What do we mean by the term machine learning? The definition is quite difficult, to do so, we are asking large field of scientists to help. We can mention an artificial intelligence pioneer's quote:

 – Arthur Samuel

Machine learning is about training a model or an algorithm with data and then using the model to predict any new data. For example, a toddler is taught how to walk from his crawling phase. Initially, the toddler's parents hold the toddler's hand to help him up, and he is taught through the data that is given. On the basis of these procedures, if an obstacle presents itself in the toddler's path or if there is a turn somewhere, the toddler is able to navigate on his own after the training. The data used for training is the training data and the recipient continues to learn even after the formal training.

Machines too can be taught like toddlers to do a task based...

Supervised learning is a learning method where there is a part of the training data which acts as a teacher to the algorithm to determine the model. The machine is taught what to learn from the target data. The target data, or dependent or response variables, are the outcome of the collective action of the independent variables. The network training is done with the target data and its behavior with patterns of input data. The target labels are known in advance and the data is fed to the algorithm to derive the model.

Most of neural network usage is done using supervised learning. The weights and biases are adjusted based on the output values. The output can be categorical (like true/false or 0/1/2) or continuous (like 1,2,3, and so on). The model is dependent on the type of output variables, and in the case of neural networks, the output layer is built on the type of target variable.

In unsupervised learning (or self organization), the output layer is trained to organize the input data into another set of data without the need of a target variable. The input data is analyzed and patterns are found in it to derive the output, as shown in the following figure. Since there is no teacher (or target variable), this type of learning is called unsupervised learning.

The different techniques available for unsupervised learning are as follows:

To summarize, the two main types of machine learning are depicted in the following figure:

For neural networks, we have both the types available, using different ways available in R.

Reinforcement learning is a type of machine learning where there is constant feedback given to the model to adapt to the environment. There is a performance evaluation at each step to improve the model. For neural networks, there is a special type called Q-learning, combined with neuron to implement reinforcement learning in the backpropagation feedback mechanism. The details are out of scope of this book.

The following are the three types of learnings we have covered so far:

Training and testing the model forms the basis for further usage of the model for prediction in predictive analytics. Given a dataset of 100 rows of data, which includes the predictor and response variables, we split the dataset into a convenient ratio (say 70:30) and allocate 70 rows for training and 30 rows for testing. The rows are selected in random to reduce bias.

Once the training data is available, the data is fed to the neural network to get the massive universal function in place. The training data determines the weights, biases, and activation functions to be used to get to output from input. Until recently, we could not say that a weight has a positive or a negative influence on the target variable. But now we've been able to shed some light inside the black box. For example, by plotting a trained neural network, we can discover trained synaptic weights and basic information about the training process.

Once the sufficient convergence is achieved, the...

The data forms a key component for model building and the learning process. The data needs to be collected, cleaned, converted, and then fed to the model for learning. The overall data life cycle is shown as follows:

One of the critical requirements for modeling is having good and balanced data. This helps in higher accuracy models and better usage of the available algorithms. A data scientist's time is mostly spent on cleansing the data before building the model.

We have seen the training and testing before deployment of the model. For testing, the results are captured as evaluation metrics, which helps us decide if we should use a particular model or change it instead.

We will see the evaluation metrics next.

Evaluating a model involves checking if the predicted value is equal to the actual value during the testing phase. There are various metrics available to check the model, and they depend on the state of the target variable.

For a binary classification problem, the predicted target variable and the actual target variable can be in any of the following four states:

When we have the predicted and actual values as same values, we are said to be accurate. If all predicted and actual values are same (either all TRUE or all FALSE), the model is 100 percent accurate. But, this is never the case.

Since neural networks are approximation models, there is always a bit of error possible. All the four states mentioned in the previous table are possible.

We define the following terminology and metrics for a model:

As we saw in Chapter 1, Neural Network and Artificial Intelligence Concepts, neural networks is a machine learning algorithm that has the ability to learn from data and give us predictions using the model built. It is a universal function approximation, that is, any input, output data can be approximated to a mathematical function. 

The forward propagation gives us an initial mathematical function to arrive at output(s) based on inputs by choosing random weights. The difference between the actual and predicted is called the error term. The learning process in a feed-forward neural network actually happens during the backpropagation stage. The model is fine tuned with the weights by reducing the error term in each iteration. Gradient descent is used in the backpropagation process.

Let us cover the backpropagation in detail in this chapter, as it is an important machine learning aspect for neural networks.