Options
Turnpikes in Deep Learning: beyond ResNets and neural ODEs?
Publikationstyp
Conference Paper
Date Issued
2026-07
Sprache
English
Start Page
1910
End Page
1916
Citation
European Control Conference, ECC 2026
Contribution to Conference
Publisher Link
Publisher
IEEE
ISBN of container
979-8-3315-5755-3
978-3-907144-13-8
It is well known that deep learning, and in particular the training of ResNets and neural ODEs, can be formalized and analyzed from an optimal control perspective. In this work, we extend the dynamic systems and optimal control perspectives to fully connected neural networks with ReLU activations and no skip connections. By exploiting equivalence relations between ResNets and networks without skip connections, we show that the corresponding training problems exhibit the turnpike property under conditions analogous to those established for ResNets. We illustrate our findings with numerical experiments for the two-spirals dataset and MNIST.
Subjects
Arrays
Equations
Steady-state
Training
Bars
Stars
Costing
Costs
Labeling
Printing
DDC Class
620: Engineering
006.32: Neural Networks