The Julia program below uses a multi-layer perceptron to classify images from the MNIST dataset. The program should run in several minutes and achieve an accuracy just under 97%. It has a few notable differences from the previous program that only used logistic regression.

This program uses the FluxML library. Run Pkg.add("Flux") before running it. The FluxML API is easier to use than TensorFlow because it does not separate graph construction and execution.

This program has tracked parameters featureThetas and featureTheta0 for the feature detection models as well as outputThetas and outputTheta0 for the classification model.

featureThetas = param([scale(r, 784, 500) for r in rand((784, 500))]) # param(zeros((784, 500)))
featureTheta0 = param([scale(r, 1, 500) for r in rand((1, 500))]) # param(zeros((1, 500)))
outputThetas = param([scale(r, 500, 10) for r in rand((500, 10))]) # param(zeros((500, 10)))
outputTheta0 = param([scale(r, 1, 10) for r in rand((1, 10))]) # param(zeros((1, 10)))

In FluxML, tracked parameters are like any other numerical or array type, with a notable exception: When you perform mathematical operations on a tracked parameter, it is able to record the gradients of the operations with respect to itself.

Feature detection happens with logistic regression models operating on raw inputs:

features(example) = 1.0 ./ (1.0 .+ exp.(example * featureThetas .+ featureTheta0))

Classification happens with a softmax model operating on features:

model(example) = NNlib.softmax(features(example) * outputThetas .+ outputTheta0)

During training, the back!() function calculates the cost with the side effect of accumulating gradients. The gradients are retrieved with calls to grad().

  update!(featureThetas, -learningRate .* grad(featureThetas))
  update!(featureTheta0, -learningRate .* grad(featureTheta0))
  update!(outputThetas, -learningRate .* grad(outputThetas))
  update!(outputTheta0, -learningRate .* grad(outputTheta0))

Number of Features

The model implemented by this program detects 500 features of the input. The number 500 is called a hyperparameter of the model (in contrast to the parameters ). The choice of hyperparameter values is a matter of experience and experimentation. Generally, if the number of features is smaller than the number of inputs, then the features are a “compressed” representation of the input.

Initialization of

The previous program initialized thetas and theta0 values to zero. For multi-layer perceptrons, this is not a good idea. When their parameter values are exactly the same, the feature detection models are all detecting the same feature! Their gradients will be the same and they will be updated by the same amounts, rendering them largely redundant. Researchers have searched for good initial values of , and this program uses random values in a range recommended by Yoshua Bengio and Xavier Glorot in their AISTATS 2010 paper, Understanding the difficulty of training deep feedforward neural networks. To convince yourself of the value of careful initialization, replace the random initializations with the calls to zeros in the comments. You will notice a decline in accuracy.