Workshop on Deep Neural Networks in Java | by Sogeti

Why use Java for ML?

  • Java is the most used language
  • There are C++/Python wrappers
  • Java shortcoming: lack of efficient parallelism on the JVM. Lowest abstraction level are thread objects.
  • ND4J
    • C++ engine (libnd4j)
    • JNI (Java Native Interface) binding to integrate with the JVM

DL4J concepts

  • Vectors and tensors
  • Load data with DataVec
  • INDArray: [nr of examples x nr of features]
  • Supports mini-batches
  • Train with .fit() method
  • Build per layer
  • Optimizer splits from updater
  • Evaluation class present
  • Visualisation and monitoring tools