Machine learning is a class of artificial intelligence methods that allow a computer to learn how to solve a problem by precedents. In simpler words – a set of approaches and algorithms that deploys historical data to search for specific patterns and uses these patterns to solve specific problems.
There is often confusion in the terminology: artificial intelligence, machine learning, and deep learning. Let’s take a look at the historical development. The term artificial intelligence appeared in the early 1950s and was used to refer to machines and programs. In the 1980s, the term machine learning appeared – the emergence of intelligent programs that learn from data without explicit programming. In 2010, a new term appears – deep learning – an algorithm that refers to machine learning using multi-layer neural networks. Neural networks are not a new invention, as they are often presented, but a class of algorithms that has been in existence for a very long time. However, thanks to our recent capacity to process a huge amount of data we are now able to run these types of algorithms.
Thus, these three types of technologies cannot be considered different. Artificial intelligence includes machine learning, which in turn includes deep learning algorithms. There are three types of machine learning – supervised learning, unsupervised learning and reinforcement learning.
Supervised learning is the most common type of machine learning. The algorithm is trained on an existing set of historical data on the object, for example, the readings of the sensors of the machine unit, such as temperature, humidity, pressure, as well as on the set of answers, for example, the useful output of this unit. These algorithms are useful whenever there is a correlation between the readings of the sensors at the machine and the responses, although the relationship is unknown, and when the number of precedents or the number of object-response pairs, or the data set, is known. Based on the historical data, the aim is to determine the hidden dependency and to build an algorithm capable of producing a fairly accurate answer for any object.
This type of machine learning allows you to solve a number of problems called classification and regression.
For example, a classification task would apply to determine whether a selected group of people has COVID-19 based on the results of x-rays. X-ray classification allows you to assign the probability of belonging to people with a positive result (infected) and people with a negative result (healthy). To build such an algorithm, we need to collect a training sample – x-rays with the marks of sick people and healthy one. Labels are affixed by a professional doctor. Next, we can teach our algorithm to automatically predict the presence of COVID-19 disease. To enrich our algorithm, we can also use regression – for each person in the selected group we predict numerical values. In our study of x-rays, we can enrich our algorithm with geographic location, gender, population density, and human body temperature and thus predict the spread of the virus. Unfortunately, at the moment there is not enough quality data on COVID-19 for the effective training of automatic image determination systems.
Another machine learning method is the unsupervised learning. In this method, the algorithm is taught to look for clusters and associations on unallocated data. This method is applied when there is information about a particular process that allows you to find a useful hidden structure in the data describing this process. For example, to determine the abnormal behavior of the sensors, which can lead to damage to the machine unit. These algorithms are used in the preventive protection of equipment from breakdowns!
Lastly, reinforcement learning is a type of machine learning in which an algorithm is an agent that interacts with an environment, such as a machine unit, and learns how to optimally control it with respect to some parameter, for example, the electricity consumption. The agent makes a control action and receives a reward if his actions which in turn leads to an improvement in the process. For example, to achieve energy savings while respecting a minimum output performance of the process.
The choice of algorithm largely depends on the problem being solved. There is no universal algorithm that can solve the problem better than other. Only through tests and experiments can we determine which algorithm works better or worse. In general, machine learning is a very interesting technology and different people take part in its development and rollout. From data architects and engineers that gather and clean-up data, to data scientists that apply the algorithms and business analyst that are able to identify the business need to be addressed by these new technologies. The creation of these algorithms often requires the adoption of innovative solutions and a creative approach that can impact people and processes as well as the tools that need to be used. Expert freelancers can be sourced to setup the optimal team or to upskill your organization to develop this function.
There is a phobia that soon artificial intelligence including robots will crowd out many people in the workplace. Indeed, a number of functions such as document management, procurement of goods, search for candidates, HR functions and accounting among others are now actively being replaced by robots. But many other functions and roles will appear that will require the irreplaceable human touch.