Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Lu, Hao, Koyun, Onur C., Guo, Yongxin, Zhu, Zhengjie, Alili, Abbas, Gurcan, Metin Nafi
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2602.18904
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866910029101137920
author Lu, Hao
Koyun, Onur C.
Guo, Yongxin
Zhu, Zhengjie
Alili, Abbas
Gurcan, Metin Nafi
author_facet Lu, Hao
Koyun, Onur C.
Guo, Yongxin
Zhu, Zhengjie
Alili, Abbas
Gurcan, Metin Nafi
contents Vector-quantized autoencoders deliver high-fidelity latents but suffer inherent flaws: the quantizer is non-differentiable, requires straight-through hacks, and is prone to collapse. We address these issues at the root by replacing VQ with a simple, principled, and fully differentiable alternative: an online PCA bottleneck trained via Oja's rule. The resulting model, PCA-VAE, learns an orthogonal, variance-ordered latent basis without codebooks, commitment losses, or lookup noise. Despite its simplicity, PCA-VAE exceeds VQ-GAN and SimVQ in reconstruction quality on CelebAHQ while using 10-100x fewer latent bits. It also produces naturally interpretable dimensions (e.g., pose, lighting, gender cues) without adversarial regularization or disentanglement objectives. These results suggest that PCA is a viable replacement for VQ: mathematically grounded, stable, bit-efficient, and semantically structured, offering a new direction for generative models beyond vector quantization.
format Preprint
id arxiv_https___arxiv_org_abs_2602_18904
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle PCA-VAE: Differentiable Subspace Quantization without Codebook Collapse
Lu, Hao
Koyun, Onur C.
Guo, Yongxin
Zhu, Zhengjie
Alili, Abbas
Gurcan, Metin Nafi
Machine Learning
Computer Vision and Pattern Recognition
Vector-quantized autoencoders deliver high-fidelity latents but suffer inherent flaws: the quantizer is non-differentiable, requires straight-through hacks, and is prone to collapse. We address these issues at the root by replacing VQ with a simple, principled, and fully differentiable alternative: an online PCA bottleneck trained via Oja's rule. The resulting model, PCA-VAE, learns an orthogonal, variance-ordered latent basis without codebooks, commitment losses, or lookup noise. Despite its simplicity, PCA-VAE exceeds VQ-GAN and SimVQ in reconstruction quality on CelebAHQ while using 10-100x fewer latent bits. It also produces naturally interpretable dimensions (e.g., pose, lighting, gender cues) without adversarial regularization or disentanglement objectives. These results suggest that PCA is a viable replacement for VQ: mathematically grounded, stable, bit-efficient, and semantically structured, offering a new direction for generative models beyond vector quantization.
title PCA-VAE: Differentiable Subspace Quantization without Codebook Collapse
topic Machine Learning
Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2602.18904