Latent Functional Maps for Medical Imaging

Abstract:

Neural Networks (NNs) learn to represent high-dimensional data as elements of lower-dimensional latent spaces. Modeling the relationships between these representational spaces is an ongoing challenge. Successfully addressing this challenge could enable the reuse of representations in downstream tasks, reducing the need to retrain similar models multiple times. Recently, Fumero et al. leveraged the internal geometry of representations and proposed applying latent functional maps to align representations across distinct models, demonstrating its relevance for comparing representations. However, these kinds of approaches have not yet been explored in the context of medical imaging datasets, where aligning multimodal representa- tions could significantly enhance the effectiveness of models in medical applications. This project aims to use latent functional maps to align multimodal medical representations (e.g., text and vision). The first part of the thesis will involve a literature review on representation similarity. This will be followed by experimenting with the latent functional maps approach on a toy dataset of medical images and later applying it to real medical imaging tasks.