Uncertainty Quantification in Neural Networks with Keras

Talk

Go to NumFOCUS academy page.

If we train a dog/cat classifier, and test it on a human image, what can be a right answer? How can our model let us know that it cannot produce an answer? In this talk we will cover methods that extend machine learning models to estimate their uncertainty, like Bayesian Neural Networks, and how they can be implemented in python using Keras and other frameworks.

Speaker

Matias Valdenegro-Toro

Dr. Matias Valdenegro (male) is a researcher at Interactive Machine Learning Team of the Robotics Innovation Center, German Research Center for Artificial Intelligence. He received his PhD in Electrical Engineering from Heriot-Watt University (Edinburgh, Scotland) in 2019, a M.Sc in Autonomous Systems from the Bonn-Rhein-Sieg University of Applied Sciences (Sankt Augustin, Germany) in 2014, and his Computer Engineering degree from Universidad Tecnologica Metropolitana (Santiago, Chile) in 2009.

His PhD research is on the problem of detecting Marine Debris on Forward-Looking Sonar images using Neural Networks. His research interests include Robot Perception, Reinforcement Learning, Uncertainty in Machine Learning, and their applications to Robotics.

Matias participates in the organization of the LatinX in AI series of workshops, being Visa Chair for the NeurIPS 2019 workshop, and Workshop Advisor for the ICML 2020 workshop. He participates as reviewer of many Robotics and Machine Learning conferences such as AISTATS, ICML, ICRA, and IROS, and many journals such as IEEE Access and IET Image Processing. Matias has 5 years of professional experience as a software engineer and more than 20 peer-reviewed publications.