Pulvinar pathways as skip connections in deep neural networks for vision

The pulvinar is thought to regulate cortico-cortical communication through transthalamic pathways, yet the computational consequences of such modulation remain unclear.

A minimal pulvinar-inspired long-range skip pathway in a hierarchical convolutional vision model was tested, implemented as a learned projection from early to late feature maps coupled with a gain-controlled gating mechanism. The architecture was evaluated in two complementary experiments: CIFAR-10 categorization and a near-threshold contrast-detection task with noisy backgrounds.

In CIFAR-10 categorization, the pulvinar-augmented model did not primarily act as a uniform accuracy booster; instead, it reshaped contrast-dependent response scaling and stabilized internal representations, consistent with gain control. These effects persisted under matched representational regimes, indicating that performance differences cannot be explained by changes in representational geometry alone. In the near-threshold detection task, pulvinar-like modulation systematically redistributed detection outcomes, consistent with shifts in decision criterion under uncertainty, and in specific regimes supported improved discriminability.

Together, these findings indicate that pulvinar-inspired connectivity contributes primarily to gain and decision-criterion control. We propose that the pulvinar regulates how sensory evidence is weighted and converted into decisions, rather than directly enhancing task accuracy.