Hybrid quantum-classical VAE beats SOTA topic models using only 10 qubits
First proof-of-concept shows quantum-enhanced topic modeling is viable on NISQ hardware with just 10 qubits.
A new proof-of-concept paper from Ivan Kankeu demonstrates the first hybrid classical-quantum variational autoencoder (VAE) for neural topic modeling. The architecture embeds parameterized quantum circuits within the VAE's inference network while retaining a classical topic-word decoder. To manage quantum hardware constraints, the author introduces a modified Gaussian Softmax posterior that decouples latent space dimensionality from the number of topics, allowing the model to run on a low-resource 10-qubit NISQ device.
On the AgNews dataset, the hybrid VAE outperforms state-of-the-art neural topic models, achieving a C_v coherence score of 0.71 and an NPMI score of 0.20 with high topic diversity. Notably, a fully classical variant also surpasses existing models and shows clear class separation in latent space. These results confirm that hybrid VAEs are computationally viable on near-term quantum hardware and represent a promising direction for quantum-enhanced topic modeling in NLP.
- First hybrid quantum-classical VAE for neural topic modeling, using parameterized quantum circuits in the inference network
- Achieves C_v coherence of 0.71 and NPMI of 0.20 on AgNews, outperforming classical state-of-the-art models
- Operates on a 10-qubit NISQ device thanks to a decoupled latent space via modified Gaussian Softmax posterior
Why It Matters
Enables scalable semantic discovery with NISQ-era quantum devices, opening new frontiers for quantum-enhanced NLP applications.