 Today we are going to speak on the topic approaches to machine learning. At the end of the session, students will be able to demonstrate various approaches to machine learning. We have seen machine learning involves creating a model, training that model, testing new data and finally generating results. So machine learning techniques are required to improve the accuracy of predictive models. There are different approaches which are based on the type of data which is used and the volume of the data. Let us now think and answer what may be the different approaches to machine learning. Here we have some categories of machine learning approaches. The first one is supervised learning followed by unsupervised learning. Then we go for reinforcement learning. Then we apply neural networks to learn a particular system. And finally we go for advances in neural networks to permit deep learning. Supervised learning involves beginning with an established set of data which is already available. Now we try to understand how this data is to be classified. So it involves classification and then we intend to find patterns in that data that can be applied to an analytical process. Data now has labeled features and these features determine the meaning of the particular data. For example, the label might be animal and the data might be a cat. Training the model data so that it fits the details of the label is necessary. Once we have completed this training in a required amount of percentage of data from that which is involved. For example, 80-20, 80 for training and 20 for testing. When this label is continuous then it is called a regression environment. When the data comes from the finite set of values it is known as classification and we put that particular data into a particular class. Therefore regression is used for supervised learning and helps you to understand the correlation between the variables which are involved in a particular system. The process to be followed are consisting of the steps. The algorithm are trained using pre-processed examples and we see the follows of these particular algorithms generating results at intermediate steps. The performance of this algorithm is evaluated with the test data which I spoke about being 20% of the total data. Overfitting means that our model which we are dealing with is precisely tuned for a training of the particular data. Not being applicable for large sets of unknown data. So we are not making it a generalized learning algorithm and this testing needs to be done against unforeseen or unknown labeled data. So even if we do not have the complete data we should be able to predict what data is missing in our particular environment. Using this unforeseen data for the test set can help you to evaluate the accuracy of the model in predicting the outcomes and the results which are expected. Some of the examples of supervised learning are fraud detection, recommendation of solutions to a particular problem, speech recognition and risk analysis. Unsupervised learning is best suited when the problem requires massive amount of data that is unlabeled. So this consists of large amount of unlabeled data where there are far many variables involved in the particular environment. And therefore we have to group them into clusters. Hence it is based on clustering and the association between elements in this particular cluster. The process for unsupervised learning involves segmenting data into groups of examples which are called as clusters or groups of features which the particular objects are possessing. The unlabeled data creates the parameter values and classification of data. The process adds label to the data so that it becomes supervised. So we are going from unsupervised to reducing a problem to the supervised environment. The unsupervised learning can determine the outcome when there is a massive amount of data. So the developer doesn't know what is the context of the data being analyzed. So labeling isn't possible at this stage. So we can be using this particular format as the first step because it gives an ease for analysis and can be used as the raw or the first step before passing the data to a supervised learning process. Unsupervised learning algorithm can help the business understand large volume of new or unlabeled data. By new we mean that we have not visited these patterns before and by unlabeled data we mean that they are still not knowing what are the features of this particular data. So we look for patterns in the data. The difference is that the data is not already understood beforehand and therefore unsupervised learning approach can help determine outcomes more quickly than supervised learning approach. Hence is used as the first step for supervised learning approach many a times. Now let us try after understanding the concepts of supervised and unsupervised learning what might be the examples of unsupervised learning where large amount of data is involved, large amount of label is involved and the data is very voluminous. Such data may be that which is under the social media applications which is generated by a Twitter or Instagram or a Snapchat. Healthcare applications also have huge amount of data due to the number of diseases which are involved and the type of healthcare that has to be given for these particular diseases. The next type of learning is reinforcement learning which believes on having behavioural learning model experience. The algorithms used here receives feedback from the analysis of the data. A user is guided to the best outcome by giving some sort of reward and attracting it to a particular good goal. The system isn't trained with the sample data set. The system learns through trial and error till it gets a successful result and a sequence of successful decisions are generated which leads you to a particular goal. This will result in a process being reinforced, a sort of annealing process of hitting and hammering and then generating magnetism into the particular environment. This learning algorithm has to be able to discover an association between the items of our particular environment. Even in a complex scenario, the algorithm can be optimised over time to find ways to adapt to the state where actions are rewarded. The examples for reinforcement learning are robotics, game playing and self-driving cars. Now we go into doing learning using neural networks. A neural network is a network that has more than two layers of learning. The first layer called as the input layer and the second layer called as the output layer. Between these two layers there is another intermediate layer called as the hidden layer. Data is ingested into the particular environment through the input layer. Data is modified and transformed in a particular form by the hidden layer and the output layer is based on the weights applied to these nodes. Thus there is a weighted output which is generated for this particular environment. Neural networks consist of thousands or even sometimes millions of simple processing nodes called as neurons. It uses a hierarchical neural network to learn for a combination of both supervised and unsupervised learning algorithms. It learns from unlabeled and unstructured data to generate the required results. The examples of using neural network learning are image recognition, computer vision, voice recognition and malfunctioning of a particular machine. The references used for this particular creation of the video are Thank you.