Machine Learning: Unveiling the Magic Behind Intelligent Systems

I am Kartik, a Senior Chief Engineer at Samsung Electronics, one of the world's technology leaders in telecommunications equipment and digital convergence. My experience is focused on Information Technology and encompasses leading engineering roles in different companies. In this article, I would like to unveil the topic of Machine Learning: a mysterious part of Artificial Intelligence, which brings opposed feelings to different people. Some think it’s dark and dangerous, while others suppose it enables humanity to go further in an area we’ve never been to before. Either way, it’s a given, and I’m aiming at showing the beauty of Machine Learning, as well as its limitations and obstacles it encounters in the current state of affairs.

Before jumping into the topic, it’s crucial to understand that Machine Learning, or ML for short, recently has been something that lies in the foundation of many everyday things we use. Maps that you use employ ML for finding the best route to avoid traffic jams, and the smart replies you make use of in every mail you send are also available thanks to ML. Much global, but still important for the convenience of our everyday lives are examples of how major financial or governmental institutes prevent fraud or even the way you use your Mobile Banking app when you don’t have time to visit a local bank branch. The ML algorithms are the bedrock for all those things.

Let me then start with the basics.

Intelligent Systems and Machine Learning: the ABCs

Intelligent systems, which sense, analyse and respond to the world around them, are technologically advanced machines that can appear in surprisingly various forms: from home automation systems like robot vacuums and voice assistant speakers to biometric monitoring systems like wrist bands and skin patches.

Machine Learning is the ability of an Intelligent System to learn based on previous results. The system adapts to specific data without programming. In other words, it teaches itself. That’s how nowadays one can insert verbal description and get the output of the visual depiction of it — the system basically processes the collected data and chooses the best to generate the response. Do not freak out yet though, as there is still human contribution necessary to make it work at all.

How does ML work

Unlike computer programs that use raw data to produce useful output, ML works with much more complex tasks. Accordingly, it requires more than just a rule and input data.

A data scientist gives the computer the input (or training) data and explains the desired, or expected output result). The computer uses this information to analyse and create the model, one that has certain parameters that are present in the initial input data. At this point of the process, there is still a need for human assistance: a data scientist selects the most relevant parameters to increase the accuracy of the future output.

Then, having enough of the initial training data, the computer starts to compare the further inserted data with the examples that it has in the model developed. That’s where the ML diverges from a simple program: it produces the prediction rather than a strict causality. This prediction is compared to the result, and the process sometimes can be iterated over and over again, until the desired output is achieved. This allows ML algorithms to learn on their own and produce optimal answers gradually, which inevitably increase in accuracy over time.

However, the ML algorithms vary depending on the nature of the result of the operation. Let me give you a brief explanation of each algorithm.

Types of learning algorithms

• The algorithm of ML working described in the previous section is something called a Supervised learning algorithm. That means that there are two mandatory things: the defined desired result and a data scientist’s oversight.

• Unlike a Supervised one, an Unsupervised learning algorithm doesn’t require either one’s oversight or strictly labelled data. It is employed to find trends, patterns or groups in a dataset in which these components are not known.

• There is also a Semi-supervised algorithm that basically is a fusion of the aforementioned two, and is used in the case where data is partly labelled.

• Another algorithm is called the Reinforcement learning algorithm and is based on the errors, or rather successes produced by the computer. It seeks to find the best solution in a particular environment. The result is evaluated and the feedback is given. Successful actions are reinforced, and that’s how the machine teaches itself.

A little deeper into the diverse world of Machine Learning…

As said earlier, Machine Learning is something different than just a program that follows a strict rule, and there are plenty of sub-fields of it to prove the point. The Neural Network (NN), as an instance, is an extremely complex model the structure of which resembles the intricate structure of the human brain. It comprises nodes that are spread across input layers, hidden layers and output layers. These nodes shall be activated and data transmitted to another layer in the network as soon as output from a single node exceeds its specified verge. It allows NN to multitask and work at a high speed with huge amounts of data. These multiple layers are what make NN a Deep Learning model. In this regard, the network's accuracy may be improved if the number of different layers and nodes is increased.

In the same manner, as a software engineer, I experienced the beauty of Deep Learning in my own work. For example, face detection is one of the areas of Deep Learning that has been extensively explored but still is uncharted territory. NN is capable of detecting the face in a cluttered scene and on variable scales. Thanks to pattern recognition theory, Intelligent Systems (take as an example Face ID in iPhones) are able to make sense of particular features of one’s face. The same is with speech recognition, which is something we use in our ordinary lives for past several years.

Deep Machine Learning is what drives many Intelligent Systems and technologies. As the volume of data grows, so does the computing power. Although ML continues to get deeper and more efficient, there are still some limitations to it. NN requires a lot of training data to start working and learning, and sometimes this data can be biased or inaccurate. Lastly, ML works on predictions, which can possibly mean that some of them might be off.

However, this only tells us that there is still room for work. ML’s importance grows with each day, and there are even greater goals to be achieved with its assistance.