Research overview | cazaudorian

My scientific background is in fundamental physics, audio signal processing and general acoustic engineering, with which I have more recently combined statistical modeling and machine learning methods. My research can be summarized as the development of original physics-based machine learning methods to answer concrete questions from different acoustic-using communities, including oceanography, bioacoustics and music. Following are some details on my profile.

Oceanography

______________

I have been working on various ocean sciences. In marine meteorology, I have developed passive acoustic systems to estimate above surface wind speed (Cazau et al., IEEE, 2018; Cazau et al., JAOT, 2017) using mobile observation platforms (Slocum gliders, ARGO profilers, bio-logged Southern Elephant Seals). In physical oceanography, I have worked on the estimation of sea state conditions combining accelerometer and magnetometer data of opportunistic ocean observation platforms (Cazau et al., Oceanography, 2017). In marine bioacoustics, I have studied the vocal repertoires of humpback whales with a production-based approach that aims to develop computational models of vocal production mechanisms to better understanding the motivational factors behind certain vocal features (Cazau et al., ScientificReports, 2016; Cazau et al., JASA, 2015).

Machine learning models

______________

Currently at IMT Atlantique, with my supervisor Ronan Fablet, I have been working on multimodal deep learning approaches for Ocean observation, with the idea of learning multimodal representations jointly from different sensor types (satellite, hydrophones) and signal modalities (image, sound). During my Phd, I was mainly interested in Bayesian statistical methods, with a special focus on latent class models that attempt to explain observations as having been drawn from a set of latent classes, each with its own distribution, like Probabilistic Latent Component Analysis (PLCA) methods (Cazau et al., AIAAS, 2016; Cazau et al., JASA, 2015).

Big ocean data processing

______________

Within the project Ocean Data Explorer, we are developing innovative technological tools to process large volume of ocean observation data in context (i.e. associated with their metadata and explanatory auxiliary data) (Nguyen et al., 2019). So far, we have developed a backend for intensive feature computations using the distributed frameworks Hadoop/Spark deployed on the infrastructure DATARMOR of IFREMER. We have also developed a collaborative web annotation frontend. All our tools are made open source and validated by the scientific community within the project OSmOSE (Open Sciences meets Ocean Sound Explorers).