Skip to content

Benchmarking & Validation

Jgocunha edited this page Jun 4, 2026 · 1 revision

Jgocunha/dynamic-field-theory-software compares dynamic-neural-field-composer against three other Dynamic Field Theory frameworks (Cedar, Cosivina, cosivina-python) on two axes: runtime performance and numerical correctness. dnf-composer is the fastest implementation tested at every problem size, and passes all algebraic-equivalence and behavioural-reliability checks.


Throughput

Each benchmark creates N independent neural fields and measures wall-clock simulation steps per second (median of 3 runs × 5 000 steps). The relevant comparison is against Cedar, the next-fastest framework and the only other C++ implementation:

N dnf-composer Cedar dnf-composer gain
10 8 190 5 798 +41%
50 1 601 1 240 +29%
100 794 616 +29%
500 142 118 +20%
1000 70 61 +15%

dnf-composer is fastest at every problem size, delivering ~15–41% higher throughput than Cedar — while running in full float64 precision (Cedar uses float32). The MATLAB (Cosivina) and Python (cosivina-python) implementations are several times slower than both C++ frameworks.


Algebraic equivalence & behavioural reliability

The validation study runs 100 simulations across 5 architectures (detection, selection, memory, insufficient, multi-peak), each in stimulus-ON and stimulus-OFF phases, and compares the resulting activation fields across frameworks.

  • Algebraic equivalence: the Cedar-vs-dnf-composer pairs (same activation-function family) all PASS, with maximum deviation max |Δu| ≤ 1×10⁻⁴ — bounded by Cedar's float32 rounding. The Cosivina / cosivina-python sigmoid pairs also pass within float64 tolerances.
  • Behavioural reliability: 800 / 800 (100%) comparisons show qualitative agreement across all frameworks, simulation types, phases, and activation-function families — no case where a bump forms in one framework but not another.

Full data

The complete methodology, per-size statistics, cross-family deviation analysis, and figures live in the benchmark repository:

Note: the published comparison was run against dnf-composer v2.4.1.

Clone this wiki locally