Imagine you're organizing your toys. You could put them in boxes with straight lines dividing them (like a grid), but that's boring and doesn't match how toys naturally group together.
Instead, imagine each toy has a magnetic field around it that attracts similar toys. These fields create swirling, vortex-like territories where similar toys naturally gather.
The โต-product creates these "vortex territories"! Each neuron creates a swirling field around its weight vector, and inputs are classified based on which territory they fall into. It's like having smart, organic boundaries instead of boring straight lines.
Traditional linear models create unbounded, half-space partitions:
The โต-product creates localized, vortex-like territories with smooth, organic boundaries that better match natural data distributions.
In a classification setting with $K$ classes, we have $K$ weight vectors $\{\mathbf{w}_1, ..., \mathbf{w}_K\}$. The decision rule is:
This creates a Voronoi-like partition, but with smooth, curved boundaries instead of straight lines. Each region is:
The fiber bundle structure comes from viewing each class as a "fiber" over the input space, with the โต-product defining the connection.
Each neuron creates a localized "vortex" territory around its weight vector, with smooth, curved boundaries that naturally adapt to data distribution.
The weight vectors act as "prototypes" or "centers" of their territories, making the learned representations geometrically interpretable.
Localized territories prevent overfitting to distant outliers and create more robust decision boundaries that generalize better.
The fiber bundle structure provides a rich mathematical framework for understanding how NMNs organize and partition the input space.
This theorem reveals the geometric organization of NMNs. Unlike linear models that create arbitrary half-spaces, โต-product neurons create natural territories โ regions of space that "belong" to each neuron.
This organization is:
This is why NMNs learn sharp, coherent prototypes (as seen in MNIST experiments) rather than diffuse, blurry representations โ each neuron naturally organizes into a well-defined territory.