v48.3.0 had a pure Neural Network focus:

  • Support for float
  • Possibility to “get” all individual parameters of the network
  • New file format to save trained networks to disk
  • Improved separation between building, training and invoking the network
  • Possible to have several network invokers used in different threads.
  • Support for dropouts as well as L1 and L2 regularisation when training the network

Example Code