Brief

Researchers have investigated how cross-language transfer learning improves Handwritten Text Recognition (HTR) for low-resource Arabic-script languages. Their studies indicate that sequence modeling, rather than just shared visual representations, is key to these improvements, especially in data-scarce scenarios. Experiments on Arabic, Urdu, and Persian datasets showed that CRNN models, which combine convolutional and sequence modeling, significantly outperformed CNN-only models when trained on multiple scripts. This suggests that contextual understanding plays a crucial role in effective transfer learning for HTR in low-resource settings. AI

Researchers have developed Convolutional Recurrent Neural Network surrogate models to simulate crystal growth dynamics. These models are trained on data from Allen-Cahn dynamics and can account for variable supersaturation levels. The study compared two architectures: one that implicitly infers supersaturation from a mini-sequence of frames, and another that takes supersaturation as an explicit input. Results indicate that explicit parameter conditioning yields the most accurate predictions, though the implicit method can achieve comparable results with larger training datasets. AI