Quantum Neural Networks Preserve Plasticity in Continual Learning

This repository contains the official implementation for the paper Intrinsic preservation of plasticity in continual quantum learning.

In this work, we demonstrate that standard deep learning methods suffer from a fundamental "loss of plasticity" in continual learning settings, wherein networks gradually lose their ability to learn from new data. In contrast, Deep Quantum Neural Networks (QNNs) naturally overcome this limitation due to the intrinsic geometric constraints of their unitary parameter manifold. We validate this advantage systematically across four diverse experimental settings ranging from classical computer vision to deep reinforcement learning and quantum-native tasks.

📂 Project Structure

The codebase is organized into four independent scripts in /src, each corresponding to a major experiment in the paper:

• mnist.py: Supervised continual learning on Permuted MNIST.
• cifar.py: Supervised continual learning on Split CIFAR-100.
• rl.py: Deep Reinforcement Learning on MuJoCo Ant-v4.
• qdata.py: Continual learning on Quantum-Native Data.

🛠️ Installation & Requirements

This project relies on TensorCircuit-NG for quantum simulation, JAX and TensorFlow for automatic differentiation/training, and Stable-Baselines3 for reinforcement learning. Specifically, the scale of experiments in this work is only possible with the help of high performance TensorCircuit-NG.

Prerequisites

• Python 3.10+
• CUDA-enabled GPU (Recommended for Deep QNN simulation)

Setup

1. Clone the repository:

   git clone https://github.com/sxzgroup/quantum-plasticity.git
   cd quantum-plasticity

2. Install dependencies:

   pip install -r requirements.txt
   # the below is only needed for the RL experiment
   pip install "stable-baselines3[extra]" "gymnasium[mujoco]"

🚀 Usage & Experiments

Each script is self-contained including data preparation and training for both classical and quantum models. Below are the details for reproducing the results reported in the paper. The results will be automatically saved to .npz files.

1. Permuted MNIST (Supervised)

Compares a Classical MLP against a QNN on 1,000 sequential tasks where pixels are randomly permuted.

• Model: MLP vs. Deep QNN (Hardware-Efficient Ansatz).
• Metrics: Test Accuracy, Weight Norms, Gradient Norms.

python mnist.py

Output: Saves classical_model_results.npz and quantum_model_results.npz containing accuracy and norm trajectories.

2. Split CIFAR-100

Evaluates scalability on 3,000 binary classification tasks derived from CIFAR-100. Includes Fisher Information Matrix (FIM) analysis.

• Model: Deep QNN (up to 30 layers) vs. Wide MLP.
• Key Metric: Trace of the Fisher Information Matrix calculated on a fixed probe dataset to measure effective learnability.

python cifar.py

Output: Saves results including FIM traces to classical_cifar100_results.npz and quantum_cifar100_results.npz.

3. Deep Reinforcement Learning (Ant-v4)

A challenging continuous control task where the agent must adapt to a composite reward function (velocity, survival, control cost).

• Agent: Proximal Policy Optimization (PPO).
• Comparison: Standard MLP Policy vs. Hybrid Quantum-PPO Policy.
• Environment: Ant-v4 (MuJoCo).

You nedd install pip install "stable-baselines3[extra]" "gymnasium[mujoco]" for this RL task, better in a separate environment.

python rl.py

Output: Saves training rewards in .npy files and logs.

4. Quantum-Native Data

Binary classification of many-body eigenstates generated from a 1D Heisenberg XXZ Hamiltonian with varying anisotropy $\Delta$.

• Data: 2,000 tasks generated from quantum phase transitions.
• Finding: Demonstrates "Dual Advantage" — QNNs learn better (higher accuracy) and longer (no plasticity loss) than classical models on quantum data.

python qdata.py

Output: Generates dataset cache eigenstate_dataset.pkl (first run) and saves validation accuracies to .npy files.

---

📜 Citation

If you find this code useful in your research, please consider citing our paper:

@article{plasticity2025,
  title={Intrinsic preservation of plasticity in continual quantum learning},
  author={Shi-Xin Zhang and Yu-Qin Chen},
  journal={arXiv preprint arXiv:2511.17228},
  year={2025}
}

📄 License

This project is licensed under the Apache License 2.0 - see the LICENSE file.