A new research paper questions the efficacy of current foundation models (FMs) in seismology, arguing they function more as glorified feature extractors than true FMs. The study, which experimented with nine Mamba-based model designs, found that varying self-supervised objectives redistributed downstream skills rather than improving overall performance. The researchers propose that the limitation lies in the training targets, which are derived from audio and vision models and do not adequately capture the stochastic nature of earthquake occurrences. They suggest that a learning target defined across events is necessary to develop more capable seismic FMs. AI
IMPACT Challenges current approaches to foundation models in specialized domains, suggesting a need for domain-specific training targets.
RANK_REASON Research paper published on arXiv discussing foundation models. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
