ODE Solvers and Experimentations with pytorch
- 0 Collaborators
The primary idea behind Neural Ordinary Differential Equations is that certain types of neural networks are analogous to discretized differential equations. These networks can be thought of as dynamical systems with each layer corresponding to propagation by a single time step. ...learn more
Project status: Under Development
Groups
DeepLearning
Intel Technologies
MKL,
Intel Opt ML/DL Framework
Overview / Usage
ODE networks are much more memory efficient, with fewer parameters and backpropagation being more efficient. These networks are adaptive in the sense that they have a tolerance level which controls the measure of performance at the cost of more function evaluations. Overall, ODE networks open up a really exciting field of research that emphasizes more on a continuous and normalized flow instead of a discrete structure.
Methodology / Approach
The authors of the original paper suggested the use of adjoint method over Euler or Higher-order variants for solving the ODE.
Technologies Used
Intel distribution for python; Intel MKL; Pytorch
Repository
https://github.com/Konsang/Neural-ODE (Locked | Under Development)