Skip to content

CLAUDE.md

Latest commit

 

History

History
71 lines (50 loc) · 3.23 KB

CLAUDE.md

File metadata and controls

71 lines (50 loc) · 3.23 KB

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Project Overview

This project reproduces morphogenesis simulations from Friston et al. 2015 ("Knowing one's place: a free-energy approach to pattern regulation"). 16 cells start at identical positions and self-assemble into a target morphology (head-body-tail) over 32 time steps using active inference (free energy minimization).

Goal: Understand spm12/toolbox/DEM/DEM_morphogenesis.m deeply enough to create active inference models in other domains:

  • Psychological phenomena (first-person energy dynamics)
  • Cult dynamics (unifying/self-policing forces)
  • Relationship dynamics (mutual investment patterns)
  • Startup growth (specialization, alignment as free energy minimization)

Running the Simulation

# Requires GNU Octave: brew install octave
octave --gui run_morphogenesis.m

This runs the morphogenesis demo, pauses for figure inspection, then generates morphogenesis.mp4 from saved frames.

Architecture

Entry Point

  • run_morphogenesis.m - Sets up paths and runs DEM_morphogenesis

Core Algorithm

  • spm12/toolbox/DEM/DEM_morphogenesis.m - Main morphogenesis simulation
    • Defines target morphology as a 2D grid with cell types (0-4 encoding RGB signals)
    • Sets up generative process (G) and generative model (M) hierarchies
    • Calls spm_ADEM() to run active inference

Key SPM12 Components

  • spm12/spm_ADEM.m - Active Dynamic Expectation Maximization solver. Integrates generative process and model inversion in parallel, enabling active inference with action variables
  • spm12/toolbox/DEM/ - DEM toolbox containing demos and supporting functions
  • spm12/spm_softmax.m, spm_vec.m, spm_unvec.m, spm_detrend.m - Utility functions used throughout

DEM Structure (passed to spm_ADEM)

DEM.M  % Generative model hierarchy (recognition model)
DEM.G  % Generative process hierarchy (world model with actions)
DEM.C  % Causes (exogenous inputs)
DEM.U  % Prior expectations on causes

Model Hierarchy (M and G)

Each level specifies:

  • .g - Observation function: y = g(x,v,P)
  • .f - State transition: dx/dt = f(x,v,P)
  • .V - Precision (noise)
  • .pE - Prior expectations

Key Concepts

Active Inference: Agents minimize variational free energy - the gap between predictions/beliefs and sensory evidence. In morphogenesis, each cell infers its identity relative to others and behaves accordingly.

Identity Logits: Hidden causes representing each cell's belief about which target position it corresponds to. Transformed via softmax to get identity beliefs.

Morphogenesis Function: Computes signal concentration at each position based on distance-weighted sum of all cell signals (exponential decay with distance).

Development Workflow

Mathematical Collaboration: Use gpt-5.2 via mcp__pal__chat for mathematical topics and modeling discussions. Supports multi-turn conversations via continuation_id - pass it from one response to the next to maintain context.

Notes

  • spm12/spm_platform.m was patched to support Apple Silicon (arm64)
  • Output frames saved to morphogenesis_frames/, video to morphogenesis.mp4
  • SPM12 source: https://github.com/spm/spm12