Reproduce “MRN: Multiplexed Routing Network for Incremental Multilingual Text Recognition” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tianlun Zheng, Zhineng Chen, BingChen Huang, Wei Zhang, Yu-Gang Jiang
The Multiplexed Routing Network (MRN) is introduced to address Incremental Multilingual Text Recognition (IMLTR), a new task where models incrementally learn new languages without forgetting old ones. MRN achieves substantial accuracy improvements, ranging from 10.3% to 35.8%, over general-purpose incremental learning methods on MLT17 and MLT19 datasets, demonstrating superior robustness against catastrophic forgetting.
Key insights
- Decoupling language-specific feature learning and freezing parameters effectively mitigates data and class imbalance, ensuring fair treatment of all languages, including data-poor ones.
- The DM-Router's soft-voting mechanism explicitly models and utilizes inter-lingual dependencies, allowing for knowledge transfer and robust language identification rather than a hard selection of a single model.
- Directly addressing the 'rehearsal-imbalance' challenges inherent in IMLTR through a tailored architectural design is crucial for preventing catastrophic forgetting in multilingual text recognition.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.