2019 February 25

Location: Coors Tech

Attending:

Thomas Iga Jihyun Bane Hayden Andy Antoine

Key takeaways

  • Hayden has lots of feedback in the comments.

Problems

  • Hayden has a ridiculously high number of feature dimensions
  • How to validate with low number of wells?
  • If trained successfully, would a model from here work in the field?

Comments

  • Polar coordinate interpolation to avoid nans
  • PCA to reduce dimensionality
  • Multi-dimensional scaling for data dimensionality reduction
  • Check error on slide 4 (normalize between -1 and 1)
  • A super low training loss may indicate over training.
  • Consider cross-validating across wells and models per well.
  • A dropout layer may help you reduce over training.
  • Consider leveraging spatial correlation with convolutional layers.
  • Space filling curves may help reorder cells into an ordered 1D sequence.
  • Consider autoencoding data then training network (alternative to PCA).
  • Consider intelligently selecting validation set (not randomly).

Topics of discussion for next week

  • Bin will discuss his project and get feedback

Comments