MOSAIC: Modality-Specific Adaptation for Incrementa... | AI Research

MOSAIC: Modality-Specific Adaptation for Incremental Continual Learning in Parkinson’s Disease Gait Assessment
Parkinson’s disease assessment relies on a variety of sensors—such as clinical walkways, depth sensors, and wearable devices—to capture complex motor symptoms. However, clinical systems rarely collect data from all these sensors at once. Because of patient privacy laws and storage limits, historical data cannot be kept to retrain models whenever a new sensor is introduced. MOSAIC is a new continual learning framework designed to solve this problem by allowing clinical AI systems to integrate new sensor types one by one without forgetting how to use the sensors learned previously.

The Challenge of "Toxic Teachers"

When a new sensor is added to an existing AI model, the system often tries to align the new data with the old model using a process called Knowledge Distillation. The researchers discovered a "Toxic Teacher" phenomenon: because the new sensor’s data is uncalibrated, it initially produces chaotic, high-entropy outputs. If the model tries to force this new data to match the old model immediately, it injects "noise" that degrades the accuracy of the entire system. MOSAIC addresses this by using a "Modality-Specific Warm-Up" stage, which stabilizes the new sensor’s data before attempting to integrate it into the shared model.

Decoupling Statistics for Better Performance

Standard AI models often force all data through the same normalization layers, which causes problems when sensors have vastly different physical characteristics (like 1D motion data versus 2D spatial data). This leads to "statistical coupling," where the model overwrites the historical knowledge of previous sensors. MOSAIC introduces a "Statistics-Decoupled MSBN" (Modality-Specific Batch Normalization) architecture. This allows the model to maintain a shared "backbone" for understanding gait while using independent settings for each sensor. This isolation prevents the sensors from interfering with one another, ensuring that the model can learn from new inputs without corrupting its existing knowledge.

Balancing Stability and Flexibility

To ensure the model remains useful as it grows, MOSAIC uses a "Plasticity Recovery" strategy. While the model needs to preserve what it has already learned, it also needs the flexibility to adapt to the unique physical properties of new sensors. The framework uses a curriculum-guided repulsive objective that prevents the model from becoming too rigid. By dynamically adjusting the balance between preserving old knowledge and allowing for new learning, the model can maintain a stable foundation while expanding its capacity to handle diverse clinical data.

Clinical Impact

The researchers tested MOSAIC on three different multimodal Parkinson’s disease gait datasets. The results show that the framework consistently improves the final diagnostic performance of the system and significantly reduces "catastrophic forgetting"—the tendency for AI to lose its ability to perform old tasks when learning new ones. By enabling a more modular and privacy-conscious approach to clinical AI, MOSAIC provides a viable path for updating diagnostic tools as medical technology evolves.