keras autoencoder vs PCA

Question

I am playing with a toy example to understand PCA vs keras autoencoder

I have the following code for understanding PCA:

import numpy as np
import mat         


        

Answer 1

First of all, the aim of an autoencoder is to learn a representation (encoding) for a set of data, typically for the purpose of dimensionality reduction. So, the target output of the autoencoder is the autoencoder input itself.

It is shown in [1] that If there is one linear hidden layer and the mean squared error criterion is used to train the network, then the k hidden units learn to project the input in the span of the first k principal components of the data. And in [2] you can see that If the hidden layer is nonlinear, the autoencoder behaves differently from PCA, with the ability to capture multi-modal aspects of the input distribution.

Autoencoders are data-specific, which means that they will only be able to compress data similar to what they have been trained on. So, the usefulness of features that have been learned by hidden layers could be used for evaluating the efficacy of the method.

For this reason, one way to evaluate an autoencoder efficacy in dimensionality reduction is cutting the output of the middle hidden layer and compare the accuracy/performance of your desired algorithm by this reduced data rather than using original data. Generally, PCA is a linear method, while autoencoders are usually non-linear. Mathematically, it is hard to compare them together, but intuitively I provide an example of dimensionality reduction on MNIST dataset using Autoencoder for your better understanding. The code is here:

from keras.datasets import mnist 
from keras.models import Model 
from keras.layers import Input, Dense 
from keras.utils import np_utils 
import numpy as np

num_train = 60000
num_test = 10000

height, width, depth = 28, 28, 1 # MNIST images are 28x28
num_classes = 10 # there are 10 classes (1 per digit)

(X_train, y_train), (X_test, y_test) = mnist.load_data()

X_train = X_train.reshape(num_train, height * width)
X_test = X_test.reshape(num_test, height * width)
X_train = X_train.astype('float32') 
X_test = X_test.astype('float32')

X_train /= 255 # Normalise data to [0, 1] range
X_test /= 255 # Normalise data to [0, 1] range

Y_train = np_utils.to_categorical(y_train, num_classes) # One-hot encode the labels
Y_test = np_utils.to_categorical(y_test, num_classes) # One-hot encode the labels

input_img = Input(shape=(height * width,))

x = Dense(height * width, activation='relu')(input_img)

encoded = Dense(height * width//2, activation='relu')(x)
encoded = Dense(height * width//8, activation='relu')(encoded)

y = Dense(height * width//256, activation='relu')(x)

decoded = Dense(height * width//8, activation='relu')(y)
decoded = Dense(height * width//2, activation='relu')(decoded)

z = Dense(height * width, activation='sigmoid')(decoded)
model = Model(input_img, z)

model.compile(optimizer='adadelta', loss='mse') # reporting the accuracy

model.fit(X_train, X_train,
      epochs=10,
      batch_size=128,
      shuffle=True,
      validation_data=(X_test, X_test))

mid = Model(input_img, y)
reduced_representation =mid.predict(X_test)

out = Dense(num_classes, activation='softmax')(y)
reduced = Model(input_img, out)
reduced.compile(loss='categorical_crossentropy',
          optimizer='adam', 
          metrics=['accuracy']) 

reduced.fit(X_train, Y_train,
      epochs=10,
      batch_size=128,
      shuffle=True,
      validation_data=(X_test, Y_test))


 scores = reduced.evaluate(X_test, Y_test, verbose=1) 
 print("Accuracy: ", scores[1])

It produces a $y\in \mathbb{R}^{3}$ ( almost like what you get by decomposition.PCA(n_components=3) ). For example, here you see the outputs of layer y for a digit 5 instance in dataset:

  class  y_1    y_2     y_3     
  5      87.38  0.00    20.79

As you see in the above code, when we connect layer y to a softmax dense layer:

mid = Model(input_img, y)
reduced_representation =mid.predict(X_test)

the new model mid give us a good classification accuracy about 95%. So, it would be reasonable to say that y, is an efficiently extracted feature vector for the dataset.

---

References:

[1]: Bourlard, Hervé, and Yves Kamp. "Auto-association by multilayer perceptrons and singular value decomposition." Biological cybernetics 59.4 (1988): 291-294.

[2]: Japkowicz, Nathalie, Stephen Jose Hanson, and Mark A. Gluck. "Nonlinear autoassociation is not equivalent to PCA." Neural computation 12.3 (2000): 531-545.