What is machine learning?

Artificial intelligence, neural networking and the quest for automated knowledge

In the age of information, data has become a key element of our modern life. It surrounds us, it guides us and in some occassions it even challenges us. The amount and pace of data we are confronted with every day, has increased noticeable over the last years and it becomes harder each day to digest and categorize all these information that bombard us through various channels. To master these challenges, information technology developed several ways to sort, evaluate and separate valuable information from worthless. On this page, we want to give you a short overview what artificial intelligence, machine and deep learning are actually means and how they can be used to intelligently manage the tasks of the modern man.

What is intelligent computing?

Before we dive into the nuances of artificial intelligence, we should reason a bit about intelligence itself. Intelligence is a characteristic usually assigned to highly developed creatures e.g. humans or some primates. Even though there is no general valid definition, intelligence is considered to be the ability to understand and adapt these findings to future actions and decisions. Unlike animals humans are able to put previous experiences in a wider context, to assimilate them, to abstract them and adapt them to future plots. While a dog can only be trained and conditioned, a human learns certain behavior through an abstract educational process.

Computers, as man-made machines, use sophisticated procedures which to most of its users keep being incomprehensible and lack the ability to "think" abstract and intelligent. Instead, they act to  pre-defined principles and patterns (algorithms) to solve complex assignments. In all cases, a computer is only as smart as the developer, who designed and constructed the machine. Machines operate in a tight set of mathematical constraints and lack a minimum of abstract creativity. Still the concepts and method saw a steep rise during the last few years and improvements in the fields of storage and processing technology lowered durations of arithmetic operations dramatically making machine learning finally affordable.

The term "artificial intelligence" is closely linked to the works of the British mathematician Alan Turing who developed the test named after him to describe the brink to artificial intelligence. In a blind conversation (no direct contact given) a human questioner engages into a conversation with one or more participants where one is a machine. If the questioner is not able to determine whether its counterpart is human or not, the machine wins and proves that it features intelligence. At first glance this might sound trivial. The huge achievement however rests in the ability of the machine to answer unforeseen questions and to react reasonably. Subjects, nuances, the whole conversational skills or timing are parameters a human can manage through learning and intelligence. A machine needs to assign far more technical background to evaluate all parameters.

Artificial intelligence lately experiences an higher attention in science and the media. Nearly daily new developments are announced, conference are being held, companies are started and breakthroughs are made. Every aspect of our life seems to be affected. Still, artificial intelligence is currently merely a buzzword than an exact technology.

Machine learning and neuronal networks

Under the umbrella of artificial intelligence, various technologies have been developed, using modern concepts of data processing to handle information faster and more comprehensively. Machine learning is one of the more promising methods. Common to all is the usage of larger data sets and the autonomous procession under various conditions.

Machine learning pursues the approach to enable a computer to autonomously process data using experiences from a prior training without later human interference To do this, the machine creates a model consisting of decision guidelines that either confirm or contradict previous assumptions. What seem to be easy for simple yes or no decisions turns out very complex if the question rests on multiple parameters requiring statistical analysis.
Advanced machine learning is made possible through enhanced computer hardware processing countless parallel calculations and through modern programming concepts using a wider data base. Let alone the sheer amount of texts and images posted on the Internet and social media. This generates vast quantities of exploitable information.

To process information with machine learning, data runs trough several steps during evaluation - a network so to say - based on the pre-trained model. Let's say you try to train a machine to recognize images of green T-Shirts. The trainings data consists of positive as well as false samples. If the machine delivers a wrong result during training, the trainer reacts to it and marks the data. The machine adds the experience to its model and draws its conclusions. At a certain stage the correct results predominate and the model turns out to be robust enough. This is the time the machine might operate autonomously and the results are considered trustworthy.

Cognitive Search

One case of application of machine learning technology can be found with search engines. Unlike exact search which is based on matching search terms, machine learning supports an fuzzy, concept-based approach. Interesting for larger data pools where terms are sometimes not consistently used, this technology clears the way. If you are looking for "electrical bicycle", the search engine might skip all documents using "e-bike" instead. This problem had previously been solved by using synonymous lists (electrical bicycle = E-Bike) which are time-consuming to maintain and costly.
Cognitive search views documents holistic and transfers the sum of embedded information into a mathematical model (a multi-dimensional vector space). Every document is unique and with the help of some basic operations data can be compared. Documents located close to each other might have a certain similarity and are considered relevant for the search results. Not single terms are key anymore but the context itself. This way, cognitive search extends the search functionality into a conceptional search.