Neural ordinary differential equations - NODEs (DS4DS 4.07)

Hosts:
Sebastian Peitz - orcid.org/0000...
Oliver Wallscheid - / wallscheid
Programming examples / Julia code via GitHub:
github.com/DS-...
Supported by the ‘digi-Fellows’ program granted by Paderborn University:
www.uni-paderb...

[@looper6394]

10:15 unstable identification -> potential fix, limit rhs output eg via tanh, but also limits the dynamical behaviour; another approach -> standardize data

[@gabrielnicolosi8706]

It is excellent to see these things being shown in Julia! Most of these tutorials are still heavily relying on Python, so this is great! You might have mentioned it somewhere else, but is there any preference for Lux over Flux? Thanks for the free content.
Gabriel

[@UPB_DS4DS-bu8ec]

Thanks for the positive feedback! The choice of Lux over Flux was mostly motivated by the closer relation of Lux to the SciML ecosystem, which we have relied on during most parts of the course series. However, with mild code modifications the examples will also work with Flux.

[@nabin6976]

Hi Do you have any reference for Neural ODE being an ANN + ODE? If I remember correctly the original neural ODE paper from NeurIPS 2018 defined NODE as the continuous limit of a residual neural net?

[@fabymike]

Hi, this is callled physics informed NN?

[@looper6394]

no bc no prior knowledge about the structure of f or any physical constraint, eg constant energy, within the loss function L is used