Spectral Complex Autoencoder Pruning: A Fidelity-Guided Criterion for Extreme Structured Channel Compression

We propose spectral complex autoencoder pruning (SCAP), a structured channel pruning framework that estimates channel redundancy from input-output interaction descriptors. For each convolutional layer, SCAP constructs a complex interaction field by pairing the full multichannel input activation as the real component with a single spatially aligned output-channel activation as the imaginary component. The interaction field is transformed into the frequency domain, where a low-capacity autoencoder is trained to reconstruct normalized spectra. The resulting reconstruction fidelity provides a scale-invariant cue of descriptor compressibility: easily reconstructed interaction fields tend to be more regular under the learned spectral model, whereas hard-to-reconstruct fields indicate less compressible channel-specific interactions. Since this fidelity cue does not encode the absolute magnitude of the corresponding filter, SCAP combines the fidelity-derived mismatch with a layer-normalized filter $\ell _{1}$ term to form the default channel importance score. The fused score is then used for threshold-based structured pruning followed by standard fine-tuning. On VGG16 trained on CIFAR-10, SCAP at threshold $\tau =0.6$ achieves 91.27% floating-point operation (FLOP) reduction and 96.32% parameter reduction (PR) with a 1.98% Top-1 accuracy drop from a 93.44% baseline. Empirical results on CIFAR-10, CIFAR-100, and ImageNet-1k, together with an additional DeiT-Tiny/CIFAR-100 validation, show that SCAP provides a competitive accuracy-compression trade-off under aggressive structured pruning and can be instantiated beyond conventional convolutional neural network (CNN) backbones.