Researchers Propose Gradland to Study Phenomenal Experience via Neural Networks

A new preprint published on arXiv investigates the hypothesis that the structure of phenomenal experience is characterized by first-order interaction gradients and Jacobians. The study is conducted in an idealized environment called Gradland, inhabited by neural networks where physical dynamics are known and functions are mostly differentiable.

The authors introduce two metrics of Jacobian structure based on Kirchhoff complexity: effective rank and cohesion. By applying these metrics across practical examples, the framework accounts for the duration of experience over hundreds of milliseconds, the distinction between vivid and obscure perceptions, and the sensation of texture.

Furthermore, the hypothesis provides mathematical explanations for what learning feels like, distinguishes between clear and confused thoughts, models early sensory confusion, and clarifies the overall function of dense, rich experience.