Tweek the algorithms
Develop new algorithms that read documents by creating a high-fidelity representation of documents. To train neural networks to predict such representations, create algorithmically accurate and comprehensive ground truth data by obfuscating information from real-world documents and enabling data augmentation and synthetic data generation. Use state-of-the-art techniques from representation learning, active learning, transfer learning, and regression constrained learning to develop a robust system that works on complex and messy real-world documents. Evaluate your algorithms at the level of the entire customer base and leverage swarm intelligence.