Quantum continual learning of quantum data realizing knowledge backward transfer

TL;DR

The paper tackles catastrophic forgetting in sequential learning on quantum data and proposes a quantum continual-learning framework that combines gradient episodic memory (GEM) with a variational quantum classifier. The method preserves prior-task performance during sequential training on six quantum-state classification tasks and enables positive backward transfer, outperforming elastic weight consolidation (EWC) and plain training. Key contributions include a GEM-based training routine tailored to quantum classifiers, a hardware-efficient variational ansatz, and demonstrations of improved accuracy and backward transfer across multiple task orders. This work advances quantum continual learning by showing practical, knowledge-forward and backward transfer benefits on quantum data, suggesting pathways toward more robust quantum AI systems.

Abstract

For the goal of strong artificial intelligence that can mimic human-level intelligence, AI systems would have the ability to adapt to ever-changing scenarios and learn new knowledge continuously without forgetting previously acquired knowledge. When a machine learning model is consecutively trained on multiple tasks that come in sequence, its performance on previously learned tasks may drop dramatically during the learning process of the newly seen task. To avoid this phenomenon termed catastrophic forgetting, continual learning, also known as lifelong learning, has been proposed and become one of the most up-to-date research areas of machine learning. As quantum machine learning blossoms in recent years, it is interesting to develop quantum continual learning. This paper focuses on the case of quantum models for quantum data where the computation model and the data to be processed are both quantum. The gradient episodic memory method is incorporated to design a quantum continual learning scheme that overcomes catastrophic forgetting and realizes knowledge backward transfer. Specifically, a sequence of quantum state classification tasks is continually learned by a variational quantum classifier whose parameters are optimized by a classical gradient-based optimizer. The gradient of the current task is projected to the closest gradient, avoiding the increase of the loss at previous tasks, but allowing the decrease. Numerical simulation results show that our scheme not only overcomes catastrophic forgetting, but also realize knowledge backward transfer, which means the classifier's performance on previous tasks is enhanced rather than compromised while learning a new task.

Figures (2)

• Figure 1: The learning curves for (a) plain training without any continual learning strategy (b) EWC method (c) GEM method. The task sequence is 234561 as shown in the legends, corresponding to the second row of Table \ref{['tab:ACCBWT']}. There are 6 epoches in each plot. In the $i$-th epoch, the classifier is trained on the $i$-th task. The testing data of a task is used to evaluate the test accuracy from the moment the training of that task begins until the training of the last task finishes.
• Figure 2: The learning curves for (a) plain training without any continual learning strategy (b) EWC method (c) GEM method. The three plots of each row correspond to the same task sequence shown on the far left. The horizontal and vertical axes represent the epoch and test accuracy respectively.