Purely functional ML Framework for Scala

Right now it is just a basic concept. You are welcome to contribute to bring ML into the Scala ecosystem.

Examples

ANN

There is an example of training ANN on MNIST dataset:

"neural network" should "classify MNIST data set" in {
    val training = MNIST.loadTrainingSet(10000)
    val test = MNIST.loadTestSet(100)
    val (factor, input) = normalize(training.input)
    val model = Dense(50, Sigmoid(), kernelReg = L2(0.005)) |&| Dense(10, Sigmoid(), kernelReg = L2(0.005))
    val weights = Adam(rate = 0.01, batch = 100)
      .minimize(nnError(model), horzcat(input, training.labels))
      .through(epoch(10))
      .observe(logStdOut)
      .observe(plotToFile("MNIST.png"))
      .runSync.vars
    val classifier = nn(model)(weights)
    val prediction = classifier(normalize(test.input, factor))
    val accuracy = binaryAccuracy(test.labels, prediction)
    accuracy should be > 0.9
  }

Features:

• Stochastic Gradient Decent (SGD)
• SDG features (momentum, nesterow)
• Advanced Gradient Decent optimizations (AdaGrad, AdaDelta, RMSProp, Adam, NAdam, Adamax)
• Linear regression
• Logistic regression
• Artificial NN (Neural Network)
• Performance
• Documentation
• Convolutional NN
• Metrics
• Preprocessors
• Visualize NN learning process
  • REST service
  • UI
• Recurrent NN
• Tensor Flow as a backend