Skip to content

Memory Instability

Jump to bottom
anonymous-author-submissions edited this page Feb 3, 2026 · 7 revisions

Memory Instability

Task description

The Memory Instability task evaluates whether an evolved Dynamic Neural Field (DNF) architecture can maintain a working-memory representation after the external stimulus has been removed.

The desired behaviour is:

  1. An external stimulus induces a localised activation peak in the input field.

  2. The input field peak is self-stabilised only during stimulus presentation.

  3. Activity is transmitted to the output field.

  4. After stimulus offset:

    • The input field relaxes to baseline.
    • The output field maintains a self-sustained peak, representing working memory.

This task tests the ability of the system to transition from pure detection to persistent internal activation.

Highest-fitness minimal phenotype

The highest-fitness solutions again converge to a minimal topology, identical in structure to the detection task but differing in field dynamics:

  • Fields: 2 (input → output)

  • Hidden fields: 0

  • Connections: 1 excitatory feedforward coupling

  • Dynamics:

    • Input field: Mexican-hat lateral interactions (transient peak)
    • Output field: Gaussian kernel enabling self-sustained activation
phenotype-memory-wbg

The figure above shows:

  • Blue: activity during stimulus presentation
  • Red: activity after stimulus removal

Only the output field preserves the peak after stimulus offset, satisfying the working-memory criterion.

Experiment-level statistics across runs

Analysed 100 runs; 100 reached the fitness threshold (100.0% success, threshold = 0.950)

Generations to threshold (successful runs)

  • Mean: 2.43
  • Median: 2.00
  • Std: 1.23

Convergence speed

  • Mean convergence rate (fitness gain/gen): 0.2241
  • Mean fitness improvement/gen: 0.2241

Architecture (successful solutions)

  • Hidden fields (mean ± std): 0.00 ± 0.00
  • Enabled connections (mean ± std): 1.00 ± 0.00

Notable runs

label run value
Highest max fitness 2026-01-23 16h12m54s 0.962585
Lowest max fitness 2026-01-23 15h55m09s 0.950034
Fastest to threshold 2026-01-23 15h49m05s 1.0
Slowest to threshold 2026-01-23 16h28m09s 7.0
Most hidden fields 2026-01-23 15h47m59s 0.0
Most enabled connections 2026-01-23 15h47m59s 1.0

Best-run evolutionary dynamics

This section reports detailed statistics from the evolutionary run that produced the highest-fitness solution for the Memory Instability task.

Fitness statistics

Final generation: g = 2

  • Best fitness: 0.9626
  • Target fitness: 0.950
  • Average fitness: 0.5870

Overall run statistics

  • Max best fitness: 0.9626 (reached at generation 2)
  • Mean best fitness over run (AUC): 0.7986
  • Mean average fitness over run (AUC): 0.4609
  • Longest stagnation period: 0 generations
  • Target fitness first reached: generation 2 (best fitness ≈ 0.9626)

Species statistics

Final generation (g = 2)

  • Species: 4
  • Active species: 4

Across run

  • Total distinct species created: 3
  • Species extinct by final generation: 0
  • Average species per generation: 2.67
  • Average active species per generation: 2.67
  • Max active species in a generation: 4 (at g = 2)

Species lifetime & size

  • Average species lifespan: 1.33 generations
  • Longest-lived species: 0 (lifespan 2)
  • Average max members per species: 336.00
  • Average offspring per species: 333.33

Topology statistics

Final generation (g = 2)

  • Avg genome size: 2.80
  • Avg field genes: 2.00
  • Avg connection genes: 0.80

Growth over run

  • Genome size change: +0.80 (≈ +0.401 / gen)
  • Field genes change: +0.00 (≈ +0.000 / gen)
  • Connection genes change: +0.80 (≈ +0.401 / gen)

Ratios

  • Avg connections per field (final gen): 0.40

Population-level kernel usage

  • Field kernels:

    • Gaussian: 3079 (77.0%)
    • Mexican-hat: 921 (23.0%)

  • Interaction kernels:

    • Gaussian: 107 (77.5%)
    • Mexican-hat: 31 (22.5%)

Genome of the highest-performing solution

Generation: 2 Fitness: 0.9626

Field genes

Field Role Kernel type Field parameters Kernel parameters
nf 1 Input Mexican-hat h = -13.19, τ = 45.00 A_exc = 22.82, σ_exc = 23.14, A_inh = 10.61, σ_inh = 34.35, A_glob = -0.08
nf 2 Output Gaussian h = -12.65, τ = 162.45 A = 24.62, σ = 1.95, A_glob = -0.16

Interaction genes

Interaction gene From → To Kernel parameters
gk cg 1-2-0 nf 1 → nf 2 A = 28.66, σ = 6.15, A_glob = 0.00

memory

Interpretation

Compared to Detection Instability, the best Memory Instability solution:

  • Requires non-trivial speciation to explore parameter space
  • Introduces Mexican-hat dynamics to control transient vs persistent activity
  • Exhibits early but non-instantaneous convergence (generation 2)

Despite identical topology, dynamic specialisation—not structural growth—is what enables working memory.

Takeaway

Memory Instability remains consistently solvable, but requires more precise parameter tuning than pure detection:

  • Convergence is still fast, but no longer instantaneous.
  • The evolved solution introduces self-sustaining dynamics without increasing topological complexity.
  • This task marks the first transition from reactive to representational behavior.

Clone this wiki locally