Decoding Covert Human Attention in Multidimensional Environments

Abstract

In complex environments, available information does not uniquely define state, requiring attention learning to identify features relevant for learning and decision-making. As a result, human decisions often reflect reasoning that cannot be directly observed from choice. This dual opacity, at the level of agent and observer, poses a fundamental challenge for understanding naturalistic behavior. We inferred latent attention during learning and decision-making by training recurrent neural networks on synthetic data generated from two classes of attention learning models: feature-based reinforcement learning (FRL), in which attention emerges through retrospective value updating, and serial hypothesis testing (SHT), in which discrete hypotheses are prospectively sampled. A network trained on hybrid (FRL+SHT) synthetic data outperformed single-model networks, decoding latent human attention with more than 80% accuracy. This work provides a new approach for decoding latent attention and suggests a mechanism of attention learning wherein value-derived hypotheses are continuously tested against incoming evidence.