Domain Adaptation for Neural Data Integration – DANDI

Developing machine learning methods suitable for data-limited systems is a core concern of artificial intelligence (AI) for health which typically suffers from sparse data. One promising approach is to leverage related data-rich systems through transfer learning; for example, by pre-training models that can be fine-tuned for clinical settings. Importantly, however, this approach can degrade performance if target domains have limited data. Human neuroimaging (e.g., functional magnetic resonance imaging [fMRI] data) is one such domain as data points are expensive to acquire and therefore commonly undersampled in clinical populations. In parallel, researchers have invested in creating large repositories of fMRI data from a small number of healthy adult participants with the goal of studying normative brain function. Leveraging these data-rich research samples to augment inferences in clinical samples would thus be highly desirable.
We propose to jointly evaluate the feasibility of this approach, leveraging our expertise in information transfer techniques and in developing models of individual fMRI datasets using AI architectures. As research datasets vary along multiple dimensions, we propose to investigate methods to overcome multiple sources of distribution shift, including in population characteristics as well as task.
We will build on existing work leveraging Optimal Transport (OT), extending to new across sample alignments. These goals entail the development of novel transfer procedures, either stemming from OT theory at large, or leveraging data structure, e.g. graph convolutional networks. As an extension of these approaches, we will also investigate aggregation through multi-source learning to improve the scalability of these techniques and enhance privacy-preservation.
Our proposed work will provide new perspectives on the challenge of domain adaptation with limited target data and create new opportunities to leverage previously unconsidered data sources.