On De-Individuated Neurons: Continuous Symmetries Enable Dynamic Topologies

Researcher George Bird has published a groundbreaking paper titled 'On De-Individuated Neurons: Continuous Symmetries Enable Dynamic Topologies' that fundamentally rethinks neural network architecture. The work introduces isotropic activation functions—a new class of symmetry-principled primitives that enable networks to dynamically grow and shrink their neuronal structure in response to task demands. This approach mathematically equates connectivity pruning with neurodegeneration while maintaining computational invariance through symmetry reparameterizations, effectively decoupling network topology from individual neuron identity.

The technical breakthrough lies in the isotropic primitives' basis independence property, which eliminates the traditional elementwise functional form of neurons. This enables a layer-wise diagonalization procedure where dense layers and convolutional kernels can be re-expressed with one-to-one neuron connectivity, revealing which connections impact functionality. The method introduces a tunable 'intrinsic length' parameter to ensure analytical invariance during structural changes. Most notably, Bird demonstrates that isotropic dense networks can achieve 50% sparsity while retaining exact functionality, offering both efficiency gains and new interpretability pathways through mechanistic analysis of network communications.