FLAIR: A Foundation Model for Grapheme Recognition in Ancient Scripts with Few-Shot Learning
Keywords: Foundation Model, Few-Shot Learning, Prototypical Networks, Encoder Network, Indus Valley Civilization Script, Omniglot Dataset
TL;DR: Updating prototypical networks and applying few-shot learning for Indus Valley script recognition with Omniglot dataset testing and validation
Abstract: The Indus Valley Civilization (IVC) left behind an undeciphered script, posing a significant challenge to archaeologists and linguists. This paper introduces FLAIR, a few-shot learning approach that aims to establish a foundational model for recognizing and identifying individual graphemes from the limited available Indus script. As a foundational model, FLAIR is designed to be versatile, supporting multiple potential applications in script recognition and beyond. It leverages prototypical networks combined with a modified proposed encoder network for segmentation, ProtoSegment to extract intricate features from the grapheme images. We evaluate FLAIR’s ability to generalize from minimal data using IVC grapheme classification tasks and further experiment with pre-trained Omniglot models for fine-tuning. Additionally, we simulate real-world data scarcity by intentionally restricting training data on the Omniglot dataset. Our experiments demonstrate FLAIR’s accuracy in digitizing and recognizing Indus Valley seal graphemes, outperforming traditional machine learning classification approaches. These results underscore FLAIR's potential not only for the digitization of ancient scripts with limited labeled datasets but also for broader applications where data is scarce. FLAIR’s success in grapheme recognition highlights its promise as a foundational model capable of extending to other undeciphered writing systems, thereby contributing to the integration of classic scientific tools and data-driven approaches.
Primary Area: foundation or frontier models, including LLMs
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 7841
Loading