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

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Large-batch training is standard for scaling multimodal neural networks integrating diverse data types (visual, textual, physiological).
• However, large batches reduce sampling fluctuations, potentially trapping models in modality-dominant minima and causing generalization gaps.

Purpose of the Study:

• To introduce Geometric Anisotropic Noise Injection (GANI), a curvature-aware optimization framework designed to mitigate generalization gaps in large-batch multimodal learning.
• To restore the exploration dynamics of small-batch training while maintaining computational efficiency.

Main Methods:

• GANI employs an exponential moving average of gradients to approximate local curvature.
• It injects structured, anisotropic noise aligned with curvature during parameter updates.
• This approach decouples deterministic descent from stochastic geometric exploration.

Main Results:

• GANI accelerates escape from sharp, modality-specific basins and promotes flatter minima.
• Experiments show GANI reduces the generalization gap and improves convergence stability in multimodal settings.
• The method maintains robust performance against visual noise and missing textual data, outperforming standard optimizers.

Conclusions:

• GANI offers an efficient, interpretable method for robust heterogeneous data processing by linking optimization geometry and multimodal dynamics.
• The framework has potential applications in uncertainty-aware multisensory integration and brain-inspired perception.
• It enables scalable multimodal learning under challenging, noisy, or incomplete sensory conditions.