Description

Standard deep-learning algorithms are based on a function-fitting approach that do not exploit any domain knowledge or constraints. This makes them unsuitable in applications that have limited data or require safety or stability guarantees, such as robotics. By infusing structure and physics into deep-learning algorithms, we can overcome these limitations. There are several ways to do this. For instance, we use tensorized neural networks to encode multidimensional data and higher-order correlations. We combine symbolic expressions with numerical data to learn a domain of functions and obtain strong generalization. We combine baseline controllers with learnt residual dynamics to improve landing of quadrotor drones. These instances demonstrate that building structure into ML algorithms can lead to significant gains.

Speakers

Location

1-5-1 Marunouchi, Chiyoda-ku, Tokyo
Shin-Marunouchi Building, Room 902

Schedule

April 22, 2019 at 01:00PM - 01:50PM