VCMM: Variance-Calibrated Momentum for Multimodal Learning
Variance-Calibrated MomentuM is proposed, which adapts gradient memory to modality-specific gradient dynamics and applies exact bias correction for the time-varying first moment, enabling adaptive gradient memory without extra network passes or explicit learning-rate scaling.
Zhong-Jing Gu, Chen-Yang Huang, Yu-Fa Feng et al.
· 0 citations