Vision-language models (VLMs) exhibit impressive zeroshot transfer, but remain static and cannot adapt when exposed to new tasks. Meanwhile, conventional continual learning often overlooks preserving this zero-shot capability during adaptation. In this work, we present DREAM, a parameter-efficient framework that enables continual adaptation of VLMs while minimizing forgetting and preserving zero-shot performance. DREAM employs a dynamic prompt system with lightweight, task-specific parameters managed by two modules: a prompt composition module that dynamically generates prompts to adapt the VLM, and a query-key module that uses learned token weights to reliably activate the appropriate parameters at inference. To enhance robustness, we propose GuidedMix, which creates semantically meaningful mixed images, and pair them with mixture-aware text embeddings to strengthen representation learning through image-text alignment. We further leverage the GuidedMix samples to estimate task-specific query-key similarity thresholds that identify samples of unseen tasks and and prevent spurious prompt usage on the VLM, thereby safeguarding its zero-shot behavior. Experiments show that our method adapts efficiently, mitigates forgetting, and maintains strong zero-shot transfer with substantially fewer trainable parameters, showing consistent gains even under partial supervision.

Continual learning is essential for adapting models to new tasks while retaining previously acquired knowledge. While existing approaches predominantly focus on uni-modal data, multi-modal learning offers substantial benefits by utilizing diverse sensory inputs, akin to human perception. However, multi-modal continual learning presents additional challenges, as the model must effectively integrate new information from various modalities while preventing catastrophic forgetting. In this work, we propose a pre-trained model-based framework for multi-modal continual learning. Our framework includes a novel cross-modality adapter with a mixture-of-experts structure to facilitate effective integration of multi-modal information across tasks. We also introduce a representation alignment loss that fosters learning of robust multi-modal representations, and regularize relationships between learned representations to preserve knowledge from previous tasks. Experiments on several multi-modal datasets demonstrate that our approach consistently outperforms baselines in both class-incremental and domain-incremental learning, achieving higher accuracy and reduced forgetting.

Class-incremental continual learning is a core step towards developing artificial intelligence systems that can continuously adapt to changes in the environment by learning new concepts without forgetting those previously learned. This is especially needed in the medical domain where continually learning from new incoming data is required to classify an expanded set of diseases. In this work, we focus on how old knowledge can be leveraged to learn new classes without catastrophic forgetting. We propose a framework that comprises of two main components: (1) a dynamic architecture with expanding representations to preserve previously learned features and accommodate new features; and (2) a training procedure alternating between two objectives to balance the learning of new features while maintaining the model’s performance on old classes. Experiment results on multiple medical datasets show that our solution is able to achieve superior performance over state-of-the-art baselines in terms of class accuracy and forgetting.

OBJECTIVE: To develop a deep learning algorithm (DLA) to detect diabetic kideny disease (DKD) from retinal photographs of patients with diabetes, and evaluate performance in multiethnic populations. MATERIALS AND METHODS: We trained 3 models: (1) image-only; (2) risk factor (RF)-only multivariable logistic regression (LR) model adjusted for age, sex, ethnicity, diabetes duration, HbA1c, systolic blood pressure; (3) hybrid multivariable LR model combining RF data and standardized z-scores from image-only model. Data from Singapore Integrated Diabetic Retinopathy Program (SiDRP) were used to develop (6066 participants with diabetes, primary-care-based) and internally validate (5-fold cross-validation) the models. External testing on 2 independent datasets: (1) Singapore Epidemiology of Eye Diseases (SEED) study (1885 participants with diabetes, population-based); (2) Singapore Macroangiopathy and Microvascular Reactivity in Type 2 Diabetes (SMART2D) (439 participants with diabetes, cross-sectional) in Singapore. Supplementary external testing on 2 Caucasian cohorts: (3) Australian Eye and Heart Study (AHES) (460 participants with diabetes, cross-sectional) and (4) Northern Ireland Cohort for the Longitudinal Study of Ageing (NICOLA) (265 participants with diabetes, cross-sectional). RESULTS: In SiDRP validation, area under the curve (AUC) was 0.826(95% CI 0.818-0.833) for image-only, 0.847(0.840-0.854) for RF-only, and 0.866(0.859-0.872) for hybrid. Estimates with SEED were 0.764(0.743-0.785) for image-only, 0.802(0.783-0.822) for RF-only, and 0.828(0.810-0.846) for hybrid. In SMART2D, AUC was 0.726(0.686-0.765) for image-only, 0.701(0.660-0.741) in RF-only, 0.761(0.724-0.797) for hybrid. DISCUSSION AND CONCLUSION: There is potential for DLA using retinal images as a screening adjunct for DKD among individuals with diabetes. This can value-add to existing DLA systems which diagnose diabetic retinopathy from retinal images, facilitating primary screening for DKD.