Machine Learning has recently seen a major upsurgence in academic & business world as well, thanks to the enormous amount of data generated everyday by the Multinational companies, instituitions & even online services. Keeping track of these data & making sense of them, is a daunting task; but thankfully, equipped with an increase in the computational powers & a proper knowledge in machine learning, we can rise to the challenge. Businesses worldwide are looking out for skilled machine learning specialists to solve these challenges for them, and the compensations for these posts are equally handsome.
But let us circle back to the basics first: what is machine learning? According to the Godfather of machine learning, Tom Mitchell himself, a machine is said to “learn from experience E with respect to some class of tasks T and performance measure P if its performance at tasks in T, as measured by P, improves with experience E”. This definition is a widely quoted one, and for good reasons; the definition captures the complete essence of machine learning without being excessively technical. Other sources may define machine learning in their own words; another popular definition explains machine learning as a method to train machines to learn or create knowledge on their own, without explicitly programming them.
We need not dive deep into the technicalities of machine learning for this brief discussion. Here, we shall merely try to concern ourselves with the various aspects, uses & types of machine learning, & see if that interests us enough to study it further. So without further ado, let us start.
Remember the time you opened up Facebook only to see an old friend in your “Suggested Friends” list? Or the time Amazon recommended you some item based on your previous purchases? How about the time Spotify surprised you with a custom playlist, based on your listening habits? How do you think these came to be? Two words can sum up the answer for all three questions: Machine Learning. Machine learning is, in fact, used everywhere nowadays, from the photo filter in your app to the fitness tracker you wear regularly. Machine learning offers us intelligent solutions & insights to very critical problems, hence almost all verticals are trying to benefit from it. Some common examples where machine learning is used for practical purposes are:
Banks regularly use machine learning methods to predict the possibility of a person defaulting on their loans. Credit card companies take advantage of machine learning to recognise abnormal patterns in a card holder’s usage to detect frauds.
Social apps like Facebook, photo editing apps like Retro cam use higly advanced machine learning models to recognise faces in images & apply filters on them. These models are trained over a long period of time on millions of faces to achieve these results.
Have you ever asked Siri for the weather, or for the current stock prices? Cortana regularly interacts with you to answer your spoken queries, right? So how can a machine recognize & understand your words? Again, the answer lies in machine learning. It is extensively used to perform voice searches, speech recognition, speech synthesis and more.
Machine learning has found recent uses in the medical also. There are ongoing researches in identifying cancerous cells & tissues using machine learning.
So how did it all begin? Well, back in 1958, an American cognitive psychologist by the name of Frank Rosenblatt devised an elctronic device, named Perceptron, capable of limited perception & learning, like humans. This device was constructed by closely mimicking the neural workings of the human brain; however this idea did not pick up pace for a long time until the late 1980s & early 1990s when, with the advent of powerful computational machines, large memory capacities & a renewed interest in statistics, machine learning suddenly became the rage. It was around this time statisical approaches were absorbed into computer science to yield astonishing results in data driven sciences. Availability of huge amounts of data also enabled researchers to build intelligent systems to perform large-scale computations & analyses that were previously thought to be impossible. This brought about a huge shift in the favour of machine learning, and this wave continues even today. With the generation of huge volumes of data everyday around the world, making sense of it is an extremely challenging task; thanks to machine learning, we have made significant progresses in this regard.
Machine learning algorithms mostly help in classification or prediction tasks, but they do so without being explicitly programmed. They are able to do so by repeatedly learning the underlying general patterns hidden in the data, and then giving outputs based on that learned knowledge. The learning is performed completely on it’s own, and the knowledge in its memory is represented as mathematical combinations & patterns that properly capture the hidden generalisations. The learning however, can be performed through a variety of methods; we discuss 3 of the most popular methods here:
In this type of learning, the learner is presented with the labeled data, that is, the system learns to map each input to it’s correct output from the dataset itself during the training phase. Supervised learning is commonly used during classification & regression purposes. During testing, the learner is asked to map new unlabeled input data to their proper labels, & the system is judged based on its accuracy.
Example: Spam detection falls under supervised learning. Initially we take some input emails & mark them “Spam” or “Not Spam” according to their contents, then we train the model on this labeled data. Once trained, we test the model by providing it with new unlabeled emails & checking its accuracy based on its predictions.
In this type of learning, the learner is presented with unlabeled, uncategorized data from the beginning, and the system has to learn by itself the mappings for each input to it’s correct output. The mapping may not be correct the first time, so the system has to learn again & again, unless the mapping is deemed satisfactory, or over some threshold value, measured using model accuracy. Artificially intelligent systems are tested in unfamiliar environments using unsupervised learning.
Example: Identifying different clusters among seemingly similar data points using K Means clustering.Types of Unsupervised learning.
In this type of machine learning, the agent, or the system, learns on its own by interacting with its environment. On taking correct steps, it is rewarded & for incorrect steps, it is similarly penalised. At the end of it’s journey, the system wants to maximise rewards; so it tries to learn by approximating correct steps in its environment through mainly dynamic programming paradigms. The agent may or may not learn have human intervention for assistance.
Example: Autonomous & self-driving vehicles learn to drive on their own using this method of reinforcement learning. If it runs into an obstacle, it is penalised & otherwise, rewarded. The agent vehicle tries to maximise its rewards, & in turn, learns to avoid obstacles in its path.