Announcing DeepLearning.scala 2.0.0

26 July 2017

Today, we are happy to announce DeepLearning.scala 2.0.0, the new stable release of DeepLearning.scala, a simple library for creating complex neural networks from object-oriented and functional programming constructs.

  • DeepLearning.scala runs on JVM, can be used either in standalone JVM applications or Jupyter Notebooks.
  • DeepLearning.scala is expressive. Various types of neural network layers can be created by composing map, reduce or other higher order functions.
  • DeepLearning.scala supports plugins. You can share your own algorithms, models, hyperparameters as a plugin, as simple as creating a Github Gist.
  • All the above features are statically type checked.

Features in DeepLearning.scala 2.0

In DeepLearning.scala 2.0, we removed the special support for differentiable ADT and Boolean types. Now differentiable computational graphs are ordinary Scala code, so all types including ADT and Boolean are avialable in these graphs.

Dynamic neural networks

Unlike some other deep learning frameworks, the structure of neural networks in DeepLearning.scala is dynamically determined during running. Our neural networks are programs. All Scala features, including functions and expressions, are available in neural networks.

For example:

def ordinaryScalaFunction(a: INDArray): Boolean = {
  a.signnum.sumT > math.random
}

def myDynamicNeuralNetwork(input: INDArray) = INDArrayLayer(monadic[Do] {
  val outputOfLayer1 = layer1(input).forward.each
  if (ordinaryScalaFunction(outputOfLayer1.data)) {
    dynamicallySelectedLayer2(outputOfLayer1).forward.each
  } else {
    dynamicallySelectedLayer3(outputOfLayer1).forward.each
  }
})

The above neural network will go into different subnetworks according to an ordinary Scala function.

With the ability of creating dynamic neural networks, regular programmers are able to build complex neural networks from simple code. You write code almost as usual, the only difference being that code based on DeepLearning.scala is differentiable, which enables such code to evolve by modifying its parameters continuously.

Functional programming

DeepLearning.scala 2.0 is based on Monads, which are composable, thus a complex layer can be built from primitive operators. Along with the Monad, we provide an Applicative type class, to perform multiple calculations in parallel.

For example, the previous example can be rewritten in higher-order function style as following:

def myDynamicNeuralNetwork(input: INDArray) = INDArrayLayer {
  layer1(input).forward.flatMap { outputOfLayer1 =>
    if (ordinaryScalaFunction(outputOfLayer1.data)) {
      dynamicallySelectedLayer2(outputOfLayer1).forward
    } else {
      dynamicallySelectedLayer3(outputOfLayer1).forward
    }
  }
}

The key construct in DeepLearning.scala 2.0 is the dependent type class DeepLearning, which witnesses a differentiable expression. In other words, given the DeepLearning type class instance, you can activate the deep learning ability of any type.

Object-oriented programming

The code base of DeepLearning.scala 2.0 is organized according to Dependent Object Type calculus (DOT). All features are provided as mixin-able plugins. A plugin is able to change APIs and behaviors of all DeepLearning.scala types. This approach not only resolves expression problem, but also gives plugins the additional ability of virtually depending on other plugins.

For example, when a plugin author is creating the Adagrad optimizer plugin, he does not have to explicitly call functions related to learning rate. However, once a plugin user enables both the Adagrad plugin and the FixedLearningRate plugin, then computation in FixedLearningRate will get called eventually when the Adagrad optimization is executed.

Plugins for DeepLearning.scala 2.0

Plugin Name Plugin Description
Builtins All the built-in plugins.
FixedLearningRate Setup fixed learning rate when training INDArray weights.
Adagrad An adaptive gradient algorithm with per-parameter learning rate for INDArray weights.
L1Regularization L1 Regularization.
L2Regularization L2 Regularization.
Momentum The Momentum and NesterovMomentum optimizer for SGD.
RMSprop The RMSprop optimizer for SGD.
Adam The Adam optimizer for SGD.
INDArrayDumping A plugin to dump weight matrices during training.
CNN A standalone CNN implementation.
Add your own algorithms, models or any cool features here.
DeepLearning.scala is an open source deep-learning toolkit in Scala created by our colleagues at ThoughtWorks. We're excited about this project because it uses differentiable functional programming to create and compose neural networks; a developer simply writes code in Scala with static typing.