[Paper Review] Quantum Tensor Network in Machine Learning: An Application to Tiny Object Classification

Fanjie Kong, Xiao-Yang Liu+1 more|arXiv (Cornell University)|Jan 8, 2021

Computational Physics and Python Applications13 references10 citations

TL;DR

This paper proposes a 2D Multi-scale Entanglement Renormalization Ansatz (MERA) tensor network model to address tiny object classification in images, leveraging quantum spin model theory to match entanglement entropy scaling with object-to-image ratios. Experiments show the MERA model achieves competitive performance—78.4% accuracy on 64×64 NeedleMNIST and 76.0% on LIDC—demonstrating tensor networks as a viable alternative to CNNs for low signal-to-noise ratio tasks.

ABSTRACT

Tiny object classification problem exists in many machine learning applications like medical imaging or remote sensing, where the object of interest usually occupies a small region of the whole image. It is challenging to design an efficient machine learning model with respect to tiny object of interest. Current neural network structures are unable to deal with tiny object efficiently because they are mainly developed for images featured by large scale objects. However, in quantum physics, there is a great theoretical foundation guiding us to analyze the target function for image classification regarding to specific objects size ratio. In our work, we apply Tensor Networks to solve this arising tough machine learning problem. First, we summarize the previous work that connects quantum spin model to image classification and bring the theory into the scenario of tiny object classification. Second, we propose using 2D multi-scale entanglement renormalization ansatz (MERA) to classify tiny objects in image. In the end, our experimental results indicate that tensor network models are effective for tiny object classification problem and potentially will beat state-of-the-art. Our codes will be available online https://github.com/timqqt/MERA_Image_Classification.

Motivation & Objective

• To address the challenge of classifying tiny objects in images with low signal-to-noise ratio, where conventional CNNs underperform.
• To apply quantum many-body physics concepts—specifically entanglement entropy and MERA—to model function complexity in image classification.
• To develop and train an end-to-end differentiable 2D MERA tensor network for image classification, enabling efficient learning of spatial hierarchies.
• To evaluate the MERA model against CNNs and hybrid tensor network models on benchmark datasets with extreme object-to-image ratios.

Proposed method

• Formalizes image classification as a quantum spin model, where the target function corresponds to a golden quantum state in a Hilbert space isomorphic to the image function space.
• Uses entanglement entropy as a theoretical measure of function complexity, linking it to object-to-image (O2I) ratio and guiding model design.
• Applies 2D Multi-scale Entanglement Renormalization Ansatz (MERA) as a hierarchical tensor network to efficiently represent multi-scale spatial correlations in images.
• Employs standard back-propagation with automatic differentiation (e.g., TensorFlow) to train the MERA model end-to-end, with hyper-parameters tuned for convergence.
• Introduces a tensor network layer in a neural network baseline (Tensor-NN) to compare performance gains from hybrid architectures.
• Uses pixel-wise input tensors with one-dimensional representations, preserving spatial structure in the MERA architecture.

Experimental results

Research questions

• RQ1 Can tensor networks inspired by quantum many-body physics effectively model the complexity of tiny object classification tasks?
• RQ2 How does the entanglement entropy of the target function scale with object-to-image ratio, and can this be leveraged to design better models?
• RQ3 Does a 2D MERA tensor network outperform or match state-of-the-art CNNs on datasets with extremely low signal-to-noise ratios?
• RQ4 How do hyper-parameters such as learning rate and weight initialization affect the training stability and convergence of MERA models?

Key findings

• The 2D MERA model achieves 78.4% accuracy on the 64×64 NeedleMNIST dataset, outperforming the baseline CNN (76.0%) and showing strong potential for tiny object tasks.
• On the LIDC dataset with 1.32% average object-to-image ratio, the MERA model reaches 76.0% accuracy, outperforming the CNN baseline (78.0%) on the same metric.
• The hybrid Tensor-NN model, which integrates a tensor network layer into a CNN, achieves 86.0% accuracy on LIDC, surpassing the standard CNN (78.0%), indicating tensor networks can boost deep learning models.
• Hyper-parameter tuning reveals that a learning rate of 1e-5 and weight standard deviation of 0.0001 are critical for MERA convergence, highlighting sensitivity to initialization and optimization.
• Training curves show that MERA training is sensitive to optimization hyper-parameters, with poor choices leading to unstable loss and accuracy, suggesting challenges in optimizing tensor networks.
• Despite lower performance than CNNs on MNIST (90.3% vs. 98.3%), the MERA model demonstrates feasibility and scalability on low-SNR datasets, where CNNs fail.