Domain-invariant representation learning for metal sheet bending

• Werner Zellinger
• Michael Zwick
• Thomas Grubinger

The domain-invariant training of neural networks is considered. In particular, the current neural network architecture of the TRUMPF project is investigated w.r.t. generalization performance on new domains (scenarios). This is done by means of domain adaptation optimization procedures that direct to domain-invariant representations in the neural networks parameter space. The key components of these optimization procedures are: (a) The central moment discrepancy (CMD), and (b) additional unsupervised objective terms for input-reconstruction (decoding). Method (a) was recently accepted as a conference paper written by the author as first author. In addition, some smaller modifications of commonly used batch learning optimization procedures are proposed that enable the application on the multi-domain settings of TRUMPF, e.g. domain balanced training procedures. The objective function terms (b) extend the work, proposed in the last report, for multiple domains. Additional analysis of the given dataset with regression models gives new insights in the data distribution w.r.t. regression accuracy.

The results show that significant improvements w.r.t. regression accuracy in the target domain, compared to the baseline model and elastic net regression, can be obtained by (b) and method (a) can be used to improve the regression accuracy in the source domains.