Self-Attention in Transformer Networks Explains Monkeys' Gaze Pattern in Pac-Man Game

Abstract: We proactively direct our eyes and attention to collect information during problem solving and decision making. Understanding gaze patterns is crucial for gaining insights into the computation underlying the problem-solving process. However, there is a lack of interpretable models that can account for how the brain directs the eyes to collect information and utilize it, especially in the context of complex problem solving. In the current study, we analyzed the gaze patterns of two monkeys playing the Pac-Man game. We trained a transformer network to mimic the monkeys' gameplay and found its attention pattern captures the monkeys' eye movements. In addition, the prediction based on the transformer network's attention outperforms the human subjects' predictions. Importantly, we dissected the computation underlying the attention mechanism of the transformer network, revealing its layered structures reflecting a value-based attention component and a component that captures the interactions between Pac-Man and other game objects. Based on these findings, we built a condensed attention model that is not only as accurate as the transformer network but also fully interpretable. Our results highlight the potential of using transformer neural networks to model and understand the cognitive processes underlying complex problem solving in the brain, opening new avenues for investigating the neural basis of cognition.