To create artificial intelligence (AI), we first need to clarify what intelligence is. There are many definitions and approaches, as intelligence is a very complex concept that refers to the ability to reason, understand, be aware, learn, emotional reason, plan, be creative and solve problems. A breakthrough in artificial intelligence research has been the realisation that intelligence is not the same as knowledge, but rather the ability to learn and generalise from learning.
Artificial intelligence in the ordinary language
In everyday spoken language, artificial intelligence is used in a broader sense, and is understood to mean several things at once. We call AI machines and software that can respond to environmental influences without constant human intervention – the high level of automation represented, for example, by self-driving cars. We also call AI systems that are able to behave in a similar way to a living being with natural intelligence, i.e. they simulate human behaviour – in this sense we can talk about the intelligence of NPCs (non-player characters) in computer games, for example. Alternatively, artificial intelligence is when the machine or system is able to change its behaviour in a purposeful and repeatable way, in other words, it is capable of learning – this is the main focus of modern AI research and is the most commonly identified with the term artificial intelligence, and the functionality of the previous two categories is achieved through the latter, machine learning.
What can artificial intelligence learn?
There are two main schools of thought in artificial intelligence research. One is the knowledge-based approach, which is basically the programming of human knowledge, but this is extremely labour- and human-intensive, and not even fully feasible. The other is a learning-based style called machine learning – which requires not just one or two examples, but thousands of pieces of data. But the accuracy of machine learning systems can never be 100%, there is always a margin of error. Within machine learning, we tend to distinguish three different methods. The first is supervised learning, where we train our model to know which picture has a cat in it based on data that we interpret, such as labelled pictures – after being presented with a sufficient number of pictures, the AI model will be able to decide with high accuracy whether or not a picture contains a cat, which it has not seen before. The second is unsupervised learning, where the model teaches itself to look for relationships between data – for example, looking for similarities and differences between pictures of fruit. And the third is reinforcement learning, where the AI is “rewarded” when it does something right – this is how they create go- or chess machines that humans can’t beat.
All three learning paradigms have their own justification. Supervised learning is applicable when we know the categories and have many labelled data or can easily produce them. Unsupervised is most effective when we don’t know or don’t want to do the categorisation, but leave it to the machine. And with reinforcement learning, we know the framework of the operation, but the machine does the design of the specific steps. The key is to find the right method to solve a business problem in AI – that’s what the Neuron Solutions team does!