Guide Journal

Sign in

Introduction to Machine Learning

Machine Learning is relatively new field of study that draws on, intersects and overlaps with other fields such as artificial intelligence. It is a technique that relies primarily on data analytics to train computers to learn from experience as much as humans or animals do, or even better. As with many concepts, different experts define Machine Learning in different ways based on what they prioritize. Nvidia, the American Technology giant defines it as “the practice of using algorithms to parse data from it, and then make a determination or prediction about something in the world.” For Stanford University, it is the science “of getting computers to act without being explicitly programmed,” that is without relying on rules-based programming. As to the Carnegie Mellon University, the field of Machine Learning attempts to build systems autonomously improving as a result of being exposed to experience and deduce the “laws that govern all learning process.”

A more inclusive definition is that given by Daniel Faggella, the founder and CEO at Emerj as he explains that “Machine Learning is the science of getting computers to learn and act like humans do, and improve their learning over time in autonomous fashion, by feeding them data and information in the form of observations and real-world interactions.” Note that definition puts the ultimate objective of Machine Learning in a nutshell, drawing on four experts that he interviewed, Dr. Yoshua Bengio from Université de Montréal, Dr. Danko Nikolic from CSC and Max-Planck Institute, Dr. Roman Yampolskiy from the University of Louisville and Dr. Emily Fox, from the University of Washington. To elaborate further, the learning process in ML relies on algorithms that enhances the performances in an adaptive way in accordance with the number of learning samples.

Applications of Machine Learning :

The objective of the machine learning algorithm is to transform data into insights and hopefully solve a problem. Imagine a retailer wants to have a forecast for next quarter’s sales, a machine learning algorithm can be quite effective here as it can utilize data related to past sales and others to predict them.

Machine Learning has ample applications in different fields, which explains the growing interests in it as well as its fast-paced development. Its algorithms’ ability to detect natural patterns and sequences and reproduce insights that can be used in predicting and decision-making is another factor that seduces experts. What increases its applicability even more is the upsurge in big data, thus rendering it essential in solving problems. In computer biology, ML is used to detect tumors, sequence DNA, discover drug. It is also used for face and object recognition and motion detection in image processing. Energy producers also use ML to forecast price and load. ML is also utilized in predictive maintenance in the automotive and aerospace manufacturing, voice recognition in natural language processing, amongst a plethora of other applications. Commenting on the prospects of Machine Learning, Bill Gates, former Microsoft chairman said “A breakthrough in machine learning would be worth ten Microsofts.”

The Three Components of Machine Learning :

When talking about Machine Learning, it is essential to know that there isn’t one algorithm. Simply put, an algorithm is a mathematic expression of data which is represented in the context of a problem. Each day, programmers and developers publish hundreds of new ML algorithms. These can be put into two main categories, learning style or methods (also form or function). The first simply refers to whether it is supervised learning, unsupervised learning, or semi-supervised learning, whereas the methods refers to classification, regression, decision tree, clustering, deep learning and so forth. In any case, all Machine learning consists of these three components: Representation, Evaluation, and optimization.

Representation: This refers to the hypothesis space, or the space of models (algorithms) allowed and takes into consideration that some models might be encoded much more easily than others in the formal language in which we are expressing models. This happens within one possible set and it is because the landscape of the field that a representation allows, or the possible models. For instance, the computational graph (3-layer feedforward neural networks) and support vector machines with RBF kernels are distinct types of representation.

Evaluation: the second component of Machine Learning refers to the way you assess or choose one model rather than the other. The models are evaluated using scoring function, utility function, fitness function or loss function. The CS professor at the university of Washingtin Pedro Domingos says “Think of this as the height of the landscape for each given model, with lower areas being more preferable/desirable than higher areas (without loss of generality). Mean squared error (of a model’s output vs. the data output) or likelihood (the estimated probability of a model given the observed data) are examples of different evaluation functions that will imply somewhat different heights at each point on a single landscape.”  

Optimization: This refers to the way better evaluations are obtained through searching the space of represented models, which is the search process. In other words, this is the strategy of getting to the “land of models,” in Domingos’ words. combinatorial optimization, convex optimization, constrained optimization are three distinct search processes.

The Ten Methods of Machine Learning :

10 methods are in use in Machine Learning: These are regression, classification, clustering, dimensionality reduction, ensemble methods, neural nets and deep learning, transfer learning, reinforcement learning, natural language processing, and word embeddings.

  1. Regression: regression methods assist in forecasting or explaining a specific numerical value in connection with past data. This can be predicting an estate price, for example, based on previous pricing data related to similar estates.
  2. Classification: This method is used to predict a class. If someone for example wants to know if purchase a service or not, we can use this method because it will classify the output into a yes or no, though the classes are not limited to two.
  3. Clustering: The goal of this method is to group observations with the same characteristics. The quality of the solution is only inspected through visualizations and the algorithm is used to define the output instead of using output information for training.
  4. Dimensionality Reduction: This method is utilized to get rid of unimportant information from a data set. This is essential as it is often the case that data set contains thousands of features. Practically, dimensionality reduction removes unimportant and redundant pixels that have no relevance in the analysis from a photo.
  5. Ensemble Methods: These methods use a combination of predictive models to give a prediction that is superior to the predictions of the individual models. Think about it as a car that you decided to build because you don’t find any of the existing models satisfactory enough, so you combine the best performing parts into one car.
  6. Neural Networks and Deep Learning: here layers of parameters are added to the model in order to meet the objective of neural networks, which is to capture non-linear patters in data and which stands in stark contrast with linear and logistic regressions. Deep learning originates from a neural net that contains numerous hidden layers and architectures.
  7. Transfer Learning: As the term implies, this method gives you the possibility of transferring the knowledge built in one model into the other. Imagine you taught a high quality model to categories photos of hatchback cars, coupés and convertibles for months. Using transfer learning you can pass this knowledge to new models to classify semi-trailer truck, straight truck and jumbo trailer truck. In other words, fractions of trained layers from the neural nets can be mixed with new layers to help you with a similar task.
  8. Reinforcement Learning: This refers to a machine learning method that uses a trial-and-error approach in a specific environment to arrive at a specific result. This learning method is similar to that used to train a mouse to find cheese in a maze: the mouse gets familiar with the maze as you expose it more and more to it, and gets to the cheese faster and faster. This method does not rely on historical data as it learns from its own experience in a specific environment.
  9. Natural Language Processing: This technique is widely used to turn texts or any forms of human language into a numerical representation used in machine learning. The objective of NLP is for the machine to understand, analyze, predict and even generate language.
  10. Word Embeddings: This method is able to capture the context of a specific word, the connections between words, their semantic and syntactic properties in order to, among other things, assign similar mathematic positions to similar words.