ContextGPT: Infusing LLMs Knowledge into Neuro-Symbolic Activity Recognition Models
Project Overview
The document explores the application of generative AI, particularly Large Language Models (LLMs), in education, focusing on a novel approach termed ContextGPT that enhances Human Activity Recognition (HAR) systems. By integrating common-sense knowledge from LLMs into Neuro-Symbolic AI frameworks, ContextGPT addresses the challenges of limited labeled data in training HAR models. The findings demonstrate that this innovative method can achieve performance levels comparable to or better than traditional logic-based systems, while also significantly decreasing the human effort required in knowledge engineering. Overall, the use of generative AI in educational contexts shows promise in improving system efficiencies and outcomes, highlighting its potential to transform how educational technologies are developed and implemented, particularly in facilitating more effective and scalable HAR systems.
Key Applications
ContextGPT: a prompt engineering approach leveraging LLMs for human activity recognition.
Context: Context-aware Human Activity Recognition (HAR) in various domains such as healthcare, well-being, and sports.
Implementation: ContextGPT transforms high-level context data into natural language descriptions for LLMs, which then outputs plausible activities consistent with the context.
Outcomes: Achieves comparable or better recognition rates in data-scarce scenarios with reduced human effort in knowledge engineering.
Challenges: Potential hallucinations and contradictions in outputs from LLMs, and the need for careful prompt engineering to ensure accurate context interpretation.
Implementation Barriers
Technical barrier
The requirement for a significant amount of labeled data to train supervised models restricts the deployment of effective HAR systems. Existing knowledge models (ontologies) require substantial human engineering effort and domain expertise to design and maintain.
Proposed Solutions: Using Neuro-Symbolic approaches to combine data-driven and knowledge-based methods to reduce the need for labeled data. ContextGPT reduces the need for complex knowledge models by utilizing LLMs to retrieve common-sense knowledge with minimal human intervention.
Project Team
Luca Arrotta
Researcher
Claudio Bettini
Researcher
Gabriele Civitarese
Researcher
Michele Fiori
Researcher
Contact Information
For information about the paper, please contact the authors.
Authors: Luca Arrotta, Claudio Bettini, Gabriele Civitarese, Michele Fiori
Source Publication: View Original PaperLink opens in a new window
Project Contact: Dr. Jianhua Yang
LLM Model Version: gpt-4o-mini-2024-07-18
Analysis Provider: Openai