Adaptive learning through surprise-gated attractor switching

Models adaptive learning via surprise-driven attractor switching in cortical networks.

• Implements biologically-plausible attractor networks (PFC, Basal Ganglia, Mortor Cortex) that track latent environmental states
• PFC switches between discrete attractor states when entropy-based surprise signals exceed threshold
• Hebbian learning links PFC attractors to motor predictions via basal ganglia pathway
• Fits model to human behavioral data from changepoint detection tasks (McGuire & Nassar 2014)
• Generates publication figures comparing model and human peri-changepoint learning rates

Module Structure

• run.py: main entry point for reproducing the analysis and figures.
• configs.py: shared paths, cache flags, figure settings, and multiprocessing settings.
• model/: network implementation, including the optional Cython-accelerated backend.
• tasks.py: helicopter-task generation for changepoint and reversal environments.
• analysis.py, fit.py, fit_line.py: behavioral analyses and subject/model fitting.
• simulations/: simulation helpers, including Figure 6 reversal analyses.
• visualization/: plotting and optional SVG figure assembly.
• data/: source data location plus generated outputs, caches, figures, and plots.

Most users should start with run.py.

Model Architecture

The SIAS network consists of four interconnected RNN layers with distinct computational roles:

• Prefrontal cortex: Attractor states representing context
• Basal ganglia: Learning layer linking abstract states and concrete actions
• Thalamus: Implements push-pull attractor switching in PFC
• Sensorimotor cortex: Represents motor output and supervisory feedback

Learning occurs via Hebbian weight updates:

• PFC-to-BG weights learn context-action associations
• SMC-to-PFC weights (flexible model only) enable reversal learning

Task Structure

Subjects and models perform "helicopter tasks" with observations drawn from a hidden Gaussian source that undergoes changepoints:

• Latent state changes with hazard rate ~0.1
• Observations are noisy samples from current latent state
• Subjects predict next observation location
• Optimal behavior requires detecting changepoints and adjusting learning rate

Two variants of these tasks are analyzed here:

• Changepoint: Standard task with discrete state changes
• Reversal: States can return to previously visited values

Usage

Create or update the conda environment:

conda env update -f environment.yml --prune
conda activate leia

Put the required source data in:

data/source/

Then open run.py, choose which parts of the pipeline you want to run, and execute:

python run.py

Generated outputs are written under data/results/, data/cache/, data/figures/, data/paper/figs, and data/plots/.

Cython Backend

The model can run with a Cython backend for faster decision and learning phases. Build it with:

python setup_cython.py build_ext --inplace

If the compiled backend is available, model/network.py loads it automatically. If not, the code falls back to the pure Python implementation.