Curvature-guided anisotropic noise injection for robust multimodal data processing in neuroscience and perception science

Introduction Large-batch training is widely used to scale multimodal neural networks that integrate heterogeneous inputs such as visual, textual, and physiological signals. However, increasing the batch size suppresses the stochastic fluctuations of mini-batch sampling, which can trap multimodal models in sharp, modality-dominant minima and produce a persistent generalization gap. Methods To address this problem, we propose Geometric Anisotropic Noise Injection (GANI), a curvature-aware optimization framework inspired by information geometry and multisensory integration. GANI decouples deterministic large-batch descent from stochastic geometric exploration. It approximates local curvature through an exponential moving average of first-order gradients and injects structured anisotropic noise during parameter updates, thereby restoring the geometry-aware exploration dynamics of small-batch stochastic gradient descent with linear computational complexity.