Page 69 - Big Data Analytics for Intelligent Healthcare Management

62 CHAPTER 4 TRANSFER LEARNING AND SUPERVISED CLASSIFIER

Class 1
Class 2
Image
Input
Convolution + Relu Pooling Convolution + Relu Pooling Class n

Flatten Fully Softmax
connected

Feature learning Classification

FIG. 4.2
Sample convolution network architecture.
Data from Mathworks.com, Convolutional Neural Network, 2018. Available from:
https://www.mathworks.com/solutions/deep-learning/convolutional-neural-network.html. Accessed 10 June 2018.
In this work, four pretrained ConvNets architectures were used: ResNet50 [27], Inception V3 [28],
InceptionResnetV2[29], and Xception [30] with their default parameter settings with an average pool-
ing implemented in keras [31] deep learning library.
4.3.1.1 Transfer learning and convolution networks
Transfer learning is a machine learning method that allows the use of a model trained on a task to per-
form another task. Since the convolution architectures released by different organizations trained on
ImageNet databases containing 1.2 million images from 1000 categories is very large, training these
types of architectures for custom datasets is not practical because datasets are not large enough in prac-
tice. So instead of training the whole network, these pretrained networks are used. There are ways of
using a pretrained a convolution network that is, doing transfer learning using convolution network.
One of them is using the pretrained convolution network as fixed feature extractors [32].

4.3.1.2 Convolution networks as fixed feature extractors
To use a convolution network as a feature extractor, remove the last fully connected layer of a pre-
trained convolution network and then use the rest of the architecture as a fixed feature extractor for
the custom dataset. Then the extracted features can be used for other purposes [32] (Fig. 4.3).

4.3.1.3 Dimensionality reduction and principle component analysis (PCA)
It is difficult to train a learning algorithm with a higher dimensional data. Here comes the importance of
dimension reduction. Dimensionality reduction is a method of reducing the original dimension of data
to a lower dimension without much loss of information. Dimension reduction techniques have two
components. One is feature selection and the other is feature extraction. Feature selection is responsible
for selecting the subset of original attributes with specified parameters and feature extraction is respon-
sible for projecting the data into a lower dimensional space that is forming a new dataset with selected
attributes [33]. PCA is one of the popular dimension reduction algorithms that uses the orthogonal