aphilippsen/thecentraltendency_model

Fork of the repository for adding result files.

Simulating the central tendency effect in different social contexts using a predictive coding model

Code for replicating the results of the paper "The world seems different in a social context: a neural network analysis of human experimental data" authored by Maria Tsfasman, Anja Philippsen, Carlo Mazzola, Serge Thill, Alessandra Sciutti, and Yukie Nagai.

Link to the arXiv version: https://arxiv.org/pdf/2203.01862.pdf

Required python packages

• chainer
• numpy
• matplotlib
• seaborn
• dtw
• scikit-learn

Steps for reproducing results

1. Run run_training.py to train the network with the human experiment data. This step can be skipped if the pretrained models provided in results/training/all_human/ should be used.

2. Run evaluate_training.py to evaluate the networks' performance and run the internal representation analysis (the paths for the network which should be evaluated needs to be adjusted). The evaluation is done automatically with a number of different H (prior/sensory integration) values. The results are written to a subfolder evaluation inside the experimental results folder. If results are already available in this folder, they are reused.

3. Run modify_H_behavior.py to test which change of the precision of the prior or on sensory information

4. Run evaluate_training_summary_across_network.py to summarize the network results and compute which H values best replicate human data and to compute the distances between neuron activations of different conditions (see Figures 8 and 9 of the manuscript).

5. Run modify_H_behavior_statistics.py which processes the results of modify_H_behavior.py and saves it in a csv format to be used for statistical analysis with R.

6. Run statistics.R with R to test statistical difference between the three conditions of the human and of the model results.

Description of source code files

nets.py

SCTRNN implementation used the the network training code and to load stored networks into the correct data structure.

run_training.py

Trains the network with human data.

evaluate_training.py

Evaluates the network performance of all trained networks, including internal representation analysis.

evaluate_training_summary_across_networks.py

Summarizes the results obtained from several networks (requires results of evaluate_training.py and of modify_H_behavior.py).

modify_H_behavior.py

Changing the precision on prior (by H) and on sensory information (by sigma_inp) to a number of different values, and compare it to human data. Compares the effect of the change of either prior or sensory information (both affect the result in similar ways).

network_activation_summary.py

Computes the distances within conditions (see Figures 8 and 9 of the manuscript).

utils

Contains helper functions, e.g. quantitative_distances.py contains helper functions to compute distances between time series (behavioral trajectory, activation history, or PCA activations), and to sort the full data results by either length and condition, or by participant and condition. calc_regression_indices.py calculates the regression index.

statistics.R

Analyses of statistical significance between human and model results (human-model-comparison.csv) using linear mixed effect models and likelihood-ratio tests.

Description of other files

Data

The human data, preprocessed for usage with the neural network model is located in the folder human_data.

• presented.npy and presented_norm.npy contain the presented stimuli (original and normalized).

• human.npy and human_norm.npy contain the human reproductions of the stimuli (original and normalized).

• subjects.npy maps the data entries to the human subject ID (0-24).

• conditions.npy maps the data entries to the experimental conditions: 0 individual, 1 mechanical and 2 social condition.

• classes.npy maps the data entries to the initial states that the network uses for training. Each subject and each condition are represented by a different initial state, thus, 75 initial states are used in total.

Results

The ten trained networks used for the analysis are stored in results/training/all_human/ which contains a subfolder with the name of the training H_prior parameter (here, always H_prior = 1). Inside, the trained network is stored as network-final.npz.

The training parameters and the trained network can be loaded using the nets.load_network() function:

params, trained_model = load_network(PATH_TO_FOLDER, model_filename='network-final.npz')

In the evaluation subfolder, all evaluation results are stored.