Neural Networks - Model inference in Keras

Data normalization

• Features should be homogeneous and standardized.

# MNIST dataset - image data
train_images <- array_reshape(train_images, c(60000, 28 * 28))
train_images <- train_images / 255

# House price dataset - numerical features
mean <- apply(train_data, 2, mean)                                
std <- apply(train_data, 2, sd)
train_data <- scale(train_data, center = mean, scale = std)       

Vectorize labels

To vectorize the labels, there are two possibilities:

• We can use one-hot encoding:

to_one_hot <- function(labels, dimension) {
  results <- matrix(0, nrow = length(labels), ncol = dimension)
  for (i in 1:length(labels))
    results[i, labels[[i]] + 1] <- 1
  results
}

one_hot_train_labels <- to_one_hot(train_labels)          
one_hot_test_labels <- to_one_hot(test_labels)            

Note that there is a built-in way to do this in Keras:

train_labels <- to_categorical(train_labels)

• We can cast the label list as an integer tensor:

The only change is that we would have to use sparse_categorical_crossentropy loss function instead of the categorical_crossentropy loss function used for one-hot vector labels.

Handling missing values

Input missing values as number zero.

• In general safe with neural networks, assuming 0 is not already a meaningful value.

Artificially generating training samples with missing entries

• (Chollet and Allaire 2018) recommends to artificially generate training samples with missing entries
  • In case you expect missing values on the test set without having them on the training set
  • This is not statistically sound. Duplicating data would artificially reduce uncertainty.

Feature engeneering

• Neural networks are capable of automatically extracting useful features from raw data.

• Does this mean we do not have to worry about feature engineering as long as you’re using deep neural networks?

  No, for two reasons:

  • Good features still allow you to solve problems more elegantly while using fewer resources.
  • Good features let you solve a problem with far less data.

Fit and evaluate

Training the model

model %>% fit(train_data, train_labels, epochs = 5, batch_size = 128)

• train_data: Input tensor with the training features.
• train_labels: Input tensor with the training labels.
• epochs: Number of passes through the data.
• batch_size: Number of samples to use for each iteration of mini-batch SGD.

Training and validating a model

# We can later plot and manipulate data in 'history'
history <- model %>% 
  fit(partial_train_data, partial_label_data, 
      epochs = 20, batch_size = 512, 
      validation_data = list(validation_train_data, validation_label_data))

• validation_data: Data on which to evaluate the loss and any model metrics at the end of each epoch.

Different way to train and validate a model

history <- model %>% 
  fit(partial_train_data, partial_train_labels,
      epochs = num_epochs, batch_size = 1) %>%
  evaluate(validation_train_data, validation_label_data)    

Information leak

• Be mindful of the following: every time you use feedback from your validation process to tune your model, you leak information about the validation process into the model.
• This makes the evaluation process less reliable, if done systematically over many iterations.