[Paper Review] QClusformer: A Quantum Transformer-based Framework for Unsupervised Visual Clustering

Xuan-Bac Nguyen, Hoang-Quan Nguyen+4 more|arXiv (Cornell University)|May 30, 2024

Retinal Imaging and AnalysisMedicine3 citations

TL;DR

QClusformer proposes a novel Quantum Transformer-based framework for unsupervised visual clustering, leveraging parameterized quantum circuits in self-attention mechanisms to enhance clustering performance on classical data. It achieves state-of-the-art results on MS-Celeb-1M and DeepFashion benchmarks, with a 74.50% Pairwise F1-score and 82.09% BCubed F1-score, demonstrating the potential of quantum-enhanced attention in large-scale clustering tasks.

ABSTRACT

Unsupervised vision clustering, a cornerstone in computer vision, has been studied for decades, yielding significant outcomes across numerous vision tasks. However, these algorithms involve substantial computational demands when confronted with vast amounts of unlabeled data. Conversely, quantum computing holds promise in expediting unsupervised algorithms when handling large-scale databases. In this study, we introduce QClusformer, a pioneering Transformer-based framework leveraging quantum machines to tackle unsupervised vision clustering challenges. Specifically, we design the Transformer architecture, including the self-attention module and transformer blocks, from a quantum perspective to enable execution on quantum hardware. In addition, we present QClusformer, a variant based on the Transformer architecture, tailored for unsupervised vision clustering tasks. By integrating these elements into an end-to-end framework, QClusformer consistently outperforms previous methods running on classical computers. Empirical evaluations across diverse benchmarks, including MS-Celeb-1M and DeepFashion, underscore the superior performance of QClusformer compared to state-of-the-art methods.

Motivation & Objective

• Address the high computational cost of classical unsupervised clustering on large-scale unlabeled visual datasets.
• Explore the potential of quantum computing to accelerate and improve unsupervised clustering through quantum machine learning.
• Design a Quantum Transformer architecture that integrates parameterized quantum circuits into self-attention and transformer blocks for end-to-end clustering.
• Demonstrate the feasibility and superiority of quantum-enhanced clustering on real-world vision benchmarks despite current NISQ hardware limitations.
• Establish a framework that leverages quantum superposition and entanglement to improve clustering accuracy and robustness on noisy or complex data.

Proposed method

• Design a Quantum Transformer architecture where the self-attention mechanism is reimagined using parameterized quantum circuits (PQCs) for query, key, and value computation.
• Implement separate PQCs for query, key, and value vectors (1Q-1K-1V), or shared PQCs (1QKV), to enable quantum state encoding and unitary evolution in attention computation.
• Utilize Pauli matrix measurements to extract classical information from quantum states, minimizing qubit and circuit depth usage while preserving clustering-relevant features.
• Train the end-to-end QClusformer framework using classical backpropagation on feature embeddings extracted via a pre-trained classical CNN, with quantum circuits optimized via parameterized gradients.
• Integrate k-nearest neighbors (k-NN) for initial clustering and refine predictions using the quantum-attention mechanism to improve cluster assignment.
• Simulate the QClusformer on NISQ-era quantum hardware with T=1 Transformer block due to hardware constraints, ensuring fair comparison with classical baselines.

Experimental results

Research questions

• RQ1 Can a Quantum Transformer architecture with parameterized quantum circuits outperform classical Transformer-based clustering models in unsupervised visual clustering?
• RQ2 How does the integration of quantum self-attention mechanisms improve clustering performance on large-scale, real-world vision datasets like MS-Celeb-1M and DeepFashion?
• RQ3 What is the impact of different PQC configurations (e.g., 1Q-1K-1V vs. 1QKV) on clustering accuracy and resource efficiency in the QClusformer framework?
• RQ4 How does QClusformer compare to prior quantum clustering methods such as QKMeans in terms of F1-score and robustness across varying numbers of clusters?
• RQ5 To what extent can quantum computing principles like superposition and entanglement enhance the representation learning and clustering capability of attention mechanisms in vision tasks?

Key findings

• On the 584K test split of MS-Celeb-1M, QClusformer achieved a Pairwise F1-score of 74.50% and BCubed F1-score of 82.09%, surpassing the classical Clustering Transformer (63.31% and 79.74%).
• On the DeepFashion dataset, QClusformer achieved a Pairwise F1-score of 35.71% and BCubed F1-score of 60.00%, outperforming classical methods like K-means and DBSCAN while remaining competitive with the classical Clusformer.
• In the 100-class setting, QClusformer improved F1-scores from 84.40% (QKMeans) to 85.36% (Pairwise) and from 88.91% to 91.15% (BCubed), demonstrating consistent gains over prior quantum clustering methods.
• In the 1000-class setting, QClusformer achieved a 54.70% Pairwise F1-score and 74.93% BCubed F1-score, significantly outperforming QKMeans (48.84% and 71.95%) under the same conditions.
• The 1QKV configuration (one PQC for query, key, and value) achieved the highest BCubed F1-score of 60.10%, indicating that shared quantum circuits can preserve clustering performance while reducing resource overhead.
• Ablation studies confirmed that the 1Q-1K-1V setting yielded the best Pairwise F1-score of 35.71%, highlighting the benefit of independent quantum processing for query, key, and value in attention computation.