This group of PyTorch experiments uses directional relationships to represent or constrain neural state. It contains three related questions, with different models and evaluation tasks.
Graph states
The initial prototypes give graph nodes local directions and learn rotations between them. Variants add spatial structure, persistent latent layouts, or completion of partially observed landmark graphs. These constructions make the relationship between local frames and the global output explicit.
Frequency-domain inputs
A separate experiment compares compact amplitude and phase representations on controlled signals and Fashion-MNIST. Ordinary MLPs provide a way to ask whether the input representation helps without attributing the result to predictive-coding dynamics. The phase models use exact gradients through their learned phases.
Local relational updates
Another track predicts phase relationships between embeddings, then compares detached local updates with global training and combinations of the two. The aim is to separate representation geometry from the learning rule.
These experiments do not establish a general solution to global credit assignment. The useful comparisons change one mechanism at a time and check whether information survives, transfers, or collapses.