The long-held observation that natural images retain their recognizability from Fourier phase alone, while magnitude carries little identity, has been a curious anomaly. This paper probes whether this asymmetry holds within the hidden layers of trained AI image classifiers.
Related startups
Phase Trumps Magnitude in Identity Encoding
By causally testing this hypothesis through phase-to-magnitude transplantation experiments across PRISM2D, GFNet, and ViT-B/16, the researchers found a consistent pattern: predictions overwhelmingly followed the phase donor. Crucially, deleting image-specific magnitude information had minimal impact on accuracy, underscoring that identity primarily rides on phase. This challenges the conventional reliance on magnitude for image recognition tasks and highlights the critical role of phase in AI image classifier phase magnitude representation.
Architectural Nuances Shape Phase Expression
While ResNet-50 initially appeared to deviate, further investigation revealed that interventions before ReLU activations exposed a strong latent sign code. The study also confirmed that the readout mechanism consumes a channel-wise spatial average. These findings indicate that different architectures, while sharing a common phase/sign identity code, expose it in distinct bases. This divergence, influenced by rectification and readout geometry, offers a mechanistic explanation for the observed texture-shape performance gap between Convolutional Neural Networks (CNNs) and attention-based models. The robust encoding of identity in phase information suggests that current methods may be overlooking a fundamental aspect of AI image classifier phase magnitude representation.
