Adversarial Noise Isolation in Multimodal Perception: A Computational Framework Inspired by Inhibitory Control

Background: Robust perception involves processing heterogeneous sensory signals, such as facial expressions, vocal prosody, and language, particularly in noisy environments. In computational modeling, a key challenge is integrating these diverse inputs while actively filtering uninformative variations. While recent deep learning models address this integration through complex fusion architectures, they typically aggregate features without explicit filtering modules analogous to inhibitory control. In this study, we propose Multi-modal Information Disentanglement (MInD), a computational framework designed to test the hypothesis that algorithmic noise isolation facilitates robust multisensory integration. Methods: Drawing conceptual inspiration from cognitive theories of modularity, our model decomposes sensory inputs into amodal (modality-invariant) and modal-specific pathways. Furthermore, we introduce an adversarial noise isolation mechanism to serve as an algorithmic analog to cognitive inhibition. Given that our model operates on pre-extracted high-level features, this mechanism functions to isolate latent distributional variance-uninformative fluctuations that persist after initial feature extraction-guiding the network to separate task-relevant affective cues from irrelevant feature variance. Results: Empirical evaluations on standard emotion recognition benchmarks indicate that this purification-before-fusion strategy is associated with competitive performance and stability across multiple metrics. Notably, the framework attains these results using simple linear integration layers, suggesting that separating representations prior to fusion may reduce the computational complexity required for subsequent integration. Conclusions: These observations highlight the computational utility of algorithmic noise suppression, illustrating how cognitive inspiration can inform efficient machine learning architectures without claiming direct neurobiological validation.