of the .mat files will need to be downloaded from More on data and model fits. The output from the Eyelink eye tracker was parsed with variants of previous scripts: https://github.com/lianaan/Microsacc/tree/master. Specifically, the raw eye tracking data uploaded in OSF was parsed with eye_text_edE5_monocular.sh, and the outputs combined to generate the pupil_all_tr_cond files.

DDM

• While analysis_all.m loads ddm_params_and_model_comp.mat and output_list_RT_allEE.mat and output_list_choices_allEE.mat to generate Figure 7, here we show how we implemented the drift diffusion model fitting with PyDDM from Shinn et al, 2020.

Pilot control experiment data

• We also share the data from Exp. 2 control, which we used to generate Figure S9. This is alldata_E2ctr.mat, which is also Nsubj * Ncond, with Nsubj being 7. There are 4 conditions:

• 1: No-Adapt-See

• 2: No-Adapt-Bias

• 3: Adapt-See

• 4: Adapt-Bias