XAI-enhanced Quantum Adversarial Networks Achieve 0.27 RMSE for Galaxy Velocity Dispersion Modeling

Balancing predictive power with model transparency remains a significant challenge in modern machine learning, particularly when analysing complex astronomical data. Sathwik Narkedimilli from Télécom Paris, Institut Polytechnique de Paris, N V Saran Kumar from Oracle Financial Services Software Ltd., and Aswath Babu H from the Indian Institute of Information Technology Dharwad, alongside their colleagues, now present a new approach to velocity dispersion modelling in galaxies from the MaNGA survey. Their work introduces a novel adversarial network that combines the strengths of quantum-inspired neural networks with classical deep learning, and crucially, incorporates explainable artificial intelligence (XAI) techniques to provide insights into the model’s decision-making process. The team demonstrates that this hybrid model achieves state-of-the-art performance across multiple regression metrics, offering a pathway to more reliable and interpretable predictions in galactic astronomy and potentially extending the reach of quantum machine learning beyond its current limitations.

Quantum Deep Learning for Galaxy Velocities

This research explores the integration of quantum-inspired techniques with classical deep learning to improve galaxy velocity dispersion estimation. Researchers developed several hybrid models, including a baseline model combining a quantum neural network with an evaluator, and models leveraging quantum generative adversarial networks to improve feature representation. A further model utilized self-supervised learning techniques within a quantum framework. These models were trained and tested on a dataset of galaxies, and performance was assessed using metrics like Root Mean Squared Error, Mean Squared Error, Mean Absolute Error, and R-squared.

The study also analyzed how increasing the number of qubits impacted model performance. Techniques like Local Interpretable Model-agnostic Explanations were used to understand the model’s decision-making process. The hybrid models achieved performance comparable to, or slightly better than, traditional deep learning methods, with the baseline model demonstrating strong results. Increasing the number of qubits generally improved performance, although benefits plateaued after a certain point. This research demonstrates the potential of integrating quantum-inspired techniques with classical deep learning for galaxy velocity dispersion estimation, highlighting the benefits of hybrid approaches and opening avenues for future research.

Quantum Machine Learning Predicts Galaxy Dynamics

Scientists developed a novel machine learning model that integrates quantum-inspired techniques with classical deep learning to predict galaxy velocity dispersion. The study harnessed a dataset of galaxies, comprising eleven features, to train and evaluate the performance of this hybrid approach. The core of the model is a quantum adversarial network, where a generator and discriminator, both implemented using parameterized quantum circuits, work together to learn complex data distributions present in galaxy spectra. This quantum component is coupled with classical deep learning layers, allowing the system to extract enhanced feature representations from the data.

A crucial innovation lies in the inclusion of an adversarial evaluator model, which simultaneously guides the quantum network by computing a feedback loss, optimizing both prediction accuracy and interpretability. Experiments employed the adversarial evaluator to refine predictions and ensure transparency in the decision-making process. The team further explored model variations, integrating Generative Adversarial Network techniques and quantum self-supervised learning, to demonstrate the robustness and adaptability of the proposed approach. Scientists developed a hybrid quantum neural network that leverages computational resources to efficiently process complex astrophysical data. Experiments demonstrate that the baseline model achieves a Root Mean Squared Error of 0. 27, a Mean Squared Error of 0. 071, a Mean Absolute Error of 0.

21, and a coefficient of determination of 0. 59. The team implemented a quantum adversarial network utilizing parameterized quantum circuits, where the generator creates synthetic velocity dispersion patterns and the discriminator distinguishes these from real observations, effectively modeling intricate correlations in stellar motions. The quantum neural network architecture begins with fully connected layers that preprocess inputs, followed by quantum layers employing parameterized rotations across a multi-qubit system, enabling efficient gradient propagation. Furthermore, the research introduces an adversarial evaluator model, providing a unique mechanism for refining predictions and ensuring interpretability by concurrently guiding the quantum neural network with feedback loss.

Quantum Enhanced Deep Learning for Prediction

This research demonstrates the successful integration of quantum-inspired techniques with classical deep learning, yielding predictive models that balance performance and interpretability. The team developed a novel framework combining a hybrid quantum neural network with an evaluator model, achieving a Root Mean Squared Error of 0. 27, a Mean Squared Error of 0. 071, a Mean Absolute Error of 0. 21, and a coefficient of determination of 0.

1. The study validates the potential of incorporating quantum circuits into traditional neural architectures. Future work will focus on refining the integration of quantum components through advanced techniques and self-supervised learning, potentially enhancing model performance and robustness further. The researchers acknowledge that while promising, the performance gains currently remain narrow, and the lightweight model size may limit scalability for more complex, real-world problems, necessitating exploration of more scalable quantum architectures and larger datasets.