Core Concepts
Large pre-trained models, known as foundation models (FMs), have the potential to revolutionize smart agriculture by offering versatile capabilities with minimal fine-tuning.
Abstract
The rapid development of machine learning and deep learning methodologies in agriculture has led to significant advancements in smart crop management, plant breeding, livestock farming, aquaculture, and agricultural robotics. Traditional ML/DL models face limitations such as reliance on labeled datasets, specialized expertise, and lack of generalizability. Large pre-trained models or foundation models (FMs) have shown success across various domains by training on vast data from multiple sources. These FMs can perform diverse tasks with minor fine-tuning and minimal labeled data. Despite their effectiveness, there is limited exploration of applying FMs in agriculture AI. This study aims to explore the potential of FMs in smart agriculture by categorizing them into language FMs, vision FMs, multimodal FMs, and reinforcement learning FMs.
Stats
BERTBASE and BERTLARGE contain 110M and 340M parameters respectively.
PaLM has 540B parameters designed for efficient training.
GPT-3 has 175 billion parameters.
SAM trained on over 1 billion masks on 11M images for segmentation tasks.
GigaGAN achieves a zero-shot Fréchet Inception Distance (FID) of 9.09 on COCO2014 dataset.
Quotes
"Large pre-trained models or foundation models (FMs) have shown success across various domains by training on vast data from multiple sources."
"FMs can perform diverse tasks with minor fine-tuning and minimal labeled data."
"Despite their effectiveness, there is limited exploration of applying FMs in agriculture AI."