Towards Continual Unsupervised Data Driven Adaptive Learning

Domain Adaptation (DA) techniques are important for overcoming the domain shift between the training dataset (called source domain) and the testing dataset (called target domain). Standard DA methods assume that the entire target domain is available during adaptation, but this assumption is often violated in practice. We consider DA in a data constrained scenario, where target data become available in small batches over time, and adaptation takes place continually. Hence, continual DA is a framework to instantiate the Dynamic Data Driven Applications Systems (DDDAS) paradigm, wherein a model is developed from the data available to discern the relevant features, and subsequently when the model is deployed, it needs to be adapted (i.e., through a learning process) from the new real-world data. We discuss a novel source-free method for Continual Domain Adaptation (ConDA) that utilizes a buffer for selective replay of previously seen samples. In our unsupervised adaptation framework, we selectively mix samples from incoming batches with data stored in a buffer and use them to adapt our model as new batches are received. Our results using ConDA demonstrate the benefits of our framework when operating in data constrained environments.