arXiv:2610.07196
Learning Disentangled Representations with Quantum Variational Autoencoders
Gaoyuan Wang, Jerry Tan, Mark Gerstein
Abstract
Variational autoencoders are powerful representation learning models that map complex data into low-dimensional latent spaces, enabling the discovery of interpretable and disentangled factors. Such representations can facilitate the interpretation and controllable generation of data describing complex scientific systems. Understanding how these factors are organized and encoded in latent space is therefore important for developing reliable representation learning models. Recently, quantum variational autoencoders (QVAEs) have been proposed as quantum representation models, demonstrating informative latent representations and improved latent-space occupancy through quantum regularization. However, it remains unclear whether and how QVAEs can learn disentangled and interpretable latent factors. A key challenge in investigating quantum latent factors is that a small number of qubits spans an exponentially large Hilbert space, making the notion of an individual quantum latent dimension nontrivial. Here, we investigate what constitutes an individual quantum latent dimension and whether it can encode a distinct factor. We develop theoretical insights into quantum latent dimensions and support them with empirical studies on representative synthetic problems, including MNIST variants. Across three datasets, we demonstrate that QVAEs can discover factorized and semantically interpretable latent representations, with individual qubits functioning as meaningful latent factors. These results establish a foundation for understanding quantum latent spaces and their potential for structured and interpretable representation learning.
AI Summary
The paper asks whether quantum variational autoencoders (QVAEs) can encode distinct, interpretable data factors in separate latent subsystems, and whether individual qubits provide useful latent coordinates. It studies an existing QVAE architecture with a unitary encoder, discarded trash qubits, and an inverse-unitary decoder. Training combines negative reconstruction fidelity with a $\beta$-weighted quantum Jensen–Shannon divergence between each latent state and the maximally mixed prior. Qubit-local gates are proposed heuristically to favor alignment of factors with individual qubits. Experiments amplitude-encode 64-dimensional classical inputs into six qubits, retaining two or three latent qubits. The datasets comprise downsampled MNIST digits, randomly rotated digits 3 and 7, and synthetic Gaussian spectral lines with varying position and width. Factor information is assessed through classical or quantum predictors trained on reduced latent states, supplemented by Bloch-coordinate visualizations. Metrics are averaged over nine runs. Regularization around $\beta=0.1$–$0.2$ improves digit-class prediction or rotation prediction from the most informative qubit. For spectra, position is predominantly associated with one qubit and width with the remaining two; position-prediction MSE decreases from 0.1217 to 0.0766 at $\beta=0.2$, approximately 37%. The analysis is heuristic rather than a general characterization of disentanglement. The paper notes that regularization sacrifices reconstruction quality, latent-qubit counts require empirical adjustment, and two latent qubits were insufficient for accurate spectral-width encoding. Width prediction does not improve significantly with regularization. Some reported improvements concern visual organization rather than averaged prediction error; spectral visualizations use a selected best-performing run.
Central claim may not hold: The stated spectral-generation formula makes width only an amplitude multiplier that amplitude encoding removes, leaving the reported width-learning demonstration inconsistent with the specified inputs unless the formula is incorrect. The balanced five-class MNIST experiment incorrectly treats 0.5 accuracy as random guessing instead of 0.2, undermining its conclusion that the remaining qubits contain minimal class information. Label-prediction tests on reduced states do not establish the paper's stronger claim that the joint latent representation decomposes into independent subsystems. Jensen–Shannon regularization toward the maximally mixed state is invariant under entangling latent unitaries and therefore does not itself select qubit-aligned factor encodings.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.