Root / CYBERNETICS AND PHYSICS / Volume 15, 2026, Number 1 / Neural classifier inversion: decoding latent visual semantics

Neural classifier inversion: decoding latent visual semantics

Sergey Salishev, Elina Vladimirova

This paper presents semantic inversion of pretrained image classifiers as an experimental probe of the view that late neural representations behave as class-
dependent error-correcting codewords. The method reconstructs ImageNet validation images from late classifier activations by optimizing pixels to match one tail-proximal feature distribution under a symmetric Kullback–Leibler (Jeffreys) objective. The protocol is not a generative-prior attack and not a data-free synthesis method: it uses only the frozen classifier, a smoothed optimization path, a feature-mean constraint, a total-variation prior, and a class-logit KL constraint. The experiment evaluates 21 Torchvision models spanning RegNet, ResNet, EfficientNet, ConvNeXt, and ViT families on 48 class-capped high-margin validation samples per model. Reconstruction quality is measured objectively with OpenCLIP ViT-L/14 image-image retrieval metrics, feature-KL reduction in bits, and top-1 classification accuracy of the reconstructed images. The results show strong semantic recovery for ViT-B/16 and EfficientNet models, moderate recovery for RegNet models and ConvNeXt-Tiny, weak retrieval for the larger ResNet and ConvNeXt variants, and adversarial nonsemantic failures for ResNet18 and ResNet34. Qualitative reconstructions further illustrate that successful recoveries preserve class-level form and color without copying exact pixel-level appearance.
CYBERNETICS AND PHYSICS, VOL. 15, NO. 1, 2026, 85–92
https://doi.org/10.35470/2226-4116-2026-15-1-85-92

File: download