WHY DEEP LEARNING IS REQUIRED?

Deep learning algorithms are the subset of machine learning algorithms. There are certain drawbacks for machine learning algorithms which a deep learning technique can overcome.

Some of the reasons why deep learning is better than machine learning algorithms are as follows:

PROBLEM WITH FEATURE EXTRACTION AND LEARNING IN MACHINE LEARNING ALGORITHMS:

Typical machine learning algorithms extract the features which are manually given by the programmer. It is a challenging task to select what features have best discriminating power in classification or prediction. If the manually extracted features are not good enough, then the machine learning algorithm will fail. The decision regarding whether what set of predefined-features to select for a given training set comes from either experience or literature knowledge of a machine learning programmer. The features, thus selected may or may not produce accurate results for a different set of examples.

A machine learning algorithm never has the ability to learn from its own features. The algorithm will work only for the set of features which is feed to it. It cannot develop new features from manually extracted features.

HOW DEEP LEARNING OVERCOMES THE PROBLEM

Deep learning techniques are feature learning approaches. They learn a feature from already existing features. Deep learning techniques are good at automatically extracting features from an input data.

Deep learning architecture is designed in such a way that the learning process takes place at a hierarchical manner. Each level of hierarchy in a deep learning system represents a concept or feature in real world which are linked to each other. The learning process takes place in such a way that the concepts at one level are linked to other concepts at higher level. When a concept is learned at one level, it automatically finds the other concepts which can be inferred from already learned concepts. This is how learning occurs in human brain. The deep learning algorithms are trying to do that. Thus, we can extract many good features from input data in a dynamic manner.

A SMALL EXAMPLE OF “MANGO TREE OR JACKFRUIT TREE CLASSIFICATION FROM IMAGES” TO JUSTIFY HOW WELL DEEP LEARNING WORKS

Machine learning approach:

Initially, the features like width, height, shape of leaves are manually selected. The features are extracted from the leaves and a classifier is only built for those features.

Deep learning approach:

The input image of the leaves is fed into the learning architecture. The learning takes place in such a way that it extracts the useful features by observing the training examples. In this case, the first layer looks for useful edges or boundaries that can be used for classification. The information regarding edges from this layer is then passed to the next layer. In next layer, it finds out what combinations of these edges can be used as feature. It tries to find whether the edges which are formed to form the horizontal, vertical lengths can be used or not. The useful information from this layer is then passed to next layer, what horizontal or vertical lengths can be further reduced to form a useful feature. This procedure continues until a learning criterion is obtained.

PROS AND CONS ON USING DEEP LEARNING:

· Deep learning cannot be used for small training data sets. In such cases, machine learning algorithms are used.

· Deep learning algorithms are really good when compared to machine learning algorithms provided the dataset is very large because as data increases, more and more number of features can be finding out by observing more examples.