Los campos de ruido espacial permiten que una sola red neural tenga múltiples funciones

Structured noise, long treated as a nuisance, can act as a spatial router that tells different parts of a neural network when to wake up.

A paper on arXiv (2606.24588) introduces the crossing activation function, implemented at sample, statistical, and analytical levels with parameter reuse across them. That's the low-level knob. The high-level trick is a virtual noise field - an auxiliary continuous space that generates spatially structured noise fields, each activating a partially overlapping subnetwork.

One Network, Many Functions

The authors test this on one-dimensional function approximation tasks. They assign each target function to a different location in the noise field. The result: a single network can store multiple functions, and the storage capacity depends directly on the spatial arrangement of the noise fields.

Memory capacity improves when the noise fields' proximity mirrors the proximity relationships among the functions. If you cluster related functions close together in noise space, the network remembers more of them. Mismatched field structure - say, scattering unrelated functions across the same noise space - reduces effective capacity.

Noise as Topology, Not Disturbance

This flips the standard assumption that noise is something to filter or regularize against. Instead, structured noise becomes a topology-defining factor for functional subnetwork selection. The network doesn't need explicit modular architecture or gating mechanisms. The noise field itself carves out which neurons participate for which task.

Parameter reuse across the three implementation levels means you can pick the trade-off between computational cost and statistical rigor. The analytical-level implementation, for instance, could be used where gradient-based optimization needs deterministic behavior.

What This Enables

Practical takeaway: multi-task models that don't require separate heads or task-specific routing modules. Just drop in a well-structured noise field and let the network self-organize. The next step is scaling this beyond 1D toy problems - image classification or language tasks where spatial structure (e.g., patch positions, token positions) could map naturally onto the virtual noise field.

Source: Spatial Partial Functionalization of Neural Networks based on Noise Fields
Domain: arxiv.org