Alapfogalmak
Large Language Models can enhance zero-shot object state classification by generating domain-specific knowledge and integrating it with pre-trained embeddings, leading to state-of-the-art performance.
Kivonat
Abstract: Investigates using Large Language Models (LLMs) for domain-specific information in Vision-based Zero-shot Object State Classification.
Introduction: Discusses limitations of deep learning methods and the potential of general-purpose semantic representations.
Methodology: Outlines stages from LLM prompting to GNN training for embedding projection.
Experimental Evaluation: Compares the proposed approach with existing VLMs and achieves superior performance.
Ablation Study: Analyzes impact of corpus size, types of embeddings, fusion techniques, dimensionality reduction, and CNN models.
Conclusions and Future Work: Suggests further exploration of LLM integration in constructing KGs and fine-tuning models.
Statisztikák
The Llama2 model used has 13 billion parameters.
GloVe 6B and fastText 16B achieve best scores for general-purpose embeddings.
GloVe is the best representation for domain-specific embeddings generated by LLM.
Idézetek
"Domain-specific knowledge can mitigate limitations in AI tasks."
"Integration of LLM-based embeddings leads to substantial performance improvements."