Research Project Title:
Robustness of Neural ODE
abstract:In 2018, Chen et al. proposed a new type of machine learning model called Neural ODE. Unlike the traditional CNNs, Neural ODEs are continuous-time models which directly learn gradients instead of layer weights. Now, as Neural ODEs have become increasingly popular and are applied on noisier and noisier data, it has become important that the algorithms we develop for them are robust to potentially worst-case noise. The aim of this project is then to develop such strategies to adversarially train Neural ODEs and certifiably increase the robustness of the model. I started working on this project this spring, and I hope to continue on it for the next year to develop a practical and mathematically proven theory of robustness for the Neural ODEs.
I am a CS major with a profound interest in its many fields. Robustness is one of them, and I look to explore it to its greatest depth through this SuperUROP. On the one hand, this project involves rigorous reasoning; on the other, it gives exposure to frontier research on topics like Neural ODEs and stability. The fact that I can use my previous knowledge of CS and ML to develop something very valuable drove me to take up this project.