Few-Shot Learning for physical performance prediction using 3D CNNs


In previous works, we have shown that we can train a 3D Convolutional Neural Network to predict aerodynamic performances of 3D CAD models directly from a surface mesh representing the object. [ICML 2018 Paper] However, this requires to collect a relatively large dataset for each new type of experiment -- different flow-speed, new simulator etc...--. In this project, we propose to use transfer learning methods inspired from works on few-shot learning in order to train Neural Networks that can learn very quickly in a new environment while leveraging on previously seen datasets.

Online Pressure and Drag prediction with GCNN.

Starting point

The student will be implementing his work on top of a large state-of-the-art code base, which includes a Deep-Learning library for 3D shapes in the TensorFlow framework.

The reference papers are the following:

Please send an email to pierre.baque@epfl.ch or timur.bagautdinov@epfl.ch

The work will be done in collaboration with the EPFL startup Neural Concept .