Learning Path: R: Complete Machine Learning & Deep Learning

This is give you brief information about the course.

R must be first installed on your system to work on it.

RStudio makes the process of development with R easier.

R packages are an essential part of R as they are required in all our programs. Let's learn to do that.

You must know how to give data to R to work with data. You will learn that here.

Data manipulation is time consuming and hence needs to be done with the help of built-in R functions.

R is widely used for statistical applications. Hence it is necessary to learn about the built in functions of R.

To communicate information effectively and make data easier to comprehend we need graphical representation. You will learn to plot figures in this section.

Because of some limitations, it is a good practice to get data from external repositories. You will be able to do just that after this video.

Reading a dataset is the first and foremost step in data exploration. We need to learn to how to do that.

In R, since nominal, ordinal, interval, and ratio variable are treated differently in statistical modeling, we have to convert a nominal variable from a character into a factor.

Missing values affect the inference of a dataset. Thus it is important to detect them.

After detecting missing values, we need to impute them as their absence may affect the conclusion.

After imputing the missing values, you should perform an exploratory analysis to summarize the data characteristics.

The exploratory analysis helps users gain insights into how single or multiple variables may affect the survival rate. However, it does not determine what combinations may generate a prediction model. We need to use a decision tree for that.

After constructing the prediction model, it is important to validate how the model performs while predicting the labels.

Another way of measuring performance is the ROC curve.

When there are huge datasets, we can find the characteristics of the entire dataset with a part or sample of the data. Hence data sampling is essential.

Probability distribution and statistics are interdependent. To provide a justification to the statistical information, we need probability.

Univariate statistics deals with a single variable and hence is very simple.

To analyze the relation among more than two variables, multivariate analysis is done.

Assessing the relation between dependent and independent variables is carried out through linear regression.

To validate that the experiment results are significant, hypothesis testing is done.

To compare means of two different groups, one- and two-sample t-tests are conducted.

Comparing a sample with a reference probability or comparing cumulative distributions of two data sets calls for a Kolmogorov- Smirnov test.

The Wilcoxon Test is a non-parametric test for null hypothesis.

To check the distribution of categorical variables of two groups, Pearson's chi-squared test is used.

To examine the relation between categorical independent variables and continuous dependent variables, Anova is used. When there is a single variable, one-way ANOVA is used.

When there are two categorical values to be compared, two-way ANOVA is used.

Linear regression is the simplest model in regression and can be used when there is one predictor value.

To obtain summarized information of a fitted model, we need to learn how to summarize linear model fits.

It would be really convenient for us if we could predict unknown values. You can do that using linear regression.

To check if the fitted model adequately represents the data, we perform diagnostics.

In the case of a non-linear relationship between predictor and response variables, a polynomial regression model is formed. We need to fit the model. This video will enable you to do that.

An outlier will cause diversion from the slope of the regression line. In order to avoid that, we need to fit a robust linear regression model.

We will perform linear regression on a real-life example, the SLID dataset.

GLM generalizes linear regression by allowing the linear model to be related to the response variable via a link function and by allowing the magnitude of the variance of each measurement to be a function of its predicted value.

GLM allows response variables with error distribution other than a normal distribution. We apply the Poisson model to see how that is done.

When a variable is binary, we apply the binomial model.

GAM has the ability to deal with non-linear relationships between dependent and independent variables. We learn to fit a regression using GAM.

Visualizing a GAM helps it to understand better.

You can also diagnose a GAM model to analyze it.

Training and testing datasets are both essential for building a classification model.

A partitioning tree works on the basis of split condition starting from the base node to the terminal node.

Plotting the classification tree will make analyzing the data easier. You will learn to do this now.

Before making a prediction, it is essential to compute the prediction performance of the model.

There can be parts in a dataset which are not essential for classification. In order to remove these parts, we have to prune the dataset.

Conditional inference trees are better than traditional classification trees because they adapt the test procedures for selecting the output.

Visualizing a conditional inference tree will make it easier to extract and analyze data from the dataset.

Like the prediction performance of a traditional classification tree, we can also evaluate the performance of a conditional inference tree.

k-nearest neighbor classifier is a non parametric lazy learning method. Thus it has the advantages of both the types of methods.

Classification in logistic regression is done based one or more features. It is more robust and doesn't have as many conditions as the traditional classification model.

The Naïve Bayes classifier is based on applying Bayes' theorem with a strong independent assumption.

Support vector machines are better at classification because they can capture complex relations between data points and provide both linear and non-linear classifications

To control our training errors and margins, we use the cost function. The SVM classifier is affected by the cost.

To visualize the SVM fit, we can use the plot function.

We can use the trained SVM to predict labels on a model.

According to the desired output, you may need to generate different combinations of gamma and cost to train different SVMs. This is called tuning.

A neural network is used in classification, clustering and prediction. Its efficiency depends on how well you train it. Let's learn to do that.

We can use the trained SVM to predict labels on a model.

Similar to other classification models, we can predict labels using neural networks and also validate performance using confusion matrix.

Nnet provides the functionality to train feed-forward neural networks with backpropagation.

As we have already trained the neural network using nnet, we can use the model to predict labels.

The k-fold cross-validation technique is a common technique used to estimate the performance of a classifier as it overcomes the problem of over-fitting. In this video we will illustrate how to perform a k-fold cross-validation.

In this video, we will illustrate how to use tune.svm to perform 10-fold cross-validation and obtain the optimum classification model.

In this video we will demonstrate how to perform k-fold cross validation using the caret package.

This video will show you how to rank the variable importance with the caret package.

In this video, we will illustrate how to use rminer to obtain the variable importance of a fitted model.

In this video we will show how to find highly correlated features using the caret package.

In this video, we will demonstrate how to use the caret package to perform feature selection.

To measure the performance of a regression model, we can calculate the distance from the predicted output and the actual output as a quantifier of the performance of the model. In this video we will illustrate how to compute these measurements from a built regression model.

In this video we will demonstrate how to retrieve a confusion matrix using the caret package.

In this video, we will demonstrate how to illustrate an ROC curve and calculate the AUC to measure the performance of a classification model.

In this video we will use the function provided by the caret package to compare different algorithm-trained models on the same dataset.

In this video we will see how to measure performance differences between fitted models with the caret package.

The adabag package implements both boosting and bagging methods. For the bagging method, the package first generates multiple versions of classifiers, and then obtains an aggregated classifier. Let's learn the bagging method from adabag to generate a classification model.

To assess the prediction power of a classifier, you can run a cross validation method to test the robustness of the classification model. This video will show how to use bagging.cv to perform cross validation with the bagging method.

Boosting starts with a simple or weak classifier and gradually improves it by reweighting the misclassified samples. Thus, the new classifier can learn from previous classifiers. One can use the boosting method to perform ensemble learning. Let's see how to use the boosting method to classify the telecom churn dataset.

Similar to the bagging function, adabag provides a cross validation function for the boosting method, named boosting.cv. In this video, we will learn how to perform cross-validation using boosting.cv.

Gradient boosting creates a new base learner that maximally correlates with the negative gradient of the loss function. One may apply this method on either regression or classification problems. But first, we need to learn how to use gbm.

A margin is a measure of certainty of a classification. It calculates the difference between the support of a correct class and the maximum support of an incorrect class. This video will show us how to calculate the margins of the generated classifiers.

The adabag package provides the errorevol function for a user to estimate the ensemble method errors in accordance with the number of iterations. Let's explore how to use errorevol to show the evolution of errors of each ensemble classifier.

Random forest grows multiple decision trees which will output their own prediction results. The forest will use the voting mechanism to select the most voted class as the prediction result. In this video, we illustrate how to classify data using the randomForest package.

At the beginning of this section, we discussed why we use ensemble learning and how it can improve the prediction performance. Let's now validate whether the ensemble model performs better than a single decision tree by comparing the performance of each method.

Hierarchical clustering adopts either an agglomerative or a divisive method to build a hierarchy of clusters. This video shows us how to cluster data with the help of hierarchical clustering.

In this video we demonstrate how to use the cutree function to separate the data into a given number of clusters.

In this video, we will demonstrate how to perform k-means clustering on the customer dataset.

We will now illustrate how to create a bivariate cluster plot.

In this video we will see how to compare different clustering methods using cluster.stat from the fpc package.

In this video we will see how to compute silhouette information.

In this video we will discuss how to find the optimum number of clusters for the k-means clustering method.

In this video, we will demonstrate how to use DBSCAN to perform density-based clustering.

In this video, we will demonstrate how to use the model-based method to determine the most likely number of clusters.

A dissimilarity matrix can be used as a measurement for the quality of a cluster. In this video, we will discuss some techniques that are useful to visualize a dissimilarity matrix.

In this video, we will demonstrate how clustering methods differ with regard to data with known clusters.

Before starting with a mining association rule, you need to transform the data into transactions. This video will show how to transform any of a list, matrix, or data frame into transactions.

The arule package uses its own transactions class to store transaction data. As such, we must use the generic function provided by arule to display transactions and association rules. Let's see how to display transactions and association rules via various functions in the arule package.

Association mining is a technique that can discover interesting relationships hidden in transaction datasets. This approach first finds all frequent itemsets and then generates strong association rules from frequent itemsets. In this video, we see how to perform association analysis using the apriori rule.

This video will show you how to determine the number of principal components using the Kaiser method.

Besides listing rules as text, you can visualize association rules, making it easier to find the relationship between itemsets. In this video, we will learn how to use the aruleViz package to visualize the association rules.

An apriori algorithm performs a breadth-first search to scan the database. So, support counting becomes time consuming. Alternatively, if the database fits into the memory, you can use the Eclat algorithm, which performs a depth-first search to count the supports. Let's see how to use the Eclat algorithm.

In addition to mining interesting associations within the transaction database, we can mine interesting sequential patterns using transactions with temporal information. This video demonstrates how to create transactions with temporal information.

In contrast to association mining, we should explore patterns shared among transactions where a set of itemsets occurs sequentially. One of the most famous frequent sequential pattern mining algorithms is the Sequential Pattern Discovery using Equivalence classes (SPADE) algorithm. Let's see how to use SPADE to mine frequent sequential patterns.

This video will give you an introduction on how to perform feature selection with the FSelector package.

Principal component analysis (PCA) is the most widely used linear method in dealing with dimension reduction problems. This video will show you how to use it.

This video demonstrates how to determine the number of principal components using a scree plot. Let's have a look at it.

This video will show you how to determine the number of principal components using the Kaiser method.

Let's see how to use biplot to plot both variables and data

In MDS, you can either use a metric or a nonmetric solution. This video illustrates how to perform MDS on the swiss dataset.

You may require several times, reducing the dimension of matrices while working on datasets. Let us see how we could do this with SVD.

Let's see how to perform SVD on the classic image processing material, Lenna.

This video will show you how to perform a nonlinear dimension reduction with ISOMAP. This is one of the approaches to manifold learning and generalizes linear frameworks to nonlinear data structures.

This video will give you a short introduction of how to use LLE on an s-curve data

In order to prepare the RHadoop environment we need to download the Cloudera and QuickStart VM.

Installation of R and the rmr2 package is essential to perform MapReduce, which is used in performing data processing and analysis.

In order to access HDFS resources you need to install rhdfs on every task node.

You can easily operate HDFS from the R console with the help of rhdfs.

You will understand how rmr2 is used for word count in this video.

Comparing Hadoop and a standard R program can help us decide which language is best suited for our needs.

Since running a MapReduce program will require a considerable amount of time, testing and debugging become very important.

Plyrmr makes data manipulation operations easy.

Writing a MapReduce program can be difficult for non-developers. Hence plyr-like operations can be used to manipulate data.

You can perform machine learning operations with RHadoop.

Multinode clusters can be deployed with the help of Amazon EMR on RHadoop.

This video provides an overview of the entire course.

The main objective is to understand the fundamental concepts and key features that make it so special and different from the classical Machine Learning approach.

The goal of this video is to learn more about Artificial Neural Networks and their vast world of variations, explore the basic architectures of ANNs in detail and talk about their possible implementations in R.

Applying what you have learned about the Multilayer Perceptron algorithm to a real-world application, which classifies handwritten digits in images.

To get probabilistic predictions using Artificial Neural Networks and specifically in the context of a multi-class classification problem.

To add multiple hidden layers to the basic Multilayers Perceptron algorithm in order to build more complex models of the world and increase the accuracy of our predictions.

The goal of this video is to learn the best practices for tuning the hyper-parameters of an ANN and being able to generalize well on the data we have never seen before. This would be the latest essential skill to acquire in order to get the best out of our ANN solution.

The goal of this video is to learn more about Multi-hidden-layer Neural Networks and how to use them in order to solve the practical problem of classifying handwritten digits within the R language.

The goal of this video is to apply what we have learned about Multi-hidden-layer ANNs to a new real-world problem and get more confidence in the use of the H2O package.

The main objective is to understand the optimization process behind and common to every Deep Learning model with a more formal definition with respect to what was previously introduced.

To explore with more details, the most common algorithm to minimize the loss function called Stochastic Gradient Descent.

The goal of this video is to understand how to actually learn the weights of our Deep Learning model using Stochastic Gradient Descent through Backpropagation, the standard way of computing the gradient for Artificial Neural Networks.

To get to tune the hyper-parameters automatically in order to minimize the error on the validation set.

This first video, will be an introduction to the fundamental concepts behind Convolutional Neural Networks. The main objective of this video is to motivate their use highlighting the differences from classical feed-forward neural networks.

The goal of this video is to learn more about Convolutional Neural Networks, concluding our dissertation on the layer-wise structure of a CNN and understand how to design architecture suitable for your specific problem.

The aim of this video is to understand how to actually implement CNNs in R, and use it to solve real-world problems.

The goal of this video is to learn about the concept of transfer Learning, and how we can use and exchange DL pre-trained models to solve even new tasks with a very tiny computational overhead.

The aim of this video is to introduce the fundamental concepts behind Recurrent Neural Networks. The main objective is to underline their main differences from classical feed-forward neural networks and CNNs.

The aim of this video is to learn more about a specific type of Recurrent Neural Networks, called Long Short-Term Memories, a natural extension of classical RNNs for dealing with long-term dependencies.

The aim of this video is to understand how to actually implement RNNs in R, and use it to solve real-world problems.

The aim of this video is to learn how to train and use an LSTM to solve a complex problem like predicting the next character in a sentence given the occurrences of the previous characters.

The aim of this video is to understand the main differences from classical supervised learning and how they can be combined together.

In this video, you will learn more about a specific unsupervised learning algorithm called Autoencoders. This type of Artificial Neural Networks are simple and effective solutions for learning efficient representation of data without any supervision.

The aim of this video is to learn about two very important unsupervised Deep Learning algorithms for features hierarchies: Restricted Boltzmann Machines and Deep Belief Networks.

The aim of this video is to get a quick picture on the main approaches for solving reinforcement learning tasks with Deep Learning.

The aim of this video is to learn how to train and use an Autoencoders in R with the H2O package, for solving a real-world anomaly detection task.

The aim of this video is to spark new inspiration for creatively applying Deep Learning techniques to real-world problems in Computer Vision.

The aim of this video is to creatively apply Deep Learning techniques to real-world problems in Natural Language Processing NLP.

In this video, we'll creatively apply Deep Learning techniques to real-world problems in ASP.

The aim of this video is to introduce some of the most successful applications of Deep Learning for complex multimodal tasks.

In this video, let's take a look at some of the most successful applications in other fields we didn't mention before.

The aim of this first video is to learn how to deal with models which do not behave as they should.

In this video, you will learn how to speed-up the training and deploy complex DL models.

The aim of this video is to present a complete overview on every available R package for Deep Learning and Neural Networks.

In this video, you will learn about the most interesting research directions and open question for the long-term developments of Deep Learning toward truly intelligent agents.