Research validates energy-aware neural architecture design across 2,203 experiments, finding optimal architecture depends on task modality and energy-regularized objectives yield 5-33% training-efficiency gains.

Study evaluated energy-aware learning across 2,203 experiments spanning vision, text, neuromorphic, and physiological datasets, using 10 seeds per configuration and factorial statistical analysis.

Energy-regularized objective (L = L_CE + lambda * E(theta, x)) reduced internal activation energy to 6% of baseline at moderate lambda with no accuracy loss on MNIST; architecture alone explained negligible variance in accuracy (partial eta^2 = 0.001), while architecture × dataset interaction was large (partial eta^2 = 0.44, p < 0.001), demonstrating optimal architecture depends on task modality.

Energy-first architectures inspired by action-principle framework yielded 5-33% within-modality training-efficiency gains over conventional baselines.