Talks@DCC por Miguel Martins, DCC-FCUP & INESC-Tec

Miguel L. Martins has both academic and industrial experience in the research and development of deep learning models, specifically for the domains of signal processing, natural language processing, and computer vision. Miguel is beginning the third year of his Computer Science PhD and has published/co-authored 5 peer-reviewed publications related to deep learning models in the context of biomedical modalities such as phonocardiograms, dermoscopy, esophagogastroduodenoendoscopy and colonoscopy. Miguel’s current focus is on discovering new generalistic priors that can enhance performance, stability, or interpretability of deep models, with a special focus on medical scenarios.

Title

"Markov-based Neural Networks and their application for Heart Sound Segmentation"

Abstract

We propose a novel approach to model signals over time in the form of a unified end-to-end framework between a statistical (Markovian) model and a Convolutional Neural Network. It is particularly useful when we know how the signal state is expected to behave over time. Such is the case for heart sounds since they evolve in a very predictable manner, and thus complex long-range time-relationships between states need not be inferred (and could potentially be spurious, likely compromising the physiological plausibility of the output).

In this talk, we will see that our Markov-based Neural Network outperforms other data-driven models for these tasks, such as the U-Net and Bi-LSTM+Attention. Moreover, due to the properties of the underlying Markov Chain, we will understand how one can improve segmentation performance in new observations by optimizing their likelihood given the model in an unsupervised fashion.