Transitions in event-timing responses in recurrent neural network models mirror those in the human brain

The brain exhibits transitions in responses to visual event duration and frequency through the visual processing hierarchy. Timing-tuned responses gradually emerge from monotonically increasing responses, and the properties of both response types are progressively transformed over their respective hierarchies. Here, we reveal the requirements for these neural response transitions using artificial neural networks. We find that multi-layer recurrent networks exhibit monotonic and tuned responses, even without training. In these networks, progressions of monotonic response properties between network layers resemble those between brain areas. Transitions from monotonic to tuned responses emerge after training to predict upcoming inputs. Furthermore, specifically after training on predictable sequences, progressions of tuned response properties resemble those observed in the brain. These results demonstrate that the recurrent neural computations and predictive processing inherent in sensory systems are sufficient to explain the emergence and subsequent hierarchical transformations of tuned responses to sensory event timing, without needing timing-specific processes.