Hokkien Low Density Language Capability

Femtosense is aimed at revolutionizing Hokkien-English translation through the innovative application of Femtosense technology. The Phase I was dedicated to a feasibility study examining the integration of sparse deep learning processors and neural networks for real-time translation tasks, with a focus on Technical Requirements such as constraint understanding, algorithm and hardware compatibility, performance, and design practicality. The study laid the groundwork for addressing translation challenges in low-resource languages by leveraging sparsity to compress translation networks for edge deployment and designing SPU chips for executing these networks efficiently. Building upon these foundations, this Phase II seeks to validate the feasibility of creating neural network models and hardware architectures optimized for low-resource translation at the edge. Key objectives include developing a prototype neural network that employs advanced quantization and compression to minimize memory requirements while maintaining accuracy, and a prototype SPU-002 hardware design capable of executing compressed language models. This phase will explore the performance of existing Hokkien translation models in uncompressed forms and aim to validate compression techniques for edge deployment feasibility.