Investigating Human Inference: Message Passing vs Sampling

This repository contains code for investigating whether humans use message passing or sampling to perform inference. The comparison is performed by fitting both models to human choices and reaction times from the Emotional Faces Task.

---

🧠 Task Summary

In the Emotional Faces Task, participants are presented with:

1. A low or high tone noise
2. Followed immediately by an image of either an angry or sad face

Key details:

• The tone and face occur in quick succession
• The image is shown very briefly (0.15 seconds)
• Participants respond with the left or right arrow key to indicate whether they saw an angry or sad face
• If no response is made quickly enough, "too slow" is displayed before the next trial

The objective is to determine whether human inference in this task is better explained by a message passing model or a sampling-based model.

---

---

📁 pyddm_scripts

This folder contains the drift-diffusion model (DDM) implementation and fitting code using pyddm.

fit_pyddm_emotional_faces.py

Fits the DDM to behavioral data from the Emotional Faces Task:

• Loads processed data for a given subject
• Defines a custom loss function (EF_Loss) which calls the model in DDM_model.py
• Fits the model to empirical data to estimate drift and learning parameters
• Simulates behavior from fitted parameters and re-fits to check recoverability
• Saves parameter estimates, model fit statistics, and simulated outputs

DDM_model.py

Implements the trial-by-trial Kalman filter-based DDM for the task:

• Computes drift values based on trial type (angry vs sad)
• Updates expected value estimates using a learning rate
• Fits: calculates reaction time PDFs, choice probabilities, prediction errors, and absolute RT errors
• Simulations: generates synthetic choices and RTs given model parameters
• Returns detailed trial-by-trial model statistics, including entropy measures

---

📁 RxInfer_scripts

This folder contains the message passing model of the task implemented using RxInfer and reactive message passing.
The model uses probabilistic graphical modeling to represent the task and infers latent variables via message passing rather than sampling.

---

⚙️ Dependencies

Python scripts require:

• Python 3.8+
• pyddm
• numpy
• pandas
• matplotlib
• scipy

RxInfer scripts require:

• Julia 1.8+
• RxInfer.jl