Skip to content

Roadmap

Mikko Kotila edited this page Jul 29, 2019 · 1 revision

[NOTE: this was created in 2017 may and is left here for now just as a historical account. Some of it is still relevant, and some not.]

Development Objectives

The development goals and current work include:

  • to expose 100% of Keras functionality to the user
  • to create robust synthesis of generality and performance in a single metric
  • to use the synthesis score as a metric for optimizing the hyperparameter optimization process
  • to leverage gradient based optimization approaches
  • to allow on-the-go model ensembling
  • to allow ensembling of model ensembles
  • to allow reverse inheritance of the optimization capacity i.e. models that build models that...

These goals are currently being met by systematically building simple, easy-to-understand building blocks that solve manageable parts of the challenge.

Immediate Development

  • Show false negatives / false positives / border cases
  • Robust cross-validation
  • Generalization measure
  • Prediction task agnostic performance measure
  • Adding layer generation for other than Dense layers
  • Current custom f1 implementation into Keras metric
  • Add support for fastai?