Speaker: Roberto Cilli
Date: Thursday 25th of September 2025, 1:15pm
Abstract:
The seminar will first discuss a downstream application of a pretrained SentenceBERT + Random Forest (RF) for the lithological classification of texts from borehole logs. A custom lightweight architecture relying on a single Transformer module is proposed to handle contextual and positional relationships of each lithological text description. The proposed method achieves an accuracy gain of approximately 10% compared to a RF fed with SentenceBERT embeddings. Finally, a comparison between benchmark uncertainty quantification (UQ) algorithms, including Bayesian NN (Blundell et al. 2015), MAPIE Conformal Learning (Taquet et al. 2023) and Deep Ensemble (Lakshminarayanan et al. 2016) is shown. Preliminary results indicate that the Deep Ensemble UQ method seems the most reliable while still feasible in low-resource computing environments.