Optimal transport in data-sciences: why and how.

The interest for optimal transport theory in the fields of data science, machine learning and statistics has grown tremendously in recent years. This talk will first discuss why that theory, which provides an intuitive and geometric perspective to compare probability distributions, is a natural alternative to more classical information divergences such as the Kullback-Leibler geometry. I will then show how some computational and statistical challenges posed by the original formulations of optimal transport, dating back to Monge (1782) and Kantorovich (1942), can be addressed using regularization. I will finally conclude by showcasing the scalability of novel new statistical methods built upon optimal transport, with varied applications in imaging, signal / natural language processing and image tagging.
  • Optimal transport in data-sciences: why and how.
  • 2018-06-15T12:00:00-03:00
  • 2018-06-15T13:00:00-03:00
  • The interest for optimal transport theory in the fields of data science, machine learning and statistics has grown tremendously in recent years. This talk will first discuss why that theory, which provides an intuitive and geometric perspective to compare probability distributions, is a natural alternative to more classical information divergences such as the Kullback-Leibler geometry. I will then show how some computational and statistical challenges posed by the original formulations of optimal transport, dating back to Monge (1782) and Kantorovich (1942), can be addressed using regularization. I will finally conclude by showcasing the scalability of novel new statistical methods built upon optimal transport, with varied applications in imaging, signal / natural language processing and image tagging.
  • Cuándo 15/06/2018 de 12:00 a 13:00 (America/Montevideo / UTC-300)
  • Dónde Salón de Seminarios. Centro de Matemática
  • Nombre
  • Speaker Marco Cuturi
  • Agregar evento al calendario iCal