Resource-Constrained Affect Modelling via Variance Regularisation Pruning

Researchers introduce Variance-Regularised Pruning (VR), a neural network pruning technique that reduces model size while maintaining robust performance across diverse users. The method balances computational efficiency with cross-participant stability in affective computing systems, achieving 80% sparsity without sacrificing reliability on the AGAIN emotion recognition dataset.

This research addresses a critical gap in model optimization for affective computing—the need to deploy emotion recognition systems on resource-constrained devices while maintaining fairness and reliability across different users. Traditional pruning techniques focus exclusively on minimizing prediction error, often overlooking how parameter removal affects system performance for specific user populations. VR fundamentally shifts this approach by treating cross-participant variance as a first-class optimization objective alongside accuracy.

The motivation stems from the growing deployment of affective systems in real-world interactive environments where computational resources are limited but user diversity is high. Adaptive games, assistive technologies, and edge devices increasingly rely on emotion recognition, yet these systems must function reliably regardless of individual differences in emotional expression. VR's framework explicitly encodes this robustness requirement into the pruning process, selecting parameters that contribute reliably to predictions across all users rather than just minimizing average error.

The experimental validation on the AGAIN dataset demonstrates significant practical value. Achieving competitive Concordance Correlation Coefficient performance at 80% sparsity without fine-tuning suggests the method can dramatically reduce computational demands—enabling deployment on mobile and embedded platforms—while preserving fairness properties. This is particularly important for assistive technologies serving diverse populations where model bias introduced during optimization could have real accessibility consequences.

Looking forward, this work establishes variance regularization as a viable design principle for robust ML systems beyond affective computing. The technique could extend to other domains requiring equitable performance across demographic groups or usage contexts, influencing how fairness-aware model compression becomes standard practice in production systems.

• →Variance-Regularised Pruning maintains model reliability across different users while achieving 80% sparsity without fine-tuning
• →The method prioritizes parameters that remain robust under user distribution differences, not just overall accuracy
• →Enables deployment of emotion recognition systems on resource-constrained devices without sacrificing fairness
• →Framework addresses growing need for computationally efficient yet reliable affective computing in interactive environments
• →Establishes variance regularization as a design principle for fairness-aware model compression across ML applications