Thanks to Sebastian from Retresco for sharing!
Artificial intelligence, machine learning, deep learning, and natural language processing – when the conversation turns to artificial intelligence and the ubiquitous buzzwords begin to fly, non-developers soon switch off. These terms are often used too broadly, synonymously, or simply incorrectly. To shed a little light on this matter and to help you deal with AI applications successfully, we thought we would clear up any terminological confusion.
Where better to start then at the beginning?
Artificial intelligence (AI for short) is a general term covering all fields of research investigating the application of human intelligence by machines.
Often AI is further subdivided into “strong” and “weak” AI. Weak AI, which is in use today, can effectively simulate individual human capabilities, while strong AI, which is still a long way from reality, can think, reason, and act as independently as a human being.
Machine learning (also known as ML), a term often used synonymously with artificial intelligence, is actually a sub-field of AI.
ML concerns the capacity to learn autonomously. So it is not just about imitating processes (or intelligence), but also about using experience to generate knowledge.
In the case of machine learning, this means learning statistical patterns from representative historical data. Today, this enables processes which previously had to be carried out manually to be fully or partly automated and thus optimised.
For example, all a modern machine learning-based image recognition system needs to produce reliable results as examples of correctly tagged images. The machine learning algorithm uses this representative data to learn how to identify specific objects. It would be too complex to write general if-then rules precisely defining an object’s appearance at pixel level. A rules-based expert system would probably never achieve the same performance.
Unlike machine learning algorithms, rules-based systems are also non-transferable. A system that classifies German texts using rules developed manually could probably not be used for Chinese texts. However, if data for German and Chinese are available, there is a good chance that, after language-specific pre-processing, the same machine learning algorithm could learn how to classify both German and Chinese texts as required.
Consequently, for a rules-based expert system faced with a complex problem, the initial development process is long and costly. And because the resulting systems are as complex as the original problem, maintenance and subsequent development is also laborious. By contrast, a machine learning-based system typically has significantly less code because the logic does not need to be explicitly stated, bringing advantages in terms of both cost and performance.
In situations where future decisions can be taken on the basis of knowledge drawn from historical data, machine learning is an ideal tool for full or partial automation of processes – provided that data relating to the problem at hand is already available or can be easily obtained.
Deep learning (also known as DL) is in turn a sub-domain of machine learning. DL systems take the knowledge learned by ML systems further: using large volumes of historical data, they act and evolve autonomously, i.e. with no need for existing knowledge to be programmed in beforehand.
They can do this because of the way they operate, which is inspired by the neuronal networks in the human brain. Thanks to their complex (and therefore deeper) structure, systems of this kind use relevant training data to learn basic patterns and infer more complex patterns independently. Deep learning models can become so complex that it is not possible to control or track exactly what they learn. While DL is also limited to the problems dealt with in the data it is presented with, it can under the right conditions produce significantly better results than simpler ML algorithms.
Deep learning requires enormous volumes of data to search for these patterns. Typical applications for DL today include facial, object, and speech recognition.
This makes machine learning and deep learning into tools for automating systems. At Retresco, we use these tools – machine learning at least – in the development of our products. To see how all these concepts relate to each other, it is helpful to look at how our technologies fit into the big picture.
Natural language processing (or NLP) combines linguistics and artificial intelligence to enable machines to process natural human language.
The term NLP is a general term which covers the technologies of natural language understanding (NLU) and natural language generation (NLG, also known as automated text generation).
In simple terms: turning text into data, and data into text.
At Retresco we use machine learning in virtually all our semantic applications to carry out tasks such as automated understanding of content (to categorise texts, for example), clustering of news articles, or understanding the inputs of a chatbot user. In our Explainer Engine, developed for our partner mysimpleshow.de, we use machine learning to suggest appropriate visual aids for relevant keywords. And our SaaS solution for automated text generation, textengine.io, uses machine learning to support users by providing an automatic linguistic analysis of the templates the user writes.
In summary: Artificial intelligence is a general term. Machine learning and deep learning are subfields of artificial intelligence which describe methods for automating processes. These processes are used for natural language processing, in which artificial intelligence is combined with linguistics to enable human language to be processed by machines. So, next time you find yourself in a conversation about AI you'll be able to follow along or even contribute.