Full text loading...
-
Manual Active Learning for Salt Interpretation: an Empirical Study to Avoid Forgetting During Incremental Trainings
- Publisher: European Association of Geoscientists & Engineers
- Source: Conference Proceedings, Third EAGE Digitalization Conference and Exhibition, Mar 2023, Volume 2023, p.1 - 5
Abstract
In seismic applications, the labelling is a challenging and tedious task due to the broad areas covered by the seismic data and requires expert knowledge. Consequently, finding solutions to limit the labelling effort is a priority to accelerate workflows and to optimize the human resources. The technique of active learning can help in reaching these goals. It consists in selecting the best data to label in order to improve the model performance based on an iterative approach during which, at each step, unlabeled data are chosen to be labelled and used to train the model. This process is repeated until the model reaches acceptable performances. The main challenge when incrementally training a neural network is the forgetting of the patterns learned during the previous training iterations. We showed that the choice of the old/new labels ratio in the training and validations sets, as well as the choice of the learning rate and the patience can help mitigate the knowledge loss in the case of incremental trainings.