• Optimizing SVMs in the primal:
  • Chapelle 2006 - Training a Support Vector Machine in the Primal
  • Ratliff et al 2007 - (Online) Subgradient Methods for Structured Prediction
• The idea for training neural networks with SVM loss came from Ronan Collobert's 2004 PhD Thesis and this paper: Collobert 2004 which has some conceptual errors (mainly that SVM loss is not Perceptron loss).

See Wikipedia - Structured SVM.

• Learning Structural SVMs with Latent Variables Latent Structured SVM