UniPrompt-CL: Sustainable Continual Learning in Medical AI with Unified Prompt Pools
Gyutae Oh, Jitae Shin
TL;DR
UniPrompt-CL introduces a prompt-based continual learning approach for medical AI that enhances stability, reduces inference costs, and adapts effectively to evolving medical data environments.
Contribution
It presents a minimally expanding unified prompt pool and a new regularization term tailored for medical continual learning, improving performance and efficiency.
Findings
Improves AvgACC by 1-3 percentage points in domain-incremental settings.
Reduces inference cost compared to existing methods.
Validates effectiveness through extensive experiments.
Abstract
Modern AI models are typically trained on static datasets, limiting their ability to continuously adapt to rapidly evolving real-world environments. While continual learning (CL) addresses this limitation, most CL methods are designed for natural images and often underperform or fail to transfer to medical data due to domain bias, institutional constraints, and subtle inter-stage boundaries. We propose UniPrompt-CL, a medical-oriented prompt-based continual learning method that improves prompt pool design via a minimally expanding unified prompt pool and a new regularization term, achieving a better stability-plasticity trade-off with lower computational cost. Across two domain-incremental learning settings, UniPrompt-CL effectively reduces inference cost while improving AvgACC by 1-3 percentage points. In addition to strong performance, extensive experiments clearly validate the…
Peer Reviews
Decision·Submitted to ICLR 2026
1. Clear motivation. The paper articulates well why CL techniques are critical for medical settings under data-sharing restrictions and how current PCL frameworks may fail in this domain. 2. The method is computationally efficient compared to dual-inference methods. 3. The method outperforms the considered PCL baselines for the considered diabetic retinopathy task.
1. The main weakness is limited empirical evaluation. The paper only conducts evaluation on one benchmark. To strengthen the empirical claims, the authors should consider evaluating the method on more tasks in the medical domain and ideally for longer sequences (more than 3 domains). [1] Table 6 and [2] DermCL contain tasks in the domain incremental learning setting, which are publicly available. 2. The main motivation of this work is for continual learning methods in medical settings. Howeve
1. A reasonable method is proposed to tackle an important problem, i.e., prompt-based continual learning in medical domain. 2. The experimental results show clear performance improvements over several prompt-based CL methods. 3. The writing is generally well.
1. The paper states that standard Prompt-based Continual Learning methods fail on medical data because they are designed for the broad feature space of natural images, whereas medical images require more fine-grained distinctions. However, this claim is supported by limited and largely qualitative evidence, primarily a single t-SNE visualization. The analysis lacks depth and fails to investigate the fundamental reasons for this performance gap. 2. The comparison lacks other CL families (e.g.,
1. **Addresses a Significant Problem with a Well-Reasoned Motivation** * The paper is grounded in a compelling and well-reasoned rationale: that medical images require a fundamentally different continual learning strategy than natural images. * The motivation is built on the clear distinction between the standardized nature and fine-grained analytical needs of medical data versus the high variability of natural images, providing a strong foundation for the work. *** 2. **Proposes a Use
1. **Lack of Clarity in Methodology Compromises Reproducibility** * The manuscript's most significant weakness is the ambiguous and incomplete description of the training procedure, which hinders the reader's ability to understand and reproduce the work. Specifically, the procedure for training the model on each successive continual learning (CL) stage is ambiguous. * The description in Section 4.2 is particularly confusing and logically contradictory: *"we first construct and train the
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsDomain Adaptation and Few-Shot Learning · Generative Adversarial Networks and Image Synthesis · Face recognition and analysis