Non-contrastive representation learning for intervals from well logs

The representation learning problem in the oil & gas industry aims to construct a model that provides based on logging data for well interval. Previous attempts are mainly supervised and focus similarity task, which estimates closeness between intervals. We desire build informative representations without using (labelled) data. One of possible approaches is self-supervised (SSL). In contrast paradigm, this one requires little or no labels Nowadays, most SSL either contrastive non-contrastive. Contrastive methods make similar ( positive ) objects closer distancing different xmlns:xlink="http://www.w3.org/1999/xlink">negative ones. Due wrong marking positive negative pairs, these can provide an inferior performance. Non-contrastive don’t rely such labelling widespread computer vision. They learn only pairs easier identify first introduce non-contrastive well-logging particular, we exploit Bootstrap Your Own Latent (BYOL) Barlow Twins avoid matching pairs. crucial part augmentation strategy. Our strategies adaption BYOL together allow us achieve superior quality clusterization mostly best performance classification tasks. results prove usefulness proposed interval particular.