Scalability of TPR-Attention to Complex Domains

Determine how TPR-Attention scales to higher-dimensional or noisier domains beyond the controlled setting with a small number of structured latent factors.

Background

The experiments evaluate TPR-Attention using manually provided latent factors derived from the dSprites dataset and focus on a controlled composition task with a relatively small number of structured factors. This design isolates the behavior of the architectural mechanism but does not establish its robustness or computational practicality in more complex environments.

The paper explicitly leaves unresolved whether the observed combinatorial-generalization benefits extend to domains with substantially higher dimensionality or greater noise, which are important conditions for applying the method to realistic data.