UPM Advances Trustworthy and Efficient AI with Edge and LLM Research

The SMARTY Chips JU Project continues to push the boundaries of secure and efficient AI integration. Two recent research publications within the project, authored by researchers from UPM, demonstrate significant progress in two key areas: trustworthy AI reasoning and efficient edge AI deployment.

Enhanced FIWARE-Based Architecture for Cyberphysical Systems With Tiny Machine Learning and Machine Learning Operations: A Case Study on Urban Mobility Systems

One study focuses on integrating TinyML with MLOps to enhance reproducibility, transparency, and performance monitoring in edge AI systems. The proposed architecture leverages FIWARE as the enabling data technology, allowing tiny models to operate at the edge while seamlessly transitioning to a base model in the cloud when necessary. This hybrid approach ensures optimal performance while maintaining efficiency in resource-constrained environments. The research showcases a scalable and secure deployment framework that aligns with SMARTY’s mission of enabling trustworthy AI in critical applications. Full paper available at https://ieeexplore.ieee.org/document/10754992 

Can ChatGPT Learn to Count Letters?

Another publication tackles a common challenge in large language models (LLMs): the difficulty of counting letters due to tokenization mechanisms. Through a fine-tuning approach, the study successfully reduces the failure rate of GPT-4o from 15.25% to 0.75% when identifying occurrences of specific letters in words. This advancement highlights a pathway to improving LLM reasoning, reinforcing their reliability for real-world applications. By addressing fundamental AI limitations, this research contributes to SMARTY’s vision of integrating AI solutions that are both effective and trustworthy across its use cases. Full paper available at https://ieeexplore.ieee.org/document/10896907 

These studies underscore SMARTY’s commitment to pioneering secure and scalable AI technologies.