IC-473ResNet-18 lacks a clear multifractal structure while ResNet-152 shows one with irregular shifts, and a 160M diffusion model exhibits lower degree of emergence than Pythia 160M
The paper extends NeuroMFA to non-transformer architectures. ResNet-18's spectral graph lacks a regular bell-shaped structure, indicating no clear multifractal structure or self-organization. ResNet-152 displays a noticeable multifractal structure but with irregular horizontal shifts, suggesting a stable self-organized structure did not form. A 160M diffusion model (Stable Diffusion) shows a lower degree of emergence than Pythia 160M at the same parameter size, indicating a lower level of emergent abilities for the CV task compared to the NLP task.
Evidence
correlational
Caveat
The authors describe these as preliminary validation results. The exact diffusion model checkpoint (160M) is not fully specified, making independent replication difficult.