learning-developmental-scaffoldings

# Learning Developmental Scaffoldings to Guide Self-Organisation **Paper:** Learning Developmental Scaffoldings to Guide Self-Organisation **arXiv:** 2605.14998v1 (2026-05-14) **Authors:** Milton L. Montero, Elias Najarro, Jakob Schauser, Sebastian Risi **Categories:** cs.AI, eess.SY, q-bio.QM ## Problem Statement Natural systems generate complex organization through self-organisation (local interactions → global structure without blueprint). However, biological development is NOT purely self-organizing — significant information is **offloaded to initial conditions**: - Maternal morphogen gradients in early embryogenesis - Tissue-level morphogenetic pre-patterns guiding organ formation - Positional and symmetry-breaking information encoded in starting states This is analogous to a **memory-compute trade-off** in computational systems: pre-patterns store information that the self-organizing process would otherwise need to compute. **Key question:** How do pre-patterns and self-organizing dynamics interact, and what information is distributed between them? ## Approach: Joint NCA + SIREN Model ### Architecture ``` [SIREN Pre-Pattern Generator] → [Initial Condition] → [NCA Self-Organization] → [Final Pattern] (learned coordinate- (bias/seed) (local rules) (target) based pattern gen) ``` **Novel contribution:** Both components are **trained simultaneously**, allowing their interplay to be varied and measured under controlled conditions. ### Components 1. **SIREN (Coordinate-based Pattern Generator)** - Generates spatial pre-patterns from coordinate inputs - Implicitly encodes target pattern structure - Provides initial conditions (seeds/biases) for the NCA 2. **Neural Cellular Automaton (NCA)** - Self-organizing system with local interaction rules - Evolves from pre-pattern initial state - Learns rules that complement (not replace) the pre-pattern 3. **Joint Training** - Both components trained end-to-end - Loss on final pattern drives learning of both pre-pattern and NCA rules - Enables measuring information distribution between components ## Key Findings ### 1. Information-Theoretic Analysis Joint learning reveals how information is distributed between: - **Pre-pattern component**: Encodes positional/symmetry-breaking information - **Self-organizing component**: Encodes local interaction rules The trade-off between these is measurable and tunable. ### 2. Robustness Improvements Jointly learned systems are **more robust** than purely self-organizing alternatives: - Better tolerance to noise in initial conditions - More reliable convergence to target patterns - Reduced sensitivity to perturbations during development ### 3. Encoding Capacity Pre-patterns increase the **diversity of patterns** the system can generate: - Pure self-organization: limited by local rule expressivity - With pre-patterns: global structure can be pre-specified ### 4. Symmetry Breaking Effective pre-patterns provide **symmetry-breaking signals** that: - Resolve ambiguities in self-organizing dynamics - Guide development toward specific outcomes - Enable complex patterns that pure self-organization cannot achieve ### 5. Non-Trivial Pre-Pattern Structure **Critical insight:** Effective pre-patterns do NOT simply approximate their targets. Instead, they: - **Bias the developmental dynamics** in ways that facilitate convergence - Create a **non-trivial relationship** between initial condition structure and dynamics - Provide the *right kind* of perturbation, not a crude approximation ## Technical Framework ### Information-Theoretic Metrics - **Pre-pattern information**: I(pre-pattern; target) — how much target info is encoded in initial conditions - **Self-organization contribution**: I(NCA state progression; target | pre-pattern) — what dynamics add beyond the seed - **Total mutual information**: Decomposed into pre-pattern vs. dynamics contributions ### Training Objective ``` L = ||NCA(SIREN(x), t=T) - target||² ``` Both SIREN and NCA parameters are updated simultaneously to minimize this loss. ## Applications to Neuroscience ### 1. Brain Development Modeling - **Cortical column formation**: Pre-patterns could represent molecular gradients that guide cortical area specification - **Retinotopic mapping**: Initial positional biases guide self-organizing connectivity - **Critical periods**: Information offloading may explain developmental windows ### 2. Neural Circuit Development - **Axon guidance**: Morphogenetic gradients provide pre-patterns for self-organizing synapse formation - **Cell-type specification**: Initial positional information guides differentiation programs - **Network topology**: Pre-patterns may encode structural constraints on self-organizing connectivity ### 3. Neurodevelopmental Disorders - Misaligned pre-patterns could model developmental disruptions - Understanding information distribution between genetic programs and self-organization ### 4. Neural Network Architecture Design - **Inductive biases as pre-patterns**: Structured initialization as a form of information offloading - **Developmental AI**: Models that grow rather than are trained end-to-end - **Robust initialization**: Understanding why certain initializations lead to better convergence ## Comparison with Pure Self-Organization | Aspect | Pure NCA | Joint NCA + Pre-Pattern | |--------|----------|------------------------| | Robustness | Moderate | High | | Encoding capacity | Limited by local rules | Extended by global seed | | Symmetry breaking | Random/stochastic | Guided | | Convergence reliability | Variable | Consistent | | Information source | Dynamics only | Dynamics + initial conditions | ## Related Concepts - **Morphogenetic pre-patterns**: Biological gradients that guide development - **Memory-compute trade-off**: Storing information vs. computing it - **Neural Cellular Automata**: Self-organizing systems with neural local rules - **SIREN**: Sinusoidal Representation Networks for coordinate-based pattern generation - **Information offloading**: Distributing computation across initial conditions and dynamics - **Developmental biology**: Embryogenesis, morphogenesis, cell differentiation ## Implementation Considerations 1. **NCA Design**: Local rules with sufficient expressivity for target patterns 2. **SIREN Architecture**: Frequency tuning affects pattern resolution 3. **Joint Optimization**: Gradient flow through both components requires careful balancing 4. **Information Analysis**: Requires multiple runs to estimate mutual information 5. **Pattern Complexity**: Start simple (gradients, stripes) before complex targets ## Related Skills - brain-inspired-nca - neural-cellular-automata-attractors - brain-inspired-cellular-automata - neurotrain-local-learning-snn-benchmarking ## Activation Keywords - developmental scaffoldings - self-organisation - neural cellular automata - NCA - pre-patterns - morphogenetic - developmental biology - SIREN - information offloading - memory-compute trade-off - joint learning NCA - brain development modeling - morphogenesis - developmental AI ## References - arXiv: https://arxiv.org/abs/2605.14998 - PDF: https://arxiv.org/pdf/2605.14998