Transfer of AURA’s gains to real-valued neural networks

Determine whether the performance gains of AURA carry over to real-valued neural networks, where the consistency variable is real and the signed-rotation component vanishes.

Background

AURA is designed for complex-valued neural networks and uses a complex-valued consistency variable to compare consecutive parameter updates. For real parameters, the consistency variable becomes real, so AURA reduces to a Dice-like test based only on sign and magnitude agreement, without the complex-plane rotation information used in the proposed complex-valued setting.

The paper reports improvements from AURA in several complex-valued benchmarks but does not establish whether those gains persist when AURA is applied to real-valued neural networks. The authors therefore identify the transfer of the observed gains to real-valued networks as unresolved.

References

For real parameters, $\zeta_{j,t}$ is real and $\Psi_{j,t}$ vanishes, so AURA reduces to a Dice-like test of sign and magnitude; whether the gains carry over to real-valued neural networks is an open question.

— AURA: Angular Update Rate Adaptation for training complex-valued neural networks  (2609.25914 - Ballini et al., 22 Sep 2026) in Section Conclusions