Why Neural Network Can Discover Symbolic Structures with Gradient-based Training: A Theoretic Foundation for Neurosymbolic AI with Practical Implications
Why Neural Network Can Discover Symbolic Structures with Gradient-based Training: A Theoretic Foundation for Neurosymbolic AI with Practical Implications
Zhangyang “Atlas”,Wang,Peihao Wang
0 Citations
TLDR
By interpreting network optimization as a gradient flow in measure space, it is proved that parameters concentrate onto low-dimensional, G -invariant manifolds, which endows the network with approximate ring compatibility and latent symbolic structures.