Energy First Arch Eval

energy-first-arch-eval

minAction.net: Energy-First Neural Architecture Design -- From Biological Principles to Systematic Validation — Frasch (2026) (arXiv:2604.24805, 2026)

What this evaluates

Evaluates the classification accuracy and training energy efficiency of biologically-inspired and physics-guided neural architectures against conventional baselines across diverse data modalities. It probes whether action-principle regularization yields modality-specific performance gains and reduced internal activation energy without accuracy loss.

Datasets

• Fashion-MNIST — total ?; splits: test (-1)
• CIFAR-10 — total ?; splits: test (-1)
• DVS Gesture — total ?; splits: test (-1)
• SHD — total ?; splits: test (-1)
• SSC — total ?; splits: test (-1)
• WESAD — total ?; splits: test (-1)
• DREAMER — total ?; splits: test (-1)
• SEED-IV — total ?; splits: test (-1)
• 20newsgroups — total ?; splits: test (-1)

Metrics

• Accuracy (primary) — range: percent
  • Proportion of correctly classified instances out of the total number of instances.
• mJ/correct — range: other
  • Total training energy in millijoules divided by the number of correct predictions.
• Activation Energy (relative) — range: other
  • Hardware-independent proxy computed as the expected L2 norm of layer activations, normalized to baseline.

Input / output format

Input: Standard dataset inputs (images, text, or neuromorphic spike trains) depending on the modality.

Output: Class predictions for each instance.

Scoring recipe

def compute_metrics(predictions, gold, energy_joules):
    correct = sum(1 for p, g in zip(predictions, gold) if p == g)
    accuracy = correct / len(gold)
    mJ_per_correct = (energy_joules * 1000) / correct if correct > 0 else float('inf')
    return {'accuracy': accuracy, 'mJ_per_correct': mJ_per_correct}

Common pitfalls

• Confusing task difficulty and convergence speed with intrinsic architectural energy efficiency.
• Averaging metrics across modalities, which obscures strong architecture-dataset interactions (partial η² = 0.439).
• Measuring only inference energy, whereas reported gains stem from training dynamics (fewer epochs).

Evidence (verbatim from paper)

Architecture significantly impacted energy efficiency (F(3,1381)=19.11, p<0.001). CNN showed the lowest energy-per-correct (7,059 mJ/correct ± 56 SE), compared to BimodalTrue (103,461 mJ/correct ± 6,465 SE), MLP (137,303 mJ/correct ± 10,303 SE), and Physics-Lagrangian (102,966 mJ/correct ± 6,087 SE). However, this apparent advantage is primarily attributable to task difficulty and faster convergence on easier vision benchmarks rather than intrinsic architectural efficiency, as detailed in the next paragraph.

Citation

@misc{frasch2026minactionnet,
  title={minAction.net: Energy-First Neural Architecture Design -- From Biological Principles to Systematic Validation},
  author={Frasch (2026)},
  year={2026},
  note={arXiv:2604.24805}
}

• arXiv: 2604.24805