What is machine learning ?

You must have heard about machine learning many times lately. It is also known as artificial intelligence; machine learning is a division of AI, both can be traced with their sources from MIT in the 1950s.

Machine learning is usually encountered every day, whether you believe it or not. The technology that maintains self-driving cars from crashing into things, Alexa and Siri voice assistants, Amazon and Netflix recommendations and Facebook’s and Microsoft’s facial detection all are results of advances in machine learning.

While it is not too complex for human brain, systems that depend on machine learning have accomplished some remarkable feats such as defeating human challengers at Jeopardy, Texas Hold ‘em, chess, and Go,

Neglected for decades as unrealistic and overhyped (the recognized”AI winter”), both machine learning and AI have enjoyed a great resurgence in the last few years. Thanks to many technological breakthroughs and the huge explosion in simple computing horsepower, including the price of machine learning models.

Self-trained software

What is machine learning? Let’s start by saying that machine learning is not human-programmed, conventional and hand-coded computing application.

Unlike conventional software that follows instructions but not better at improvising, machine learning systems are coded themselves efficiently and developed their instructions by general examples.

The model example is image detection. This shows that machine learning system has enough photos of labeled dogs, including images of babies, trees, bananas, cats, or any other object apart from marked "not dogs." If the system is appropriately trained, it will undoubtedly get better at recognizing canines without human being detecting what a dog is supposed to look like.

The spam filter inside your email program is the best example of machine learning in action. After it has been exposed to many spam samples, including non-spam email, it knows to recognize the significant characteristics of those unwanted and malicious messages. It is not just right, but it’s pretty accurate.

Unsupervised and Supervised learning

This type of machine learning is known as supervised learning that means that someone can uncover the machine learning algorithm to a massive set of training data and look at its output. After constantly tweaked its settings until you get the expected result when displayed on data it had not seen before. This process is similar to clicking “not spam” push button in your inbox when the filter clicks on legitimate message mistakenly. The more you follow this process, the more accurate the filter ought to get better.

The most recognize supervised learning tasks involve prediction (i.e., “regression”) and classification. Predicting stock prices is a good example of a regression problem.

The second type of machine learning is known as unsupervised learning. The system pores over large amounts of data to learn how “normal” data looks like, to recognize hidden and anomalies patterns. Unsupervised machine learning is helpful when you do not know what you are looking for and it is not possible to train the system to find it.

Unsupervised machine learning systems can detect patterns in a large amount of data many times quicker than humans. That is the reason why banks make use of them to identify fraudulent transactions; security software used them to unfriendly flag activity on a network and marketers employ them to detect customers with similar attributes.

The two examples of unsupervised learning algorithms are clustering and association. Clustering is a unique tool for customer segmentation while association rule learning can be used for recommending engines.

Restrictions of machine learning

Since every machine learning system builds its connections, the way in which one works can involve a little black box. It is not possible to reverse engineering process to find out why your system can be distinguished between a Persian and Pekingese. It doesn’t matter, as long as the system is working properly.

A machine learning system is also good as the data that has been exposed. A typical example is “garbage in, garbage out.” When a system is exposed or poorly trained to a small data set, a machine learning algorithm can generate results discriminatory apart from providing wrong information.

In 2009, HP got it trouble when facial technology recognition built into the webcam on an HP MediaSmart notebook was unable to recognize the faces of African Americans. Also, in June 2015, flawed algorithms in the Google Photos app mistakenly labeled two black Americans as gorillas.

Another remarkable example: In March 2016, Microsoft’s unfortunate Taybot experimented to check if an AI system copy human conversations by learning from tweets. In less than 24 hours, suspicious Twitter trolls had changed Tay into a hate-speech-sending out chatbot from hell.

A machine learning glossary

Machine learning is the perfect tip of the AI berg. Many other terms closely related to machine learning are deep learning, cognitive computing, and neural networks.

Neural network: A computer structural designed to copy the structure of neurons in human brains, with all artificial neuron (microcircuit) linking other neurons inside the system. Neural networks are arranged in layers along with neurons. Each layer connects data to multiple neurons in the next layer until they get to the output layer. The final layer is the region where the neural network offers its best predictions to determine the dog-shaped object along with a confidence score.

There are many types of neural networks for solving multiple types of problems. Networks with a large number of layers are known as “deep neural networks.” Neural networks are one of the essential tools used in machine learning situations, but it is not the only method.

Deep learning: This is a type of machine learning through steroids with the use of multi-layered (deep) neural networks to achieve a result based on incomplete or “imperfect” information. DeepStack is the deep learning system that defeated specialized poker players through continuous computing its approach after each round of stakes.

Cognitive computing: This is the term used by creators of Watson, IBM, and the workstation that kicked humanity’s ass at risk in 2011. The variation between artificial intelligence and cognitive computing, in IBM’s point of view, instead of changing the human intellectual, cognitive computing is intended to supplement it. These processes enable medical doctors to diagnose patients more effectively, help financial executives to make good recommendations and lawyers to check law cases faster, etc.

This is an extremely superficial impression. Also, people that want to go deeply into machine learning and the intricacies of AI can begin with this semi-wonky lesson from the University of Washington’s Pedro Domingos or InfoWorld’s Martin Heller on “The meaning of deep learning” including series of Medium posts from Adam Geitgey.

Regardless of the hype about AI, it is not an exaggeration to say that machine learning including the related technologies in placed are transforming the world as we all know it. The best time to learn more about it is now before the machines become entirely self-aware.