In the previous chapter we used a support vector machine to classify the handwritten digits in the MNIST dataset. In this section we will classify the images using an artificial neural network:

from sklearn.datasets import load_digits
from sklearn.cross_validation import train_test_split, cross_val_score
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.neural_network.multilayer_perceptron import MultilayerPerceptronClassifier

First we use the load_digits convenience function to load the MNIST dataset. We will fork additional processes during cross validation, which requires execution from a main-protected block:

>>> if __name__ == '__main__':
>>>     digits = load_digits()
>>>     X = digits.data
>>>     y = digits.target

Scaling the features is particularly important for artificial neural networks and will help some learning algorithms to converge more quickly. Next, we create a Pipeline class that scales the data before fitting a MultilayerPerceptronClassifier. This network contains an input layer, a hidden layer with 150 units, a hidden layer with 100 units, and an output layer. We also increased the value of the regularization hyperparameter alpha argument. Finally, we print the accuracies of the three cross validation folds. The code is as follows:

>>>     pipeline = Pipeline([
>>>         ('ss', StandardScaler()),
>>>         ('mlp', MultilayerPerceptronClassifier(n_hidden=[150, 100], alpha=0.1))
>>>     ])
>>>     print cross_val_score(pipeline, X, y, n_jobs=-1)
Accuracies [ 0.95681063  0.96494157  0.93791946]

The mean accuracy is comparable to the accuracy of the support vector classifier. Adding more hidden units or hidden layers and grid searching to tune the hyperparameters could further improve the accuracy.