11:15 AM - 11:30 AM
[SSS04-08] A Pioneer Study on Using ChatGPT with In-context Learning for Earthquake Detection
Keywords:Earthquake Detection, Deep Learning, Seismic Waveforms analysis, ChatGPT, Earthquake Prediction, Large Language Models
Earthquakes are among the most costly natural disasters people have ever encountered, so earthquake prediction has become a very important and challenging task for humanity. Along with the research trend of using deep learning, the school of techniques has also been used to predict earthquakes by referring to feature vectors that are distilled from seismic waveforms. However, these methods usually require a large amount of labeled data for training.
In the recent two years, large language models (LLM) have shown powerful results in natural language processing. These models were initially designed to understand and generate fluency texts. More recently, developments have also enabled them to comprehend and create images. Such an expanded functionality allows large language models to receive and interpret image inputs and answer questions based on and about the input images. Besides, an innovation of these models is that they can easily adapt to a new task without relying on a large set of label data using the in-context learning strategy. Specifically, they perform a new task by conditioning on a few pairs of inputs and corresponding labels as a prompt, then making predictions for further inputs. During the process, there is no need to train/update any model parameters of the large language models, and the new tasks can be performed solely through inference.
Motivated by the new techniques and to the best of our knowledge, there has not yet been any research into whether large language models can understand and analyze features in seismic waveform images and whether they can classify between earthquake waveforms and noise waveforms. In this study, we take ChatGPT as an example and use the in-context learning strategy to apply and examine the capabilities and abilities of large language models on the earthquake detection task on the Taiwan central weather administration seismographic network dataset (CWASN). The experiments show that the ChatGPT with in-context learning can deliver a certain level of performance.
[Reference]
S. Mostafa Mousavi, William L. Ellsworth, Weiqiang Zhu,et al (2020) “Earthquake transformer—an attentive deep-learning model for simultaneous earthquake detection and phase picking”
https://doi.org/10.1038/s41467-020-17591-w
Muhammad Shakeel, Katsutoshi Itoyama, Kenji Nishida & Kazuhiro Nakadai (2021) “Detecting earthquakes: a novel deep learning-based approach for effective disaster response”
https://doi.org/10.1007/s10489-021-02285-7
Weiqiang Zhu, Gregory C Beroza (2018) “PhaseNet: a deep-neural-network-based seismic arrival-time picking method”
https://doi.org/10.1093/gji/ggy423
Sewon Min, Xinxi Lyu, Ari Holtzman, et al (2022) “Rethinking the Role of Demonstrations: What Makes In-Context Learning Work?”
https://doi.org/10.48550/arXiv.2202.12837
Jannik Kossen, Yarin Gal, Tom Rainforth, etal (2023) “In-Context Learning Learns Label Relationships but Is Not Conventional Learning”
https://doi.org/10.48550/arXiv.2307.12375
In the recent two years, large language models (LLM) have shown powerful results in natural language processing. These models were initially designed to understand and generate fluency texts. More recently, developments have also enabled them to comprehend and create images. Such an expanded functionality allows large language models to receive and interpret image inputs and answer questions based on and about the input images. Besides, an innovation of these models is that they can easily adapt to a new task without relying on a large set of label data using the in-context learning strategy. Specifically, they perform a new task by conditioning on a few pairs of inputs and corresponding labels as a prompt, then making predictions for further inputs. During the process, there is no need to train/update any model parameters of the large language models, and the new tasks can be performed solely through inference.
Motivated by the new techniques and to the best of our knowledge, there has not yet been any research into whether large language models can understand and analyze features in seismic waveform images and whether they can classify between earthquake waveforms and noise waveforms. In this study, we take ChatGPT as an example and use the in-context learning strategy to apply and examine the capabilities and abilities of large language models on the earthquake detection task on the Taiwan central weather administration seismographic network dataset (CWASN). The experiments show that the ChatGPT with in-context learning can deliver a certain level of performance.
[Reference]
S. Mostafa Mousavi, William L. Ellsworth, Weiqiang Zhu,et al (2020) “Earthquake transformer—an attentive deep-learning model for simultaneous earthquake detection and phase picking”
https://doi.org/10.1038/s41467-020-17591-w
Muhammad Shakeel, Katsutoshi Itoyama, Kenji Nishida & Kazuhiro Nakadai (2021) “Detecting earthquakes: a novel deep learning-based approach for effective disaster response”
https://doi.org/10.1007/s10489-021-02285-7
Weiqiang Zhu, Gregory C Beroza (2018) “PhaseNet: a deep-neural-network-based seismic arrival-time picking method”
https://doi.org/10.1093/gji/ggy423
Sewon Min, Xinxi Lyu, Ari Holtzman, et al (2022) “Rethinking the Role of Demonstrations: What Makes In-Context Learning Work?”
https://doi.org/10.48550/arXiv.2202.12837
Jannik Kossen, Yarin Gal, Tom Rainforth, etal (2023) “In-Context Learning Learns Label Relationships but Is Not Conventional Learning”
https://doi.org/10.48550/arXiv.2307.12375